PAST

PAleontological STatistics

Version 5.4

Reference manual

Øyvind Hammer

Natural History Museum

University of Oslo

ohammer@nhm.uio.no

1999-2026

1

Contents

Welcome to the PAST! ............................................................................................................................ 8

Installation............................................................................................................................................... 9

Windows.......................................................................................................................................... 9

Mac.................................................................................................................................................. 9

Quick start ............................................................................................................................................. 10

How do I export graphics?............................................................................................................. 10

How do I organize data into groups? ............................................................................................ 10

The spreadsheet and the Edit menu ..................................................................................................... 11

Entering data ..................................................................................................................................... 11

Selecting areas................................................................................................................................... 11

Moving a row or a column ................................................................................................................ 12

Renaming rows and columns ............................................................................................................ 12

Increasing the size of the array ......................................................................................................... 12

Cut, copy, paste................................................................................................................................. 12

Remove.............................................................................................................................................. 12

Row colors and symbols.................................................................................................................... 12

Selecting datatypes for columns, and specifying groups .................................................................. 12

Remove uninformative rows/columns.............................................................................................. 13

Transpose .......................................................................................................................................... 13

Grouped columns to multivar ........................................................................................................... 13

Grouped rows to multivar ................................................................................................................. 14

Stack grouped rows into columns ..................................................................................................... 14

Rows, cols, values to table ................................................................................................................ 14

Value pairs to matrix ......................................................................................................................... 14

Samples to events (UA to RASC)........................................................................................................ 14

Events to samples (RASC to UA)........................................................................................................ 15

Loading and saving data.................................................................................................................... 15

Importing data from Excel................................................................................................................. 16

Reading and writing Nexus files ........................................................................................................ 16

Counter.............................................................................................................................................. 16

Transform menu.................................................................................................................................... 17

Logarithm .......................................................................................................................................... 17

Subtract mean ................................................................................................................................... 17

Remove trend.................................................................................................................................... 17

Row percentage ................................................................................................................................ 17

Row normalize length........................................................................................................................ 17

Box-Cox.............................................................................................................................................. 17

Compositional data transforms......................................................................................................... 18

Remove size from distances.............................................................................................................. 20

Landmarks, Procrustes fitting............................................................................................................ 20

Landmarks, Bookstein fitting............................................................................................................. 21

Project to tangent space (not yet in Past 4)...................................................................................... 21

Remove size from landmarks (not yet in Past 4)............................................................................... 21

Transform landmarks ........................................................................................................................ 22

Regular interpolation ........................................................................................................................ 22

2

Evaluate expression........................................................................................................................... 22

Plot menu .............................................................................................................................................. 24

Graph................................................................................................................................................. 24

XY graph............................................................................................................................................. 25

XY graph with error bars ................................................................................................................... 26

Histogram .......................................................................................................................................... 27

Bar chart/box plot ............................................................................................................................. 28

Pie chart............................................................................................................................................. 30

Stacked chart..................................................................................................................................... 31

Percentiles......................................................................................................................................... 32

Normal probability plot..................................................................................................................... 33

Ternary .............................................................................................................................................. 34

Bubble plot ........................................................................................................................................ 35

Matrix plot......................................................................................................................................... 36

Mosaic plot........................................................................................................................................ 37

Venn diagram .................................................................................................................................... 38

Radar chart ........................................................................................................................................ 40

Polar plot ........................................................................................................................................... 41

Vector plot......................................................................................................................................... 42

Network plot ..................................................................................................................................... 43

3D scatter/bubble/line plot............................................................................................................... 44

3D surface plot .................................................................................................................................. 45

3D parametric surface plot................................................................................................................ 46

Univariate menu.................................................................................................................................... 47

Summary statistics ............................................................................................................................ 47

One-sample tests............................................................................................................................... 50

Two-sample tests .............................................................................................................................. 52

t test and related tests for equal means ....................................................................................... 52

F test for equal variances .............................................................................................................. 55

Reference ...................................................................................................................................... 55

Mann-Whitney test for statistic equality ...................................................................................... 55

Brunner-Munzel test for stochastic equality................................................................................. 57

Mood’s median test for equal medians ........................................................................................ 57

Kolmogorov-Smirnov test for equal distributions......................................................................... 58

Anderson-Darling test for equal distributions............................................................................... 59

Epps-Singleton test for equal distributions................................................................................... 60

Coefficient of variation (Fligner-Kileen test) ................................................................................. 61

F and t tests from parameters........................................................................................................... 62

Two-sample paired tests (t, sign, Wilcoxon) ..................................................................................... 63

Several-sample tests ......................................................................................................................... 65

Several-samples repeated measures tests........................................................................................ 72

3

Two-way ANOVA ............................................................................................................................... 76

Two-way ANOVA without replication ............................................................................................... 79

Two-way repeated measures ANOVA............................................................................................... 80

One-way ANCOVA ............................................................................................................................. 81

Correlation table ............................................................................................................................... 82

Intraclass correlation......................................................................................................................... 87

Normality tests .................................................................................................................................. 89

Outlier tests....................................................................................................................................... 92

Contingency table (chi²etc.) ............................................................................................................. 94

Cochran-Mantel-Haenszel test.......................................................................................................... 96

Risk/odds........................................................................................................................................... 97

Single proportion............................................................................................................................... 99

Multiple proportion confidence intervals ....................................................................................... 100

Ratio of counts confidence interval ................................................................................................ 101

Survival analysis (Kaplan-Meier curves, log-rank test etc.)............................................................. 102

Combine errors................................................................................................................................ 104

Propagate errors ............................................................................................................................. 104

Multivariate menu............................................................................................................................... 106

Principal components...................................................................................................................... 106

Principal coordinates....................................................................................................................... 111

Non-metric MDS.............................................................................................................................. 112

Correspondence analysis................................................................................................................. 114

Detrended correspondence analysis............................................................................................... 116

Canonical correspondence.............................................................................................................. 117

Seriation .......................................................................................................................................... 118

CABFAC factor analysis.................................................................................................................... 119

Discriminant analysis....................................................................................................................... 120

Two-block PLS.................................................................................................................................. 122

Redundancy analysis (RDA)............................................................................................................. 123

Nonlinear ordination (UMAP) ......................................................................................................... 124

Nonlinear ordination (ISOMAP) ...................................................................................................... 126

Classical cluster analysis.................................................................................................................. 127

Neighbour joining............................................................................................................................ 129

K-means clustering .......................................................................................................................... 130

K-medoids clustering....................................................................................................................... 131

DBSCAN clustering........................................................................................................................... 131

K-nearest neighbors classifier ......................................................................................................... 132

Naïve Bayes classifier ...................................................................................................................... 133

Decision tree classification.............................................................................................................. 134

Random forest classifier.................................................................................................................. 135

Parsimony (cladistic)........................................................................................................................ 136

Phylogenetically independent contrasts......................................................................................... 139

Phylogenetic generalized least squares (PGLS)............................................................................... 140

Multivariate normality .................................................................................................................... 141

Box’s M............................................................................................................................................ 142

MANOVA ......................................................................................................................................... 143

One-way ANOSIM............................................................................................................................ 144

One-way PERMANOVA.................................................................................................................... 145

Two-way ANOSIM ........................................................................................................................... 147

Two-way ANOSIM without replication............................................................................................ 147

4

Two-way PERMANOVA.................................................................................................................... 147

Test for multivariate dispersion (PERMDISP).................................................................................. 148

Mantel test and partial Mantel test ................................................................................................ 149

SIMPER ............................................................................................................................................ 151

Indicator species analysis (IndVal) .................................................................................................. 152

Paired Hotelling............................................................................................................................... 153

Modern Analog Technique.............................................................................................................. 154

Weighted averaging calibration (WA) ............................................................................................. 156

Weighted averaging partial least squares (WA-PLS)....................................................................... 157

Similarity and distance indices ........................................................................................................ 160

Genetic sequence stats ................................................................................................................... 166

Model menu ........................................................................................................................................ 167

Linear, bivariate............................................................................................................................... 167

Linear, multivariate (one independent, n dependent) ................................................................... 172

Linear, multiple (one dependent, n independent).......................................................................... 173

Linear, multivariate multiple (m independent, n dependent) ........................................................ 174

Generalized Linear Model ............................................................................................................... 175

Polynomial regression ..................................................................................................................... 177

Nonlinear......................................................................................................................................... 178

Sinusoidal regression....................................................................................................................... 181

Smoothing spline............................................................................................................................. 183

LOESS smoothing............................................................................................................................. 184

Mixture analysis .............................................................................................................................. 186

Abundance models.......................................................................................................................... 188

Species packing (Gaussian).............................................................................................................. 190

Fit spiral ........................................................................................................................................... 191

Changepoint modeling .................................................................................................................... 192

Diversity menu .................................................................................................................................... 194

Alpha diversity indices..................................................................................................................... 194

Quadrat richness ............................................................................................................................. 199

Beta diversity................................................................................................................................... 202

Taxonomic ....................................................................................................................................... 204

Individual rarefaction ...................................................................................................................... 206

Shareholder Quorum Subsampling (SQS) ....................................................................................... 209

Sample rarefaction (Mao’s tau)....................................................................................................... 210

SHE analysis..................................................................................................................................... 212

Diversity permutation test .............................................................................................................. 213

Diversity t test ................................................................................................................................. 214

Diversity profiles.............................................................................................................................. 215

Time series menu ................................................................................................................................ 216

Simple periodogram........................................................................................................................ 216

REDFIT spectral analysis .................................................................................................................. 217

Multitaper spectral analysis ............................................................................................................ 219

Walsh transform.............................................................................................................................. 220

Evolutive Fourier transform ............................................................................................................ 221

Wavelet transform .......................................................................................................................... 222

Wavelets for unequal spacing ......................................................................................................... 224

Point events spectrum .................................................................................................................... 225

Autocorrelation ............................................................................................................................... 227

Autoassociation............................................................................................................................... 228

5

Cross-correlation ............................................................................................................................. 230

Mantel correlogram (and periodogram) ......................................................................................... 231

Runs test.......................................................................................................................................... 233

Mann-Kendall trend test ................................................................................................................. 234

Point events..................................................................................................................................... 235

Markov chain................................................................................................................................... 237

ARMA (and intervention analysis)................................................................................................... 238

Simple smoothers............................................................................................................................ 240

FIR filter ........................................................................................................................................... 242

Insolation (solar forcing) model ...................................................................................................... 244

Date/time conversion...................................................................................................................... 245

Geometrical menu............................................................................................................................... 246

Circular (one sample) ...................................................................................................................... 246

Circular (two samples)..................................................................................................................... 249

Circular correlation.......................................................................................................................... 251

Spherical (one sample).................................................................................................................... 252

Point pattern analysis - nearest neighbours ................................................................................... 254

Ripley’s K point pattern analysis ..................................................................................................... 256

Correlation length analysis.............................................................................................................. 258

Minimal spanning tree analysis....................................................................................................... 259

Kernel density.................................................................................................................................. 260

Point alignments.............................................................................................................................. 261

Quadrat counts................................................................................................................................ 262

Spatial autocorrelation (Moran’s I) ................................................................................................. 264

Gridding (spatial interpolation)....................................................................................................... 266

Multivariate allometry .................................................................................................................... 268

PCA of 2D landmarks (relative warps)............................................................................................. 269

Thin-plate splines for 2D landmarks................................................................................................ 270

Linear regression of 2D landmarks.................................................................................................. 270

Common Allometric Component analysis for 2D landmarks .......................................................... 271

PCA of 3D landmarks....................................................................................................................... 272

Linear regression of 3D landmarks.................................................................................................. 272

Common Allometric Component analysis for 3D landmarks .......................................................... 272

Size from landmarks (2D or 3D) NOT YET IN PAST 4 ....................................................................... 273

All distances from landmarks (EDMA) NOT YET IN PAST 3 ............................................................. 274

Edit landmark lines/polygons.......................................................................................................... 275

Elliptic Fourier shape analysis ......................................................................................................... 276

Hangle Fourier shape analysis......................................................................................................... 277

Eigenshapes..................................................................................................................................... 278

Coordinate transformation ............................................................................................................. 279

Open Street Map............................................................................................................................. 280

Measure on image........................................................................................................................... 281

Stratigraphy menu............................................................................................................................... 282

Unitary Associations........................................................................................................................ 282

Ranking-Scaling................................................................................................................................ 287

Range confidence intervals ............................................................................................................. 290

Distribution-free range confidence intervals .................................................................................. 291

Sequence slotting ............................................................................................................................ 292

Diversity and rates........................................................................................................................... 293

Stratigraphic chart........................................................................................................................... 295

6

Radiocarbon calibration .................................................................................................................. 297

Grain size analysis............................................................................................................................ 300

Scripting............................................................................................................................................... 302

Language structure.......................................................................................................................... 302

The output window ......................................................................................................................... 309

Accessing the main Past spreadsheet and menus .......................................................................... 311

Array and vector operations ........................................................................................................... 312

Scalar math functions...................................................................................................................... 312

File I/O ............................................................................................................................................. 313

String operations............................................................................................................................. 314

Other functions ............................................................................................................................... 316

Calling dll functions (Windows only)............................................................................................... 317

Libraries and classes........................................................................................................................ 318

Forms and components................................................................................................................... 319

7

Welcome to the PAST!

This program was originally designed as a follow-up to PALSTAT, a software package for

paleontological data analysis written by P.D. Ryan, D.A.T. Harper and J.S. Whalley (Ryan et al. 1995).

Through continuous development for more than twenty years, PAST has grown into a comprehensive

statistics package used not only by paleontologists, but in many fields of life science, earth science,

engineering and economics.

Further explanations of many of the techniques implemented together with case histories are found

in the book “Paleontological data analysis” (Hammer & Harper 2024).

If you have questions, bug reports, suggestions for improvements or other comments, we would be

happy to hear from you. Contact us at ohammer@nhm.uio.no. For bug reports, remember to send us

the data used, as saved from PAST, together with a complete description of the actions that lead to

the problem.

The latest version of Past, together with documentation and a link to the Past mailing list, is found at

https://www.nhm.uio.no/english/research/resources/past

We are grateful if you cite PAST in scientific publications. The official reference is Hammer et al.

(2001).

References

Hammer, Ø. & Harper, D.A.T. 2024. Paleontological Data Analysis, 2^nded. Elsevier.

Hammer, Ø., Harper, D.A.T., and P. D. Ryan, 2001. PAST: Paleontological Statistics Software Package

for Education and Data Analysis. Palaeontologia Electronica 4(1): 9pp.

Harper, D.A.T. (ed.). 1999. Numerical Palaeobiology. John Wiley & Sons.

8

Installation

Windows

Download the past5setup5.zip (or with other version number) file, uncompress and run. Windows

will consider this a breach of security and will ask if you trust the software provider. If you want to

use the program, you will have to answer yes. This procedure requires administrator rights.

If you do not have admin rights, just download the file ‘Past5.1.zip’ (zipped) and put it anywhere on

your hard disk. Double-clicking the file will start the program. However, with this installation mode,

the map module (OpenStreet) will not work, and also there will be an error message on exiting Past,

which you can disregard.

Mac

Past for Mac must be downloaded from the Apple Store. Search for “Past5” there or use the link on

the Past home page. Only newer Macs with ARM processors (M1, M2, M3 etc.) are supported.

Command-line parameters

If you start Past from the command line, you can specify a data file name as the first parameter, and

optionally a scripting file name as a second parameter (this script will be loaded and run

automatically on startup).

9

Quick start

Past is to some extent self-explanatory, but a couple of important functions are a bit difficult to find:

How do I export graphics?

For publication quality, save the graphic in the SVG or PDF vector format. Click the “Graph settings”

button next to the graphic. In the graph preferences window, click the “Export picture” button

(arrow on the right). You can open SVG files in Adobe Illustrator, Corel Draw or the free program

Inkscape. SVG files are supported by most web browsers, and can be placed directly on a web page.

You can also export the picture in bitmap formats (JPG, TIF etc.), but the quality is lower, and you

cannot easily edit the graphic. Or you can copy-paste the image as a bitmap by clicking the “Copy”

button under the graphic.

Note: On Mac, you must manually provide the correct file extension (e.g. .svg or .jpg) when typing

the file name, or the file type will not be recognized by other programs.

How do I organize data into groups?

This requires a separate group column, with a group identifier for each row. In the example (1) there

is a group column “Lithology”, with two groups LS and MS. To specify that “Lithology” is a group

column, first select the “Column attributes” box above the spreadsheet, or double-click the top cell

(name cell) of any column. This will show two extra rows at the

top of the spreadsheet (2). Then click a few times on the “Type”

cell of the group column, to bring up a menu where you select

“Group” (3). Then click elsewhere to update, and you can

deselect “Column attributes” if you wish. The group column

should now be marked with a G (4).

10

The spreadsheet and the Edit menu

PAST has a spreadsheet-like user interface. Data are entered as an array of cells, organized in rows

(horizontally) and columns (vertically).

Entering data

To input data in a cell, click on the cell with the mouse and type in the data. The cells can also be

navigated using the arrow keys. Any text can be entered in the cells, but most functions will expect

numbers. Both comma (,) and decimal point (.) are accepted as decimal separators.

Absence/presence data are coded as 0 or 1, respectively. Any other positive number will be

interpreted as presence. Absence/presence-matrices can be shown with black squares for presences

by ticking the 'Square mode' box above the array.

Genetic sequence data are coded using C, A, G, T and U (lowercase also accepted).

Missing data are coded with question marks (‘?'). Unless support for missing data is specifically

stated in the documentation for a function, the function may not handle missing data correctly, so be

careful.

The convention in PAST is that items occupy rows, and variables columns. Three brachiopod

individuals might therefore occupy rows 1, 2 and 3, with their lengths and widths in columns A and B.

Cluster analysis will always cluster items, that is rows. For Q-mode analysis of associations, samples

(sites) should therefore be entered in rows, while taxa (species) are in columns. For switching

between Q-mode and R-mode, rows and columns can easily be interchanged using the Transpose

operation.

Selecting areas

Most operations in PAST are only carried out on the area of the array which you have selected

(marked). If you try to run a function which expects data, and no area has been selected, you will get

an error message.

•

A row is selected by clicking on the row label (leftmost column).

A column is selected by clicking on the column label (top row).

Multiple rows are selected by selecting the first row label, then shift-clicking (clicking with

the Shift key down) on the additional row labels.

•

Multiple columns are similarly marked by shift-clicking the additional column labels.

You can also select disjunct rows or columns by ctrl-clicking.

The whole array can be selected by clicking the upper left corner of the array (the empty grey

cell) or by choosing 'Select all' in the Edit menu.

•

Smaller areas within the array can be selected by clicking and shift-clicking.

11

Moving a row or a column

Select the ‘Drag rows/columns’ button in the ‘Click mode’ box. A row or a column can now be moved

simply by clicking on the label and dragging to the new position.

Renaming rows and columns

When PAST starts, rows are numbered from 1 to 99 and columns are labelled A to Z. For your own

reference, and for proper labelling of graphs, you should give the rows and columns more descriptive

but short names.

Select the 'Row attributes' option above the spreadsheet to see an editable column of the row

names. Select the ‘Column attributes’ option to see an editable row of the column names.

Increasing the size of the array

By default, PAST has 99 rows and 26 columns. If you should need more, you can add rows or columns

by choosing 'Insert more rows' or 'Insert more columns' in the Edit menu. Rows/columns will be

inserted after the marked area, or at the bottom/right if no area is selected. When loading large data

files, rows and/or columns are added automatically as needed.

Cut, copy, paste

The cut, copy and paste functions are found in the Edit menu. You can cut/copy data from the PAST

spreadsheet and paste into other programs, for example Word and Excel. Likewise, data from other

programs can be pasted into PAST – these need to be in a tab-separated text format.

Before pasting, select the top left cell of the spreadsheet area in Past you want to paste into. Take

care not to paste into the possibly hidden column and row attribute fields, unless you mean to.

Remove

The remove function (Edit menu) allows you to remove selected row(s) or column(s) from the

spreadsheet. The removed area is not copied to the paste buffer.

Row colors and symbols

Each row can be given a color and a symbol (dot, cross, square etc., or user-defined image). These

will be used in scatter plots and other plots. Select the ‘Row attributes’ option (or double-click the

name cell of any row) to edit the rows and colors individually, or use the ‘Row colors/symbols’

function to set all selected rows simultaneously (optionally based on the group, see below).

Selecting datatypes for columns, and specifying groups

Each column can be given a datatype using the ‘Column attributes’ mode. Select the ‘Column

attributes’ box above the spreadsheet, or double-click the name cell of any column. Then click on the

‘Type’ cell of the column a few times to bring up a small menu where the data type can be selected.

The data types are as follows:

12

Unspecified (-)

This is the default datatype.

Ordinal, nominal or binary

Specifying one of these types is only required if you wish to use mixed similarity/distance measures.

Group

In a group column, you can enter identifiers for groups of data. You can use integers, or strings such

as ‘males’ and ‘females’ (without the apostrophes). This will allow group-based polygons or ellipses

in scatter plots. A group column is also required for many analyses, such as MANOVA. It is

recommended to have rows in the same group as consecutive. Some analyses (e.g. two-way ANOVA)

require two or even more group columns.

Note that unlike the first versions of Past, there are no automatic links between colors, symbols and

groups. If you wish to use different colors and/or symbols for different groups, you can set up the

group column first and then use the ‘Row colors/symbols’ function in the Edit menu to assign

colors/symbols accordingly.

String

So far, this type is only used in the Stratigraphic chart module, for specifying names of periods, zones

etc.

Remove uninformative rows/columns

Rows or columns can be uninformative especially with respect to multivariate analyses. Such rows

and columns should be considered for removal. Several types can be searched for and removed:

Rows or columns with only zeroes, rows or columns with only missing data (‘?’), rows or columns

with only one non-zero cell (singletons), rows or columns with constant values (zero variance).

Transpose

The Transpose function, in the Edit menu, will interchange rows and columns. This is used e.g. for

switching between R mode and Q mode in cluster analysis.

Grouped columns to multivar

Converts from a format with multivariate items presented in consecutive groups of N columns to the

Past format with one item per row and all variates along the columns. For N=2, two specimens and

four variables a-d, the conversion is from

a₁b₁a₂b₂

c₁d₁c₂d₂

to

13

a₁b₁c₁d₁

a₂b₂c₂d₂

Grouped rows to multivar

Converts from a format with multivariate items presented in consecutive groups of N rows to the

Past format with one item per row and all variates along the columns. For N=2, two specimens and

four variables a-d, the conversion is from

a₁b₁

c₁d₁

a₂b₂

c₂d₂

to

a₁b₁c₁d₁

a₂b₂c₂d₂

Stack grouped rows into columns

Stacks groups horizontally along columns. This can be useful e.g. for performing univariate statistics

on pairs of columns across groups.

Rows, cols, values to table

Expects three columns of data. The first column contains categories coded as numbers or strings,

identifying rows in the output table. The second column contains categories identifying columns. The

third column contains the values to be filled into the output table.

Value pairs to matrix

Similar to “Rows, cols, values to table” but expects only two columns of numbers or strings. Each row

is one observation. Each column contains categories, e.g. Europe, Africa, Asia in the first column;

Dogs, Cats, Foxes in the second column. The occurrences of different combinations are counted,

giving a full data matrix, in this case with localities in columns and taxa in rows.

Samples to events (UA to RASC)

Given a data matrix of occurrences of taxa in a number of samples in a number of sections, as used

by the Unitary Associations module, this function will convert each section to a single row with

orders of events (FADs, LADs or both) as expected by the Ranking-Scaling module. Tied events (in the

same sample) will be given equal ranking.

14

Events to samples (RASC to UA)

Expects a data matrix with sections/wells in rows, and taxa in columns, with FAD and LAD values in

alternating columns (i.e. two columns per taxon). Converts to the Unitary Associations

presence/absence format with sections in groups of rows, samples in rows and taxa in columns.

Loading and saving data

The 'Open' function is in the File menu. You can also drag a file from the desktop onto the PAST

window. PAST uses a text file format for easy importing from other programs (e.g. Word), as follows:

The top left cell must contain a colon (:). Cells are tab-separated. There are two top rows with data

types and column names, and three left columns with colors, symbols and row names. Here is an

example:

:

-

Group

Species

0

1

Slow

4

18

10

Med

2

3

24

6

Fast

3

7

33

7

Black

Red

Dot

North

South

West

East

Red

Optional additional fields can be added to the end of the file, e.g.:

<image n>Filename Specifies an image file name to be used for the ‘Image n’ symbol, n=1 to 8.

In addition to this format, Past can also detect and open files in the following formats:

•

Excel; only the first worksheet, and only .XLS format, not .XLSX.

Nexus (see below), popular in systematics.

TPS format developed by Rohlf. The landmark, outlines, curves, id, scale and comment fields

are supported, other fields are ignored.

•

NTSYS. Multiple tables and trees are not supported. The file must have the extension ‘.nts’.

FASTA molecular sequence format, simplified specification according to NCBI.

PHYLIP molecular sequence format. The file must have the extension ‘.phy’.

Arlequin molecular sequence format. For genotype data the two haplotypes are

concatenated into one row. Not all options are supported.

•

TNT character matrix format, with some restrictions.

BioGraph format for biostratigraphy (SAMPLES or DATUM format). If a second file with the

same name but extension '.dct' is found, it will be included as a BioGraph dictionary.

RASC format for biostratigraphy. You must open the .DAT file, and the program expects

corresponding .DIC and .DEP files in the same directory.

CONOP format for biostratigraphy. You must open the .DAT file (log file), and the program

expects corresponding .EVT (event) and .SCT (section) files in the same directory.

•

If the file is not recognized, it is assumed to be a general text file with values separated by white

space, tabs or commas. The program will then ask about the format of the file.

15

Importing data from Excel

There are several ways to get data from Excel to Past.

•

Copy from Excel and paste into PAST. Make sure you click (select) the top left cell where the

data should be placed in Past before pasting. This will depend on whether row or column

attributes are included in the data.

•

Open the Excel file from PAST (only .XLS, not .XLSX).

Save as tab-separated text in Excel. The resulting text file can be opened in PAST.

Reading and writing Nexus files

The Nexus file format is used by many systematics programs. PAST can read and write the Data

(character matrix) block of the Nexus format. Interleaved data are supported. Also, if you have

performed a parsimony analysis and the 'Parsimony analysis' window is open, all shortest trees will

be written to the Nexus file for further processing in other programs (e.g. MacClade or Paup). Note

that not all Nexus options are presently supported.

Counter

A counter function is available in the Edit menu for use e.g. at the microscope when counting

microfossils of different taxa. A single row (sample) must be selected. The counter window will open

with a number of counters, one for each selected column (taxon). The counters will be initialized

with the column labels and any counts already present in the spreadsheet. When closing the counter

window, the spreadsheet values will be updated.

Count up (+) or down (-) with the mouse, or up with the keys 0-9 and a-z (only the first 36 counters).

The bars represent relative abundance. A log of events is given at the far right - scroll up and down

with mouse or arrow keys. An optional auditive feedback has a specific pitch for each counter.

16

Transform menu

These routines subject your data to mathematical operations. This can be useful for bringing out

features in your data, or as a necessary preprocessing step for some types of analysis.

Logarithm

The Log function in the Transform menu log-transforms your data using the base-10 logarithm. If the

data contain zero or negative values, it may be necessary to add a constant (e.g. 1) before log-

transforming (use Evaluate Expression x+1).

This is useful, for example, to compare your sample to a log-normal distribution or for fitting to an

exponential model. Also, abundance data with a few very dominant taxa may be log-transformed in

order to downweight those taxa.

Missing data supported.

Subtract mean

This function subtracts the column mean from each of the selected columns. The means cannot be

computed row-wise.

Missing values supported.

Remove trend

This function removes any linear trend from a data set (two columns with X-Y pairs, or one column

with Y values). This is done by subtraction of a linear regression line from the Y values. Removing the

trend can be a useful operation prior to time series analyses such as spectral analysis, auto- and

cross-correlation and ARMA.

Missing data supported.

Row percentage

All values converted to the percentage of the row sum. Missing values supported.

Row normalize length

All values divided by the Euclidean length of the row vector. This is sometimes called chord

transformation. Missing values supported.

Box-Cox

The Box-Cox transformation is a family of power transformations with the purpose of making data x

more normally distributed. The transformation has a parameter λ:

17

λ





x −1

λ ≠ 0

y =



λ





ln x

λ = 0

If the smallest input value is zero or negative (which would invalidate the transform), a constant is

added to all data such that the minimum input value becomes 1.

The default value of the parameter is calculated by maximizing the log likelihood function:

n

ˆ ²

L

(

λ

)

= − lnσ_λ+ (λ −1)

ln x

i

,

∑

2

i=1

where σ²_λis the variance of the transformed data. Alternatively, λ can be found by maximizing the

Shapiro-Wilk test statistic W for the transformed data.

This optimal value can be changed by the user, limited to the range -4 ≤ λ ≤ 4.

Missing values supported.

Compositional data transforms

Multivariate data which sum to a constant by design, such as percentages summing to 100, are called

compositional data (Aitchison 1986). Such data contain “spurious” correlations because as one value

increases, the others will have to decrease. Some multivariate analyses and tests such as PCA can be

negatively affected by this. Past includes three commonly used transforms that can be applied to

compositional data before further analysis.

The data should have the usual multivariate format, with variables in columns and items in rows. The

values in each row should sum to a constant, e.g. 1 or 100. Negative values are not allowed, and

missing data are not supported.

Additive logratio (ALR)

The input data (one row) is a vector x with N dimensions

푥₁

푥_푁

푙ꢀ

,

⋯ , 푙ꢀ

( )

alr 풙 = �

�

푥_푁

That last element is equal to zero, showing that the transformed data have dimension N-1. As the

ALR is computed with respect to the last element x_N, it may be a good idea to place a low-noise

variable with high values in the last column.

Center logratio (CLR)

푥₁

푥_푁

ln

,

… , ln

( )

clr x = �

�

( )

푔 x

( )

푔 x

18

where g(x) is the geometric mean of the data vector. This is equivalent to a simple log transform

followed by subtraction of the mean.

Isometric log ratio (ILR)

The isometric logratio transform was introduced by Egozcue et al. (2003). It has some good

theoretical properties, but the results are difficult to interpret because the transformed variables are

complicated combinations of the original variables.

Define a matrix H₀with dimensions NxN, with ones in the first row, and each subsequent row j

containing j-1 ones followed by the value –(j-1) on the diagonal followed by zeros:

1

⋮

1

−1

1

⋮

1

0

−2

1

0

−3

⋮

⋯

⋱

1

0

⋮

⎡

⎢

⎤

⎥

H₀=

⋮

⎣1

1

⋯

−(ꢁ − 1)⎦

Then, normalize each row to unit length. This gives the so-called Helmert matrix H of order N.

Remove the first row of ones, giving an (N-1) x N matrix V (we had that first row there just to make

the connection to the Helmert matrix). Finally, apply the center logratio to the data and pre-multiply

with –V to obtain the ilr:

푖푙ꢂ(풙) = −푽ꢃ푙ꢂ(풙)

The negative sign is included only to reproduce the results of the compositional package in R. Also

note that R defines the Helmert matrix as the transpose of H above.

The transformed data vector has dimensions N-1. The last column N is filled with zeros for reference,

and should not be included in further analysis.

Treatment of zero values

All three transforms involve log-transforming the original data, so zero values cannot be transformed

directly. Past treats zero values following equation (6) in Martin-Fernandez et al. (2003):

훿

if 푥_푗= 0

푧

ꢂ = �

푗

�1 − � 푥_푗if 푥_푗> 0

ꢃ

Here, δ is a small value, approximating to the lower detection limit of the measurements. This value

can be set by the user in the “Zero threshold” box (default 0.01). The z is the number of zeroes in the

original data vector, while c is the total sum, computed from the data (e.g. 100 for percentages).

References

Aitchison J. 1986. The Statistical Analysis of Compositional Data. Chapman & Hall.

19

Egozcue, J.J., Pawlowsky-Glahn, V., Mateu-Figueras, G. & Barcelo-Vidal, C. 2003. Isometric logratio

transformations for compositional data analysis. Mathematical Geology 35:279-300.

Martin-Fernandez, J.A., Barcelo-Vidal, C., Pawlowsky-Glahn, V. 2003. Dealing with zeros and missing

values in compositional data sets using nonparametric imputation. Mathematical Geology 35:253-

278

Remove size from distances

Attempts to remove the size component from a multivariate data set of measured distances

(specimens in rows, variables in columns). Three methods are available.

•

Isometric Burnaby’s method projects the set of measured distances onto a space orthogonal

to the first principal component. Burnaby's method may (or may not!) remove isometric size

from the data, for further "size-free" data analysis. Note that the implementation in PAST

does not center the data within groups - it assumes that all specimens (rows) belong to one

group.

•

Allometric Burnaby’s method will log-transform the data prior to projection, thus conceivably

removing also allometric size-dependent shape variation from the data.

Allometric vs. standard estimates allometric coefficients with respect to a standard

(reference) measurement L such as overall length (Elliott et al. 1995). This standard variable

should be placed in the first column. Each additional column is regressed onto the first

column after log-transformation, giving a slope (allometric coefficient) b for that variable. An

adjusted measurement is then computed from the original value M as



^b

L

M _adj= M 









Reference

Elliott, N.G., K. Haskard & J.A. Koslow 1995. Morphometric analysis of orange roughy (Hoplostethus

atlanticus) off the continental slope of southern Australia. Journal of Fish Biology 46:202-220.

Landmarks, Procrustes fitting

Transforms your measured point coordinates to Procrustes coordinates. There is also a menu choice

for Bookstein coordinates. Specimens go in different rows and landmarks along each row. If you have

three specimens with four landmarks in 2D, your data should look as follows:

x1 y1 x2 y2 x3 y3 x4 y4

For 3D the data will be similar, but with additional columns for z.

Landmark data in this format could be analyzed directly with the multivariate methods in PAST, but it

is recommended to standardize to Procrustes coordinates by removing position, size and rotation. A

further transformation to Procrustes residuals (approximate tangent space coordinates) is achieved

20

by selecting 'Subtract mean' in the Edit menu. You must convert to Procrustes coordinates first, then

to Procrustes residuals.

The “Rotate to major axis” option places the result into a standard orientation for convenience.

The “Keep size” option adds a final step where the shapes are scaled back to their original centroid

sizes.

The “Allow mirroring” (reflection) selection is only available for 2D. For 3D, this option is always on.

Note that the default in the procGPA function in R is “off”, while in MorphoJ it is “on”. It will usually

not make any difference, as reflection will only be optimal for data sets with very large variance.

A thorough description of Procrustes and tangent space coordinates is given by Dryden & Mardia

(1998). The algorithms for Procrustes fitting are from Rohlf & Slice (1990) (2D) and Dryden & Mardia

(1998) (3D). It should be noted that for 2D, the iterative algorithm of Rohlf & Slice (1990) often gives

slightly different results from the direct algorithm of Dryden & Mardia (1998). Past uses the former in

order to follow the “industry standard”.

Missing data is supported but only by column average substitution, which is perhaps not very

meaningful.

References

Dryden, I.L. & K.V. Mardia 1998. Statistical Shape Analysis. Wiley.

Rohlf, F.J. & Slice, D. 1990. Extensions of the Procrustes method for the optimal superimposition of

landmarks. Systematic Zoology 39:40-59.

Landmarks, Bookstein fitting

Bookstein fitting has a similar function as Procrustes fitting, but simply standardizes size, rotation and

scale by forcing the two first landmarks onto the coordinates (0,0) and (1,0). It is not in common use

today. Bookstein fitting is only implemented for 2D.

Project to tangent space (not yet in Past 4)

After Procrustes or Bookstein fitting, some statistical procedures are ideally carried out on tangent

space projected coordinates (usually it doesn’t make any difference, but don’t quote us on that!).

With d the number of dimensions and p the number of landmarks, the projection is

X = X

(

I_dp− X^t_cX_c

)

′

.

Here, X is the nxdp matrix of n specimens, X’ is the transformed matrix, I the dpxdp identity matrix,

and X_cthe mean (consensus) configuration as a dp-element row vector.

Remove size from landmarks (not yet in Past 4)

The 'Remove size from landmarks' option in the Transform menu allows you to remove size by

dividing all coordinate values by the centroid size for each specimen (Procrustes coordinates are also

normalized with respect to size).

21

See Dryden & Mardia (1998), p. 23-26.

Reference

Dryden, I.L. & K.V. Mardia 1998. Statistical Shape Analysis. Wiley.

Transform landmarks

Allows rotation of the point cloud in steps of 90 degrees, and top-bottom or left-right flipping

(mirroring), mainly for plotting convenience. The mirror operation may be useful for reducing a

bilaterally symmetric landmark data, by Procrustes fitting the left half to a mirrored version of the

right half (and optionally averaging the two).

Only for 2D coordinates.

Regular interpolation

Interpolates an irregularly sampled time series or transect (possibly multivariate) into a regular

spacing, as required by many methods for time series analysis. The x values should be in the first

selected column. These will be replaced by a regularly increasing series. All additional selected

columns will be interpolated correspondingly. The perils of interpolation should be kept in mind.

You can either specify the total number of interpolated points, or the new point spacing. Four

interpolation methods are available. The antialiasing interpolation uses a 50-point sinc (FIR) filter

with a Hamming window, low-pass filtering at half the new sampling frequency (averaged for uneven

sampling) to reduce aliasing when downsampling.

Evaluate expression

This powerful feature allows flexible mathematical operations on the selected array of data. Each

selected cell is evaluated, and the result replaces the previous contents. A mathematical expression

must be entered, which can include any of the operators +, -, *, /, ^ (power), and mod (modulo). Also

supported are brackets (), and the functions abs, atan, asin, cos, gtzer (greater than zero; 1 if x>0, 0 if

x<=0), sin, exp, ln, sqrt, sqr, round, tan, and trunc.

The following values are also defined:

•

x (the contents of the current cell)

l (the cell to the left if it exists, otherwise 0)

r (the cell to the right)

u (the cell above, or up)

d (the cell below, or down)

mean (the mean value of the current column)

median (the median of the current column)

min (the minimum value)

max (the maximum value)

n (the number of cells in the column)

i (the row index)

j (the column index)

random (uniform random number from 0 to 1)

22

•

normal (Gaussian random number with mean 0 and variance 1).

integral (running sum of the current column)

stdev (standard deviation of the current column)

iqr (interquartile range of the column)

sum (total sum of the current column)

pi

In addition, other columns can be referred to using the column name preceded by '%', for example

%A.

Examples:

sqrt(x)

Replaces all numbers with their square roots

Mean and standard deviation normalization, column-wise

Centers the values around zero

(x-mean)/stdev

x-0.5*(max+min)

(u+x+d)/3

x-u

Three-point moving average smoothing

First-order difference

Fills the column with the row numbers (requires non-empty cells, such as all

zeros)

I

Generates one period of a sine function down a column (requires non-empty

cells)

sin(2*3.14159*i/n)

5*normal+10

Random number from a normal distribution, with mean of 10 and standard

deviation of 5.

Missing values supported.

23

Plot menu

Graph

Plots one or more columns as separate graphs. You can also use a group column. It is also possible to

show each row, instead of each column, as a separate graph, with the “Plot rows” option. The x

coordinates are set automatically to 1,2,3,...

There are six plot styles available: Line, points, line with points, bars, steps and stems (vertical lines).

The 'Row labels' option sets the x axis labels to the appropriate row names.

The “Log Y” option log-transforms the values to base 10. For values <=0, the log value is set to 0.

The sequence can be smoothed with a 3-point moving average.

Missing values are disregarded.

24

XY graph

Plots one or more pairs of columns containing x/y coordinate pairs. The 'log Y' option log-transforms

your Y values (zero or negative values are set to 0). The curve can also be smoothed using 3-point

moving average.

95% concentration ellipses can be plotted in most scatter plots in PAST, such as scores for PCA, CA,

DCA, PCO and NMDS. The calculation of these ellipses assumes bivariate normal distribution. They

estimate a region where 95% of population points are expected to fall, i.e. they are not confidence

regions for the mean.

Convex hulls can also be drawn in the scatter plots, in order to show the areas occupied by points of

different groups. The convex hull is the smallest convex polygon containing all points.

The minimal spanning tree is the set of lines with minimal total length, connecting all points. In the

XY graph module, Euclidean lengths in 2D are used.

Points with missing values in X and/or Y are disregarded.

25

XY graph with error bars

As XY graph, but expects four columns (or a multiple), with x, y, x error and y error values. Symmetric

error bars are drawn around each point, with half-width as specified. If an error value is set to zero or

missing, the corresponding error bar is not drawn.

Points with missing values in X and/or Y are disregarded.

26

Histogram

Plots histograms (frequency distributions) for one or more columns. You can also use a group

column. The number of bins is by default set to an "optimal" number (the zero-stage rule of Wand

1997), with bin width

h = 3.49min(s, IQ 1.349)n^{−1 3}

where s is the sample standard deviation and IQ the interquartile range. The number of bins can be

changed by the user. When two columns are selected, they can be plotted as a bihistogram, i.e.

“mirror image” histograms which are easily compared.

The "Fit normal" option draws a graph with a fitted normal distribution (Parametric estimation, not

Least Squares).

Kernel Density Estimation is a smooth estimator of the histogram. PAST uses a Gaussian kernel with

range according to the rule given by Silverman (1986):

h = 0.9min

(

s, IQ 1.34

)

n^{−1 5}

.

Missing values are deleted.

References

Silverman, B.W. 1986. Density estimation for statistics and data analysis. Chapman & Hall.

Wand, M.P. 1997. Data-based choice of histogram bin width. American Statistician 51:59-64.

27

Bar chart/box plot

Bar plot, box plot, mean-and-whisker plot, jitter plot or violon plot for one or several columns

(samples) of univariate data. Alternatively, you can use a group column. If you include two group

columns, you get a grouped bar/box plot with the first group column specifying the group of bars,

and the second the bar within the group. There is also an option for using the first value in each

column for setting the x position of the column in the plot. Missing values are disregarded.

Bar chart

For each sample, the mean value is shown by a bar. In addition, “whiskers” can optionally be shown.

The whisker interval can represent a one-sigma or a 95% confidence interval (1.96 sigma) for the

estimate of the mean (based on the standard error), or a one-sigma or 95% concentration interval

(based on the standard deviation).

Box plot

For each sample, the 25-75 percent quartiles are drawn using a box. The median is shown with a

horizontal line inside the box. The minimal and maximal values are shown with short horizontal lines

("whiskers").

If the "Outliers" box is ticked, another box plot convention is used. The whiskers are drawn from the

top of the box up to the largest data point less than 1.5 times the box height from the box (the

"upper inner fence"), and similarly below the box. Values outside the inner fences are shown as

circles, values further than 3 times the box height from the box (the "outer fences") are shown as

stars.

The “Notches” option visualizes an approximate 95% confidence interval for the median.

The quartile methods (rounding or interpolation) are described under “Percentiles” below.

Mean and whisker plot

Similar to bar chart, but without the bar, showing the mean as a point with whiskers for the standard

error, standard deviation or 95% intervals.

Jitter plot

Each value is plotted as a dot. To show overlapping points more clearly, they can be displaced using a

random “jitter” value controlled by a slider.

Violin plot

The violin plot shows a kernel density plot (“continuous histogram”) for each sample. The plot ranges

from the minimum to the maximum value. A box plot (described above) can optionally be shown on

top of the violin.

28

Bar chart

Box plot

Violin plot with box plot

29

Pie chart

Plots a pie chart or doughnut chart from a single column of data, or several columns for multiple

charts. A sector can be emphasized by “explosion”:

With the “Coords in first rows” option, the program expects x-y coordinates in the first two rows of

each column, followed by the pie chart data. This allows a combination scatterplot/pie chart, with

multiple pies in a coordinate system.

The “Sort sizes” option will sort the slices from largest to smallest, in clockwise order.

Missing values are disregarded.

30

Stacked chart

One or more rows of data can be plotted as stacked bar chart or stacked area chart. Each bar

represents one row, and the data along columns are plotted cumulatively. The ‘Percentage’ option

converts to percentages of the row total, so that all bars will be of equal height (100%).

Missing data are treated as zero.

Stacked area chart with percentage option, axes flipped

31

Percentiles

For each percentile p, plots the value y such that p percent of the points are smaller than y. Multiple

plots can be produced using several data columns or a group column. Two popular methods are

included. For a percentile p, the rank is computed according to k=p(n+1)/100, and the value that

corresponds to that rank taken. In the rounding method, k is rounded to the nearest integer, while in

the interpolation method, non-integer ranks are handled by linear interpolation between the two

nearest ranks.

Missing values are disregarded.

32

Normal probability plot

Plots a normal probability (normal QQ) plot for one or more columns of data. A normal distribution

will plot on a straight line. For comparison, an RMA regression line is given, together with the

Probability Plot Correlation Coefficient.

(three groups were given in this example)

Missing values are disregarded.

The normal order statistic medians are computed as N(i) = G(U(i)), where G is the inverse of the

cumulative normal distribution function and U are the uniform order statistic medians:

1−U(n),

i = 1





U(i) = i − 0.3175 (n + 0.365) i = 2,3,...n −1



0.5^{1 n}

i = n





33

Ternary

Ternary plot for three columns of data, normally containing proportions of compositions. A color

map of point density (computed with a kernel density method) can also be shown. If a fourth column

is included, it will be shown using either a bubble representation or as a color/grayscale map.

Rows with missing value(s) in any column are disregarded. When using the color map option, rows

with only the fourth variable missing are included in the plot but do not contribute to the map.

34

Bubble plot

Plotting 3D data (three columns) by showing the third axis as size of disks. Negative values are not

shown. Select "Subtract min" to subtract the smallest third axis value from all values - this will force

the data to be positive. The "Size" slider scales the bubbles relative to unit radius on the x axis scale.

Rows with missing value(s) in any column are disregarded.

35

Matrix plot

Two-dimensional plot of the data matrix, using a grayscale with white for lowest value, black for

highest, or a colour scale. Includes contouring. Use to get an overview over a large data matrix.

Missing values are plotted as blanks (allowing holes and non-square boundaries).

36

Mosaic plot

Shows proportions in a two-way or three-way contingency table as areas of rectangles. A two-way

table can be given as a simple data matrix or with two group columns and a single data column (there

must be exactly one row for each combination of group levels). Column widths will reflect column

totals, and row heights will reflect row totals. The interpretation of a three-way contingency table

(specified with three group columns) is a little more complicated – see below for the standard

“Titanic” example, described on e.g. the Wikipedia page about mosaic plots.

37

Venn diagram

Plotting of Venn diagrams for two or three sets, with a number of options. Circles can be plotted with

equal sizes, or with sizes and overlaps proportional to the number of members.

Two sets

The input counts for two sets A and B can be given as a 2x2 contingency table on this form:

A and B

B, not A

A, not B

Not A, not B

The “not A, not B” value is only used when the “Show none” box is ticked.

Alternatively, the values can be given in a single column with 3 or 4 rows as follows:

A, not B

B, not A

A and B

Not A, not B (this value is optional)

Finally, you can supply a binary (0/1) matrix with any number of items in rows, and two columns A

and B corresponding to the two sets. Each item (row) can be a member of set A (value 1 in column A)

or set B (value 1 in column B), or neither (0 in both columns) or both (1 in both columns).

Three sets

The input counts for three sets A, B and C are given in a single column with 7 or 8 rows:

A, not B, not C (ABC=100)

B, not A, not C (ABC=010)

A and B, not C (ABC=110)

C, not A, not B (ABC=001)

A and C, not B (ABC=101)

B and C, not A (ABC=011)

A and B and C (ABC=111)

not A, not B, not C (ABC=000) (this value is optional)

Example data:

20

10

12

8

9

4

3

38

80

Alternatively, you can supply a binary matrix with 3 columns and any number of rows, similarly as

described above for two sets.

Plotting a fully size-proportional 3-set Venn diagram with circles is not generally possible. Past

prioritizes the total sizes of circles and the pairwise overlapping regions. The algorithm is inspired by

the “matplotlib-venn” Python code by Konstantin Tretyakov.

Four sets

Only the binary input format is available, with four columns corresponding to the four sets, and items

in rows. A full, four-set Venn diagram cannot be constructed with circles, so ellipses are used.

Proportional sizes are not attempted.

39

Radar chart

For visualizing multivariate data. One radar chart (polygon) will be plotted for each row in the data.

In the example below, the data consisted of two rows (P34 and P26) and six columns (variables). The

grid lines can be polygons (radar chart) or circles (polar chart).

The module will also accept a single column of data.

Another use of this module is for visualizing circular or cyclic data such as animal activity through 24

hours or sunlight through 12 months.

Missing values are treated as zero.

40

Polar plot

This plot accepts polar coordinates, with angles in degrees in the first column and radius values in the

second column. By default, the angles are assumed to go counterclockwise from east (see figure). By

ticking “Geographical convention” the angles are assumed to go clockwise from north.

41

Vector plot

Accepts four columns with vector start points (x and y) and vector x and y components. An optional

fifth column can specify line thickness.

There is also an option to allow the vector components (columns 3 and 4) to be specified in polar

coordinates, as angles in degrees (counter-clockwise from East) and lengths.

(In this example, the colors were specified by the row colors in the spreadsheet).

42

Network plot

This module plots networks (graphs), with nodes (rows in the spreadsheet) connected by edges. You

can specify the network with an adjacency matrix in the spreadsheet (only the lower triangle needs

to be given). In this matrix, a 1 in row i, column j implies an edge from node i to node j. All other cells

should be zero. For this type of input data, you must select “User-defined similarity” as the similarity

index. You can also specify similarities other than 1 for variable edge thickness (see below).

Alternatively, you can plot a network showing similarities between rows, using your raw data matrix

and any similarity or distance measure you choose. You can then choose a similarity cutoff (in

percent) to control the number of edges included. Zero percent cutoff will give a fully connected

(complete) network with edges between all pairs of nodes; fifty percent will show only edges

between nodes that are more than 50% similar.

The “Scale nodes by no. edges” option will set the diameter of nodes proportional to the number of

edges connected to it. The “Scale edges by similarity” will set the thickness of edges proportional to

the similarity.

The nodes can be arranged in a circle, or they can be positioned with the Fruchterman-Reingold

algorithm (1991). Using a random starting position, this algorithm will produce a new layout each

time, so click “Redraw” a few times until you get a pleasing result.

The “Bipartite linear” option plots a bipartite graph with the data rows shown as graph nodes in one

vertical column and the data columns as nodes in another column. A non-zero value in a data cell will

produce an edge from the corresponding row node to the column node. This format is typically used

for ecological data matrices with samples in rows and taxa in columns.

“Bipartite Fruchterman-Reingold” produces a similar bipartite graph but using the F-R algorithm.

Network (graph) statistics

The following statistics are calculated for the unweighted graph:

•

Number of nodes M

43

•

Number of edges N

Graph density: Number of edges divided by the maximal possible number of edges, N /

(M(M-1)/2)

•

Average number of neighbors

Graph diameter (largest eccentricity, which is the max distance from each node)

Graph radius (smallest eccentricity)

Characteristic path length (average shortest distance between any pair of nodes)

Local clustering coefficient

Reference

Fruchterman, T. M. J. & Reingold, E. M. 1991. Graph drawing by force-directed placement. Software:

Practice and Experience 21:1129–1164.

3D scatter/bubble/line plot

Requires three or four columns of data. For three columns, the data are plotted as fixed-size spheres

(or other symbols as given in the spreadsheet) with the given xyz coordinates. An optional fourth

column is shown as sizes of bubbles. The coordinate system is right-handed, with the z axis vertical

(positive up). Sticks can be added to emphasize the positions in the xy plane. Select the ‘Lines’ box to

draw lines between the points.

44

3D surface plot

Three-dimensional landscape plot of a data matrix of elevation values. Colors are assigned according

to height, and/or the surface can be gray-shaded using a lighting model. Vertical exaggeration is

adjustable. Missing values are replaced with the average. The data in the example below are the

same as for the matrix plot above.

45

3D parametric surface plot

Plots a 3D parametric surface given as a matrix of xyz triples. Such a matrix will usually be generated

by a script. Consider for example the following Past script, writing a parabolic helicoid into the

spreadsheet:

tablesize(10, 400);

for i:=1 to 100 do begin

phi:=4*2*pi*(i-1)/100;

z:=0.2*phi*phi;

for j:=1 to 8 do begin

r:=phi*(j-1)/7;

x:=r*cos(phi);

y:=r*sin(phi);

tableout(j-1, (i-1)*3, x);

tableout(j-1, (i-1)*3+1, y);

tableout(j-1, (i-1)*3+2, z);

end;

46

Univariate menu

Summary statistics

This function computes descriptive statistics for one or more samples (columns) of univariate data.

The samples can be given in one or more separate columns or with a single data column and a group

column. Each sample must have at least 3 values. The columns can have different numbers of values.

The following numbers are shown for each sample:

N:

The number of values n in the sample

The minimum value

Min:

Max:

Mean:

The maximum value

The estimate of the mean, calculated as

x

∑

i

x =

n

Std. error:

The standard error of the estimate of the mean, calculated as

s

SE_x=

n

where s is the estimate of the standard deviation (see below).

The sample variance, calculated as

Variance:

1

n −1

2

s²=

(x_i− x)

∑

.

47

Stand. dev.:

The sample standard deviation, calculated as

1

n −1

2

s =

(x_i− x)

∑

.

Median:

25 prcntil:

75 prcntil:

Mode:

The median of the sample. For n odd, the given value such that there are equally

many values above and below. For n even, the average of the two central values.

The 25^thpercentile, i.e. the given value such that 25% of the sample is below, 75%

above. The “interpolation” method is used (see Percentile plot above).

The 75^thpercentile, i.e. the given value such that 75% of the sample is below, 25%

above. The “interpolation” method is used (see Percentile plot above).

The most common value. If there is no single most common value (e.g. if all values

are different), the mode is reported as “NA” (not available)

Skewness:

The sample skewness, zero for a normal distribution, positive for a tail to the right.

Calculated as

3

(x_i− x)

n

∑

G₁=







³





(n −1)(n − 2)

1

2

(x_i− x)

∑



n −1

.

Note there are several versions of this around – Past uses the same equation as SPSS

and Excel. Slightly different results may occur using other programs, especially for

small sample sizes.

Kurtosis:

The sample kurtosis, zero for a normal distribution. Calculated as

4

2

(x_i− x)

n

(n +1)

(n −1)

∑

G₂=

− 3

⁴





(n −1)(n − 2)(n − 3)

(n − 2)(n − 3)







1

2

(x_i− x)

∑



n −1

.

Again, Past uses the same equation as SPSS and Excel.

^{1/ n}. Logarithms are used internally.

(x₁x₂x_n

)

Geom. mean: The geometric mean, calculated as

⁄∑ ⁄

1 푥_ꢄ.

Harm. mean: The harmonic mean, ꢀ

Coeff.var:

Coefficient of variation, or ratio of standard deviation to the mean, in percent:

1

n −1

2

(x_i− x)

∑

s

CV =

⋅100 =

⋅100

x

Bootstrapping

Selecting bootstrapping will compute lower and upper limits for 95% confidence intervals, using the

specified number of bootstrap replicates. Confidence intervals for the min and max values are not

48

given, because bootstrapping is known to not work well for these statistics. Three different bootstrap

methods are available (cf. Davison and Hinkley 1997):

Simple (basic):

The statistic estimated from the original sample is t. The simulated estimates from R bootstrap

*

replicates are t₁, t₂, …, t_R. For a 95% CI, we set the one-tailed error α=0.025. The simple (or basic)

bootstrapped CI is then

2t − t₍^*

₎, 2t − t₍^*

]

.

[

R+1)(1−α

R+1

)α

To ensure integer-valued subscripts, values for R such as 999, 9999 or 99999 are convenient.

Percentile:

An even simpler estimate:

t₍^*_{R+1 α}, t₍^*

]

.

[

)

R+1)(1−α

)

BCa (adjusted percentile method):

This is a complex method, but somewhat more accurate than the simple and percentile bootstrap.

Estimate a bias correction factor (called z in some texts):









{

t_r^*< t

}

w = Φ⁻¹

,









R +1

where Φ is the cumulative normal function and |⋅| is the number of elements in the set. Note we use

strictly less than, unlike some sources. Then calculate a skewness correction factor:

n

3

(

t − t

)

∑

−i

−

i=1

a =

,

3

2

n

2









6

(

t − t

)

∑

−i

−

i=1





where t_-iis the statistic computed with value i removed (jackknifed), and t₋is the mean of the

jackknifed values. With these values for w and z, compute adjusted CI endpoint values





w−1.96

a₁= Φw+





1− a

(

w−1.96

)









w+1.96

a₂= Φw+



,



1− a(w+1.96)





where 1.96 is the approximate quantile for the normal distribution corresponding to a 95% CI (the

actual value used is 1.959964). The bootstrapped confidence interval is

t₍^*_R+1, t₍^*

]

.

[

)

a₁

R+1

)

a₂

49

No interpolation is used if the index is not an integer.

Missing data: Supported by deletion.

One-sample tests

Tests for whether a single sample (single column of data) comes from a population with a given,

often hypothetical, mean or median. For example, are a number of oxygen isotope values from sea

shells (single sample) the same as average seawater composition (given mean)? The given test value

must be typed in. In addition, single-case tests are used to test whether a single value comes from

the same population as the given sample.

One-sample t test for given mean µ₀(parametric)

Sample mean and standard deviation are estimated as described above under Univariate statistics.

The 95% confidence interval for the difference in means is based on the standard error for the

estimate of the mean, and the t distribution. Normal distribution is assumed. With s the estimate of

the sample standard deviation, the confidence interval is

푠

|

� 푥

̅

− 휇₀− 푡₍

,

푥̅

− 휇₀+ 푡₍

�

⁄

훼 2,푛−1

)

⁄

훼 2,푛−1

)

ꢀ

√

ꢀ

√

Here, t has n-1 degrees of freedom, and 1-α = 0.95 for a 95% confidence interval. The t test has null

hypothesis

H₀: The sample is taken from a population with mean µ₀.

The test statistic is

푥̅

− 휇₀

푠

푡 =

ꢀ

√

One-sample Wilcoxon signed-rank test for given median M (nonparametric)

The one-sample Wilcoxon test has null hypothesis

H₀: The sample is taken from a population with median M.

All values equal to M are first removed by the program. Then the absolute values of the differences

|d_i| are ranked (R_i), with mean ranks assigned for ties. The sum of ranks for pairs where d_iis positive

is W⁺. The sum of ranks for pairs where d_iis negative is W^-. The reported test statistic is

W = max(W⁺, W^-)

(note that there are several other, equivalent versions of this test, reporting other statistics).

For large n (say n>10), the large-sample approximation to p can be used. This depends on the normal

distribution of the test statistic W:

ꢀ(ꢀ + 1)

(

)

퐸 푊 =

4

50

_ꢅ푓³− 푓

∑

ꢀ(ꢀ + 1)(2ꢀ + 1)

ꢅ

(

)

Var 푊 =

−

24

48

The last term is a correction for ties, where f_gis the number of elements in tie g. The resulting z is

reported, together with the p value.

For n<13, an exact p value is computed, by complete enumeration of all possible reassignments

(there are 2ⁿof them, e.g., 4096 for n=12). This is the preferred p value, if available.

Single-case tests

The single-case tests have null hypothesis

H₀: The given single value y is taken from the same population as the given sample.

Normal distribution is assumed. A simple z test is often used for this purpose, and is also provided by

Past. However, the z test is inaccurate because it assumes that the mean and standard deviations are

given exactly, whereas in reality, they are estimated from the sample. Therefore, Past also provides a

modified t test (Sokal & Rohlf 1995; Crawford & Howell 1998):

푦 − 푥

̅

푡 =

ꢀ + 1

�

푠

ꢀ

with s the sample standard deviation and n-1 degrees of freedom.

Binomial proportion

Expects binary data (0 or non-zero) in the given sample. The proportion of non-zeroes in the sample

is compared with the given proportion, and confidence intervals are also reported.

The same test is provided in the Single proportion module, but there a binary data column is not

required, only the observed proportion.

References

Crawford, J.R. & Howell, D.C. 1998. Comparing an individual’s test score against norms derived from

small samples. The Clinical Neuropsychologist 12:482-486.

Sokal, R.R. & Rohlf, J.F. 1995. Biometry. W.H. Freeman, San Francisco.

51

Two-sample tests

A number of classical statistics and tests for comparing two univariate samples, as given in two

columns. It is also possible to specify the two groups using a single column of values and an

additional Group column. Missing data are disregarded.

t test and related tests for equal means

Sample statistics

Means and variances are estimated as described above under Univariate statistics. The 95%

confidence interval for the mean is based on the standard error for the estimate of the mean, and

the t distribution. Normal distribution is assumed. With s the estimate of the standard deviation, the

confidence interval is



s

s 

x − t₍_{α 2,n−1}

,

x + t₍_{α 2,n−1}

)





n





.

Here, t has n-1 degrees of freedom, and 1-α = 0.95 for a 95% confidence interval.

The 95% confidence interval for the difference between the means accepts unequal sample sizes:

[

x − y − t₍_{α 2,df}₎s_D,

x − y + t₍_{α 2,df}₎s_D

]

,

where

2

SSE =

(x_i− x)

+

(y_i− y)

∑

52

df = (n₁−1)

+

(n₂−1)

MSE = SSE / df

2

n_h=

1 n₁+1 n₂

2MSE

s_D=

n_h

.

The confidence interval is computed for the larger mean minus the smaller, i.e. the center of the CI

should always be positive. The confidence interval for the difference in means is also estimated by

bootstrapping (simple bootstrap), with the given number of replicates (default 9999).

t test

The t test has null hypothesis

H₀: The two samples are taken from populations with equal means.

The t test assumes normal distributions and equal variances.

From the standard error s_Dof the difference of the means given above, the test statistic is

x − y

t =

s_D

.

Unequal variance t test

The unequal variance t test is also known as the Welch test. It can be used as an alternative to the

basic t test when equal variances cannot be assumed. Note that this version of the t test is the

default in R. The test statistic is

x − y

t =

Var(x) n₁+ Var(y) n₂

.

The number of degrees of freedom is



²

Var(x)

Var(y)

+





n₁

n₂





df =

2

[Var(x) n₁]

[Var(y) n₂

]

+

n₁−1

n₂−1

.

Monte Carlo permutation test

The permutation test for equality of means uses the absolute difference in means as test statistic.

This is equivalent to using the t statistic. The permutation test is non-parametric with few

assumptions, but the two samples are assumed to be equal in distribution if the null hypothesis is

53

true. The number of permutations can be set by the user. The power of the test is limited by the

sample size – significance at the p<0.05 level can only be achieved for n>3 in each sample.

Exact permutation test

As the Monte Carlo permutation test, but all possible permutations are computed. Only available if

the sum of the two sample sizes is less than 27.

Bayes factor

The Jeffrey-Zellner-Siow (JZS) Bayes Factor is reported in favour of the alternative, i.e. it quantifies

the evidence for the hypothesis of unequal means. It is calculated according to Rouder et al. (2009),

(

− ꢆ+1

)⁄

2

푡²

∞

⁄

(1 + ꢁ푔ꢂ²)^{−1 2}�1 +

�

2휋 ^{−1 2}푔^{−3 2}푒^{−1 2ꢅ}푑푔

(

)

∫

2

0

(

)

휈 1 + ꢁ푔ꢂ

BF₁₀

=

(

− ꢆ+1

)⁄

2

푡²

휈

�1 +

�

where N = n₁n₂/ (n₁+n₂) and ν = n₁+ n₂– 2 (degrees of freedom). The t value is the usual t statistic,

and the r value is the Cauchy scale factor, fixed at √2/2 = 0.707. The integral is approximated

numerically.

A BF value larger than 3 may be taken as “substantial” evidence for unequal means, but this is

relative to the assumed “uninformative” prior (Rouder et al. 2009). Likewise, a BF value smaller than

1/3 may be taken as substantial evidence for equal means, but this rarely happens unless N is very

large. BF>10 is reported as “strong evidence”; BF>30 as “very strong”; and BF>100 as “decisive

evidence”.

Cohen’s d

Cohen’s d (Cohen 1988) is a measure of effect size. Past calculates the classical version without bias

correction (this usually does not matter):

푥̅

− 푦�

푑 =

푠

where s is the pooled standard deviation

(

)

( )

(

)

( )

ꢀ₁− 1 Var 푥 + ꢀ₂− 1 Var 푦

�

푠 =

ꢀ₁+ ꢀ₂− 2

The textual interpretation of the value for Cohen’s d follows Sawilowsky (2009).

References

Cohen, J. 1988. Statistical power analysis for the behavioral sciences (2nd ed.). Academic Press, NY

Rouder, J.N., Speckman, P.L., Sun, D., Morey, R.D. & Iverson, G. 2009. Bayesian t tests for accepting

and rejecting the null hypothesis. Psychonomic Bulletin and Revue 16:225–237.

Sawilowsky, S. 2009. New effect size rules of thumb. Journal of Modern Applied Statistical Methods

8:467–474.

54

F test for equal variances

The F test has null hypothesis

H₀: The two samples are taken from populations with equal variance.

Normal distribution is assumed. The F statistic is the ratio of the larger variance to the smaller. The

significance is two-tailed, with n₁and n₂degrees of freedom.

Monte Carlo and exact permutation tests on the F statistic are computed as for the t test above.

Note: As usual, the permutation test assumes interchangeability. This assumption is broken if the

population means are not equal. To partly alleviate this problem, Past subtracts the sample means

before permutation. However, this hack is not ideal (Baker 1995).

Reference

Baker, R. 1995. Two permutation tests of equality of variances. Statistics and Computing 5, 289-296.

Mann-Whitney test for statistic equality

The two-tailed (Wilcoxon) Mann-Whitney U test is a non-parametric test and does not assume

normal distribution. The null hypothesis is

H₀: For randomly selected values x and y from the two populations, the probability of x>y is

equal to the probability of y>x.

If the distributions are equal except for location, the null hypothesis can be interpreted as equal

medians.

55

For each value in sample 1, count the number of values in sample 2 that are smaller than it (ties

count 0.5). The total of these counts is the test statistic U (sometimes called T). If the value of U is

smaller when reversing the order of samples, this value is chosen instead (it can be shown that

U₁+U₂=n₁n₂).

The program computes an asymptotic approximation to p based on the normal distribution (two-

tailed), which is only valid for large n. It includes a continuity correction and a correction for ties:

U − n₁n₂2 + 0.5

z =







_g



3

n n n − n −

f ³− f

∑

1

2

g





g



12n(n −1)

where n=n₁+n₂and f_gis the number of elements in tie g.

A Monte Carlo value based on the given number of random permutations (default 9999) is also given

– the purpose of this is mainly as a control on the asymptotic value.

For n₁+n₂<=26 (e.g. 13 values in each group), an exact p value is given, based on all possible group

assignments. If available, always use this exact value. For larger samples, the asymptotic

approximation is quite accurate.

Vargha-Delaney A (effect size)

The Vargha-Delaney effect size (Vargha & Delaney 2000) is an estimate of the proportion P(X > Y) +

0.5*P(X = Y). Values close to 0.5 indicate small or no difference; close to 0 indicates X smaller than Y;

close to 1 indicates X larger than Y.

56

Brunner-Munzel test for stochastic equality

The Brunner-Munzel test (Brunner & Munzel 2000; Karch 2021) has the same null hypothesis as

Mann-Whitney:

H₀: For randomly selected values x and y from the two populations, the probability of x>y is

equal to the probability of y>x.

However, the Brunner-Munzel test is less sensitive to difference in distribution and variance and is

therefore possibly superior to Mann-Whitney. The Brunner-Munzel test usually gives slightly higher p

values than Mann-Whitney. The equations used in Past were taken from Fagerland et al. (2011).

The “phat” value is an estimate of the proportion P(X < Y) + 0.5*P(X = Y). This also known as the

Vargha-Delaney A effect size (see above).

If there is no overlap between the two samples, the Brunner-Munzel statistic is not defined. Past

then sets the p value to 0.

A Monte Carlo value based on the given number of random permutations (default 9999) is also given

(Neubert & Brunner 2007)

.

References

Brunner, E., Munzel, U. 2000. The nonparametric Behrens-Fisher problem: Asymptotic theory and a

small-sample approximation". Biometrical Journal 42:17–25.

Fagerland M.W., Sandvik L., Mowinckel P. 2011, Parametric methods outperformed non-parametric

methods in comparisons of discrete numerical variables – Additional File 1: Test statistics. BMC

Medical Research Methodology 11:44.

Karch J.D. 2021. Psychologists should use Brunner-Munzel’s instead of Mann-Whitney’s U test as the

default nonparametric procedure. Advances in Methods and Practices in Psychological Science

2021;4(2).

Neubert, K., Brunner, E. 2007. A studentized permutation test for the non-parametric Behrens-Fisher

problem". Computational Statistics & Data Analysis 51:5192–5204.

Vargha, A., Delaney. H.D. 2000. A critique and improvement of the CL common language effect size

statistics of McGraw and Wong. Journal of Educational and Behavioral Statistics 25:101–252.

Mood’s median test for equal medians

The median test is an alternative to the Mann-Whitney test. It truly tests for equal medians even for

unequal distributions. The median test has low power, and the Mann-Whitney test is therefore

usually preferable. However, there may be cases with strong outliers where the Mood’s test may

perform better.

The test simply counts the number of values in each sample that are above or below the pooled

median, producing a 2x2 contingency table that is tested with a standard chi-squared test with two

degrees of freedom, and with Yate’s correction. For smaller data sets, a Fisher’s exact test is also

reported.

57

Kolmogorov-Smirnov test for equal distributions

The Kolmogorov-Smirnov test is a nonparametric test for overall equal distribution of two univariate

samples. In other words, it does not test specifically for equality of mean, variance or any other

parameter. The null hypothesis is H₀: The two samples are taken from populations with equal

distribution.

In the version of the test provided by Past, both columns must represent samples. You cannot test a

sample against a theoretical distribution (one-sample test).

The test statistic is the maximum absolute difference between the two empirical cumulative

distribution functions:

D = max S_N

(

x

)− S_N

(

x

)

1

2

x

The algorithm is based on Press et al. (1992), with significance estimated after Stephens (1970).

Define the function

∞

2

Q_KS

(

λ

)

= 2

(−1)^j−1e^{−2 j λ}

∑

j=1

.

With N_e= N₁N₂/(N₁+N₂), the significance is computed as

p = Q_KS

(

[

N_e+ 0.12 + 0.11

N_e

]

D

)

.

The permutation test uses 10,000 permutations. Use the permutation p value for N<30 (or generally).

References

58

Press, W.H., Teukolsky, S.A., Vetterling, W.T. & Flannery, B.P. 1992. Numerical Recipes in C. 2^nd

edition. Cambridge University Press.

Stephens, M.A. 1970. Use of the Kolmogorov-Smirnov, Cramer-von Mises and related statistics

without extensive tables. Journal of the Royal Statistical Society, Series B 32:115-122.

Anderson-Darling test for equal distributions

The Anderson-Darling test is a nonparametric test for overall equal distribution of two univariate

samples. It generally has higher power than the Kolmogorov-Smirnov test.

With two samples x₁.. x_nand y₁.. y_m, the pooled sample size is N = n+m. The test statistic A²_Ncan be

computed according to Pettitt (1976):

푁−1

2

(

)

1

푀_ꢄꢁ − ꢀ푖

퐴²_푁

=

�

(

)

푚ꢀ

푖 ꢁ − 푖

ꢄ=1

where M_iis the number of x’s less than or equal to the ith smallest in the pooled sample. Past uses a

slightly more complicated version of this equation, with better performance in the presence of ties

(Scholz & Stephens 1987, eq. 6).

This statistic is transformed to a statistic called Z according to Scholz & Stephens (1987). For our case

with k=2 samples, compute the variance of the statistic as follows:

1

퐻 =

+

푚

ꢀ

^푁−11

ℎ = �

푖

ꢄ=1

푁−2 푁−1

1

푔 = � �

(

)

ꢁ − 푖 ꢇ

ꢄ=1 푗=ꢄ+1

(

)

(

)

(

)

푎 = 2 4푔 − 6 + 10 − 6푔 퐻 − 4푔 + 6 = 4푔 − 6 + 10 − 6푔 퐻

(

)

(

)

(

)

푏 = 4 2푔 − 4 + 16ℎ + 2푔 − 14ℎ − 4 퐻 − 8ℎ + 4푔 − 6 = 12푔 + 8ℎ − 22 + 2푔 − 14ℎ − 4 퐻

(

)

(

)

(

)

(

)

ꢃ = 4 6ℎ + 2푔 − 2 + 2 4ℎ − 4푔 + 6 + 2ℎ − 6 퐻 + 4ℎ = 36ℎ + 4 + 2ℎ − 6 퐻

(

)

푑 = 4 2ℎ + 6 − 8ℎ = 6

푎ꢁ³+ 푏ꢁ²+ ꢃꢁ + 푑

휎_푁²=

(ꢁ − 1)(ꢁ − 2)(ꢁ − 3)

Then,

퐴²_푁− 1

푍 =

휌_푁

The p value is computed by interpolation and extrapolation in Table 1 (m=1) of Scholz & Stephens

(1987), using a curve fit to the von Bertalanffy model. The approximation is fairly accurate for

59

p≤0.25.For p>0.25, the p values are estimated using a polynomial fit to values obtained by

permutation. A p value based on Monte Carlo permutation with N=999 is also provided.

References

Pettitt, A.N. 1976. A two-sample Anderson-Darling rank statistic. Biometrika 63:161-168.

Scholz, F.W. & Stephens, M.A. 1987. K-sample Anderson–Darling tests. Journal of the American

Statistical Association 82:918–924.

Epps-Singleton test for equal distributions

The Epps-Singleton test (Epps & Singleton 1986; Goerg & Kaiser 2009) is a nonparametric test for

overall equal distribution of two univariate samples. It is typically more powerful than the

Kolmogorov-Smirnov test, and unlike the Kolmogorov-Smirnov it can be used also for non-continuous

(i.e. ordinal) data. The null hypothesis is H₀: The two samples are taken from populations with equal

distribution.

The mathematics behind the Epps-Singleton test are complicated. The test is based on the Fourier

transform of the empirical distribution function, called the empirical characteristic function (ECF).

The ECF is generated for each sample and sampled at two points (t₁=0.4 and t₂=0.8, standardized for

the pooled semi-interquartile range). The test statistic W₂is based on the difference between the

two sampled ECFs, standardized by their covariance matrices. A small-sample correction to W₂is

applied if both sample sizes are less than 25. The p value is based on the chi-squared distribution. For

details, see Epps & Singleton (1986) and Goerg & Kaiser (2009).

References

Epps, T.W. & Singleton, K.J. 1986. An omnibus test for the two-sample problem using the empirical

characteristic function. Journal of Statistical Computation and Simulation 26:177–203.

Goerg, S.J. & Kaiser, J. 2009. Nonparametric testing of distributions – the Epps-Singleton two-sample

test using the empirical characteristic function. The Stata Journal 9:454-465.

60

Coefficient of variation (Fligner-Kileen test)

This module tests for equal coefficient of variation in two samples.

The coefficient of variation (or relative variation) is defined as the ratio of standard deviation to the

mean in percent, and is computed as:

1

n −1

2

(x_i− x)

∑

s

CV =

⋅100 =

⋅100

x

.

The 95% confidence intervals are estimated by bootstrapping (simple bootstrap), with the given

number of replicates (default 9999).

The null hypothesis if the statistical test is:

H₀: The samples were taken from populations with the same coefficient of variation.

If the given p(normal) is less than 0.05, equal coefficient of variation can be rejected. Donnelly &

Kramer (1999) describe the coefficient of variation and review a number of statistical tests for the

comparison of two samples. They recommend the Fligner-Killeen test (Fligner & Killeen 1976), as

implemented in Past. This test is both powerful and is relatively insensitive to distribution. The

following statistics are reported:

T:

The Fligner-Killeen test statistic, which is a sum of transformed ranked positions of the

smaller sample within the pooled sample (see Donnelly & Kramer 1999 for details).

E(T):

The expected value for T.

z:

The z statistic, based on T, Var(T) and E(T). Note this is a large-sample approximation.

61

p:

The p(H₀) value. Both the one-tailed and two-tailed values are given. For the alternative

hypothesis of difference in either direction, the two-tailed value should be used. However,

the Fligner-Killeen test has been used to compare variation within a sample of fossils with

variation within a closely related modern species, to test for multiple fossil species (Donnelly

& Kramer 1999). In this case the alternative hypothesis might be that CV is larger in the fossil

population, if so then a one-tailed test can be used for increased power.

The screenshot above reproduces the example of Donnelly & Kramer (1999), showing that the

relative variation within Australopithecus afarensis is significantly larger than in Gorilla gorilla. This

could indicate that A. afarensis represents several species.

References

Donnelly, S.M. & Kramer, A. 1999. Testing for multiple species in fossil samples: An evaluation and

comparison of tests for equal relative variation. American Journal of Physical Anthropology 108:507-

529.

Fligner, M.A. & Killeen, T.J. 1976. Distribution-free two sample tests for scale. Journal of the American

Statistical Association 71:210-213.

F and t tests from parameters

Sometimes publications give not the data, but values for sample size, mean and variance for two

samples. These can be entered manually using the ‘F and t from parameters’ option in the menu. This

module does not use any data from the spreadsheet.

62

Two-sample paired tests (t, sign, Wilcoxon)

Three statistical tests (one parametric, two non-parametric) for two samples (columns) of univariate

data. The data points are paired, meaning that the two values in each row are associated. For

example, the test could be the for length of the left vs. the right arm in a number of people, or the

diversity in summer vs. winter at a number of sites. Controlling for a “nuisance factor” (person, site)

in this way increases the power of the test. The null hypothesis is:

H₀: The mean (t test) or median (sign test, Wilcoxon test) of the difference is zero.

All reported p values are two-tailed.

t test

Testing for mean difference equal to zero using the standard one-sample t test on the differences.

With d_i= x_i-y_i, we have

2

1

s =

(d − d

)

∑

i

n −1

,

d

t =

s

n

.

There are n-1 degrees of freedom. This test assumes normal distribution of the differences.

The exact version of the test calculates all possible group reassignments within pairs. It is only

computed for less than 26 pairs.

63

Bayesian factor

The Jeffrey-Zellner-Siow (JZS) Bayes Factor is reported in favour of the alternative, i.e. it quantifies

the evidence for the hypothesis of unequal means. For details, see the section on the (unpaired) two-

sample t test above. The same equation is used, but with N-1 degrees of freedom.

Sign test

The sign (binomial) test simply counts the number of cases n₁where x_i>y_iand n₂where y_i>x_i. The

number max(n₁, n₂) is reported. The p value is exact, computed from the binomial distribution. The

sign test may have lower power than the other paired tests, but makes few assumptions.

Wilcoxon signed rank test

A non-parametric rank test that does not assume normal distribution. The null hypothesis is no

median shift (no difference).

All rows with zero difference are first removed by the program. Then the absolute values of the

differences |d_i| are ranked (R_i), with mean ranks assigned for ties. The sum of ranks for pairs where

d_iis positive is W⁺. The sum of ranks for pairs where d_iis negative is W^-. The reported test statistic is

W = min(W⁺, W^-)

(note that there are several other, equivalent versions of this test, reporting other statistics).

For large n (say n>10), the large-sample approximation to p can be used. This depends on the normal

distribution of the test statistic W:

n(n +1)

E(W ) =

4

f ³− f

∑

g

n

(n +1)(2n +1)

g

Var(W

)

=

−

24

48

.

The last term is a correction for ties, where f_gis the number of elements in tie g. The resulting z is

reported, together with the p value.

The Monte Carlo significance value is based on 99,999 random reassignments of values to columns,

within each pair. This value will be practically identical to the exact p value.

For n<23, an exact p value is computed, by complete enumeration of all possible reassignments

(there are 2ⁿof them, i.e. more than four million for n=22). This is the preferred p value, if available.

Missing data: Supported by deletion of the row.

64

Several-sample tests

One-way ANOVA and Kruskal-Wallis tests for equality of means or medians between several

univariate samples, given in separate columns. It is also possible to specify the groups using a single

column of values and an additional Group column. Missing data are supported by deletion.

One-way ANOVA

One-way ANOVA (analysis of variance) is a statistical procedure for testing the null hypothesis that

several univariate samples are taken from populations with the same mean. The samples are

assumed to be close to normally distributed and have similar variances. If the sample sizes are equal,

these two assumptions are not critical. If the assumptions are strongly violated, the nonparametric

Kruskal-Wallis test should be used instead.

ANOVA table

The between-groups sum of squares is given by:

2

(

)

SS_bg

=

n x − x

∑

g

T

g

,

where n_gis the size of group g, and the means are group and total means. The between-groups sum

of squares has an associated df_bg, the number of groups minus one.

The within-groups sum of squares is

65

2

(

)

SS_wg

=

x_i− x_g

∑∑

g

i

where the x_iare those in group g. The within-groups sum of square has an associated df_wg, the total

number of values minus the number of groups.

The mean squares between and within groups are given by

SS_bg

MS_bg=

df_bg

SS_wg

MS_wg

=

df_wg

Finally, the test statistic F is computed as

MS_bg

F =

MS_wg

The p value is based on F with df_bgand df_wgdegrees of freedom.

Random effects (Model II) ANOVA

For balanced, one-way ANOVA, the ANOVA table, F value and p value are the same for fixed-effects

and random-effects ANOVA, so you can use the results for either type of ANOVA. If your design is of

the random-effects type (i.e. the factor levels are taken at random from a larger population instead

of being fixed by experiment), then you should also report the given variance due to random errors

and the variance due to difference between groups (can become negative):

푠²= MS_wg

MS_bg− MS_wg

푠_ꢅ²=

ꢀ

where n is the sample size for each group in the case of balanced design. For unbalanced design, set

n to

ꢈ

2

∑

1

ꢀ_ꢄ

ꢀ₀=

�� ꢀ_ꢄ−

�

퐺 − 1

ꢄ=1

where G is the number of groups. The intraclass correlation coefficient ICC gives the proportion of

variance due to group differences:

푠_ꢅ²

퐼ꢉꢉ =

푠_ꢅ²+ 푠²

66

Omega squared

The omega squared is a measure of effect size, varying from 0 to 1:

SS_bg− df_bgMS_wg

ω²=

.

SS_total+ MS_wg

Levene's test

Levene's test for homogeneity of variance (homoskedasticity), that is, whether variances are equal as

assumed by ANOVA, is also given. Two versions of the test are included. The original Levene's test is

based on means. This version has more power if the distributions are normal or at least symmetric.

The version based on medians has less power but is more robust to non-normal distributions. Note

that this test can be used also for only two samples, giving an alternative to the F test for two

samples described above.

Unequal-variance (Welch) ANOVA

If Levene's test is significant, meaning that you have unequal variances, you can use the unequal-

variance (Welch) version of ANOVA, with the F, df and p values given.

Bayes factor

The Jeffrey-Zellner-Siow (JZS) Bayes Factor is reported in favour of the alternative, i.e. it quantifies

the evidence for the hypothesis of unequal means. It is calculated according to Rouder et al. (2012).

Effect size (confidence intervals for the pairwise differences in means)

This table gives 95% confidence intervals for the differences in group means. If a confidence interval

does not include zero, the difference may be considered significant, and it is marked in pink. These

are “multiple-t” or “simultaneous” confidence intervals, which are wider than the confidence

intervals for each pairwise difference in isolation. For the two groups i and j, the confidence interval

is

1

�푥� − 푥�� ± 푡_{훼 2ꢊ}푀ꢋ

�

푤ꢅ

+

�

⁄

횤

횥

ꢀ_ꢄ

ꢀ_푗

where t_α/2mis the upper α/2m point of the t distribution with α=0.05, with m the number of

confidence statements (total number of pairwise comparisons). The degrees of freedom is equal to

the total number of values minus the number of groups.

67

Analysis of residuals

The “Residuals” tab shows properties of the residuals, in order to evaluate some assumptions of

ANOVA such as normal and homoskedastic distribution of residuals.

The Shapiro-Wilk test for normal distribution is given, together with several common plots of

residuals (normal probability, residuals vs. group means, and histogram).

Tukey’s pairwise post-hoc tests

If the ANOVA shows significant inequality of the means (small p), you can go on to study the given

table of "post-hoc" pairwise comparisons, based on the Tukey-Kramer test. The Studentized Range

Statistic Q is given in the lower left triangle of the array, and the probabilities p(equal) in the upper

right.

X _L− X _S

,

Q =

MS_wg

n

where X_Lis the larger and X_Sthe smaller mean of the two samples being compared. If the sample

sizes are not equal, their harmonic mean is used for n. Significance is estimated according to

Copenhaver & Holland (1988), with df_wgdegrees of freedom. There is also an option to use the

slightly less accurate method of Lund & Lund (1983) which was used in older versions of Past.

68

Kruskal-Wallis

The Kruskal-Wallis test is a non-parametric ANOVA, comparing the medians of several univariate

groups (given in columns). It can also be regarded as a multiple-group extension of the Mann-

Whitney test (Zar 1996). It does not assume normal distribution, but does assume equal-shaped

distribution for all groups. The null hypothesis is

H₀: The samples are taken from populations with equal medians.

The test statistic H is computed as follows:

T_g²

n_g













12

+1)

H

=

− 3(n

+1)

∑

n

(

n

g

,

where n_gis the number of elements in group g, n is the total number of elements, and T_gis the sum

of ranks in group g.

The test statistic H_cis adjusted for ties:

H

H_c=

f ³− f

∑

i

1−

n³− n

,

where f_iis the number of elements in tie i.

69

With G the number of groups, the p value is approximated from H_cusing the chi-square distribution

with G-1 degrees of freedom. This is less accurate if any n_g<5.

Mann-Whitney pairwise post-hoc tests

Mann-Whitney pairwise test p values are given for all N_p=G(G-1)/2 pairs of groups. The asymptotic

approximation described under the Mann-Whitney module is used. If samples are very small, it may

be useful to run the exact test available in that module instead. Four different views are available for

the symmetric table:

1. Raw p values, uncorrected significance: The p values from each individual pairwise test,

marked in pink if p<0.05, not corrected for multiple testing.

2. Raw p values, sequential Bonferroni significance: The p values from each individual pairwise

test are shown uncorrected for multiple testing. Significance (pink marking) is assessed by

first evaluating the smallest p value, with Bonferroni correction for N_ppairs. If significant

(pN_p<0.05) the next smallest p value is significant if p(N_p-1)<0.05, etc.

3. Bonferroni corrected p values: The values shown are p’ = pN_p. Marked as significant if

p’<0.05.

4. Mann-Whitney U: The test statistics.

Dunn’s post-hoc

The Dunn’s post hoc test (Dunn 1964) is a pairwise test often carried out after a significant Kruskal-

Wallis test. It is an alternative to the pairwise Mann-Whitney. With T_gthe sum of ranks within group

g from the Kruskal-Wallis test, calculate for each group the average rank:

푇

ꢅ

�

푇 =

ꢅ

ꢀ_ꢅ

To compare two groups A and B, calculate the z statistic

�

|

푇 − 푇_퐵

ꢌ

푧_AB

=

휌_AB

_ꢄ푓³− 푓

1

∑

ꢀ(ꢀ + 1)

^ꢄ� �

+

�

ꢄ

�

휌_AB

=

�

−

12

12(ꢀ − 1)

ꢀ_ꢌ

ꢀ_퐵

Here, n is the total sample size and f_ithe number of elements in tie i, as in the Kruskal-Wallis test. See

pairwise Mann-Whitney above for the various options in the table output (raw p values, raw p values

with sequential Bonferroni, Bonferroni corrected p values, and the z statistics).

References

Copenhaver, M.D., Holland, B. 1988. Computation of the distribution of the maximum studentized

range statistic with application to multiple significance testing of simple effects. Journal of Statistical

Computation and Simulation 30:1-15.

70

Dunn, O.J. 1964. Multiple comparisons using rank sums. Technometrics 6:241-252.

Lund, R.E., Lund, J.R. 1983. Algorithm AS 190: Probabilities and upper quantiles for the studentized

range. Journal of the Royal Statistical Society C 32:204-210.

Rouder, J.N, Morey, R.D., Speckman, P.L., Province, J.M. 2012. Default Bayes factors for ANOVA

designs. Journal of Mathematical Psychology 56:356-374.

Zar, J.H. 1996. Biostatistical analysis. 3^rded. Prentice Hall.

71

Several-samples repeated measures tests

In repeated measures ANOVA, values in each row are observations on the same “subject”. Repeated-

measures ANOVA is the extension of the paired t test to several samples. Each column (sample) must

contain the same number of values.

Missing values are not supported.

The procedure begins like the independent-samples one-way ANOVA above. In short,

2

,

(

)

SS =

n x − x

∑

bg

g

T

g

where n is the sample size. The associated df_bgis the number of groups minus one.

2

(

)

SS_wg

=

x_i− x_g

∑∑

g

i

where the x_iare those in group g. The associated df_wgis the total number of values minus the

number of groups.

The between-subjects sum of squares is

2

,

SS_sub

=

n

(x_i− x_T

)

∑

i

x_i

where the

minus one.

are means of subject i across groups. The associated df_subis the number of subjects

72

The SS_erroris simply SS_wg– SS_sub, with df_error= df_wg– df_sub

.

The mean squares are then the sum squares divided by their respective degrees of freedom:

SS_bg

MS_bg=

df_bg

SS_wg

MS_wg

MS_sub

=

df_wg

SS_sub

df_sub

SS_error

df_error

.

MS_error

=

Finally, the F ratio is MS_bg/ MS_error, with df_bgand df_errordegrees of freedom.

Sphericity estimates and corrections

An assumption of repeated measures ANOVA is sphericity, meaning equal variances of the

differences between all combinations of groups. A statistic called epsilon approaches 1 for data

meeting the sphericity assumption. For smaller values of epsilon, a correction can be applied to the

degrees of freedom of the F test, providing a corrected p value for the ANOVA. PAST provides two

versions of this procedure, Greenhouse-Geisser (Greenhouse & Geisser 1959) and Huynh-Feldt

(Huynh & Feldt 1976).

Tukey’s pairwise post-hoc tests

The "post-hoc" pairwise comparisons are based on the Tukey test. The Studentized Range Statistic Q

is given in the lower left triangle of the array, and the probabilities p(equal) in the upper right.

X _L− X _S

Q =

,

MS_error

n

where X_Lis the larger and X_Sthe smaller mean of the two samples being compared. There are df_error

degrees of freedom.

73

Friedman test

The Friedman test is a non-parametric test for equality of medians in several repeated-measures

univariate groups. It can be regarded as the non-parametric version of repeated-measures ANOVA,

or the repeated-measures version of the Kruskal-Wallis test.

The Friedman test follows Bortz et al. (2000). The basic test statistic is

k

12

2

χ ²=

T − 3n(k +1)

,

∑

j

nk(k +1)

j=1

where n is the number of rows, k the number of columns and T_jthe column sums of the data table.

The χ²value is then corrected for ties (if any):

χ ²

χ_t²_ie

=

m

1

1−

(

t³− t

)

∑

i

nk

(

k ²−1

)

i=1

where m is the total number of tie groups and t_iare the numbers of values in each tie group.

For k=2, it is recommended to use one of the paired tests (e.g. sign or Wilcoxon test) instead. For

small data sets where k=3 and n<10, or k=4 and n<8, the tie-corrected χ²value is looked up in a table

of “exact” p values. When given, this is the preferred p value.

74

The asymptotic p value (using the χ²distribution with k-1 degrees of freedom) is fairly accurate for

larger data sets. It is computed from a continuity corrected version of χ²:

2

k

n

(k +1)









S =

T −

∑

j

2

_j=1



12n(k −1)(S −1)

χ ²=

.

n²

(

k³− k

)

+ 24

This χ²value is also corrected for ties using the equation above.

The post hoc tests are by simple pairwise Wilcoxon, exact for n<20, asymptotic for n>=20. These tests

have higher power than the Friedman test.

References

Bortz, J., Lienert, G.A. & Boehnke, K. 2000. Verteilungsfreie Methoden in der Biostatistik. 2nd ed.

Springer.

Greenhouse, S.W. & Geisser, S. 1959. On methods in the analysis of profile data. Psychometrika

24:95-112.

Huynh, H. & Feldt, L.S. 1976. Estimation of the Box correction for degrees of freedom from sample

data in randomized block and split-plot designs. Journal of Educational Statistics 1:69-82.

75

Two-way ANOVA

Two-way ANOVA (analysis of variance) tests the null hypotheses that several univariate samples have

the same mean across each of two factors A and B, and that there are no dependencies (interactions)

between factors. The samples are assumed to be close to normally distributed and have similar

variances. If the sample sizes are equal, these two assumptions are not critical. The default analysis is

a fixed-effect blocked design (the most usual case). There is also an option for random-effect blocked

two-way ANOVA, and fixed-effect and random-effect, nested two-way ANOVA. There is no

interaction term for the nested design.

Past supports unbalanced designs (unequal sample sizes for each factor combination). In this case, a

“Model II” ANOVA is carried out (Fox 2016).

Three columns are needed: A group column (set data type to Group with ‘Column attributes’) with

the levels for factor A, a group column with the levels for factor B, and a column of the corresponding

measured values.

The algorithm uses weighted means.

Total sum of squares:

2

,

SS =

(x_i− x)

∑

T

i

taken over all points. The associated degrees of freedom df_Tis the total number of values minus one.

76

Within-group (error) sum of squares:

2

,

SS_wg

=

(x_i− x_{g g}

)

∑∑∑

1

2

g₁g₂

i

where the x_iare those in group (level) g₁for the first factor and g₂for the second factor, and the

mean is taken within the same group combination. The associated df_wgis the total number of values

minus the product of numbers of groups and columns.

The between-groups sum of squares SS_bg= SS_T– SS_wgcan be partitioned into three, namely the

Factor A, Factor B and interaction terms.

2

,

SS_A= N_A

(x_i− x)

∑

i

where the sum is over levels of Factor A, and the two means are the level mean and the total mean,

respectively. N_Ais the number of levels of A. The degrees of freedom is df_A= N_A– 1. Similarly, for

Factor B:

2

(

)

SS_B= N_B

x_j− x

∑

j

where now the sum is over levels of Factor B. The degrees of freedom is df_B= N_B– 1.

The interaction sum of squares is SS_AxB= SS_bg– SS_A– SS_B, with df_AxB= (N_A- 1)(N_B- 1) degrees of

freedom.

For unbalanced designs (Model II), the computations are more complex, using dummy-variable

multiple regressions (Fox 2016).

Mean squares MS are the sum of squares divided by their respective degrees of freedom.

Finally, the F ratios are

F_A= MS_A/ MS_wg

F_B= MS_B/ MS_wg

F_AxB= MS_AxB/ MS_wg

Random effects

The random-effects ANOVA is computed as fixed effects ANOVA, except

F_A= MS_A/ MS_AxB

F_B= MS_B/ MS_AxB

.

For random-effects ANOVA, the components of variance are computed as follows (only for balanced

design, where n is the sample size for each level combination):

var(A) = (MS_A– var(err) – n var(AxB)) / (nN_B)

var(B) = (MS_B– var(err) – n var(AxB)) / (nN_A)

var(AxB) = (MS_AxB– var(err) ) / n

77

var(err) = MS_wg

Nested two-way ANOVA

The main factor (factor A) sum-of-squares SS_A, degrees of freedom df_Aand mean square MS_Aare

calculated as above. For the fixed-effects model, F_Ais calculated as above, while for random-effects it

is F_A=MS_A/MS_B. The nested factor (factor B) is always random-effect.

Graph of means

The graph of means is a simple graphical device, traditionally used to see the effects of factors and

their interaction for a two-way ANOVA. The means are shown with either the A levels or the B levels

on the x axis, and the other levels as separate lines:

Tukey’s post-hoc tests

Tukey’s post-hoc tests are available for the two main factors, and for all unconfounded interactions.

For the main factors, the Studentized Range Statistic Q is given in the lower left triangle of the array,

and the probabilities p(equal) in the upper right. The interaction test uses the “adjusted k” value,

accounting for the number of unconfounded comparisons, for up to 7 levels in each factor. For larger

number of levels, the k value is not adjusted.

Missing values : Rows with missing values are deleted.

Reference

Fox, J. 2016. Applied Regression Analysis and Generalized Linear Models, 3^rded. Sage Publications.

78

Two-way ANOVA without replication

Two-way ANOVA for testing the null hypotheses that several univariate samples have the same mean

across each of two factors. This module expects only one observation for each combination of levels

for the two factors. The input data format is a table where the first factor levels enter in rows, and

the second factor level in columns, e.g. a table of veterinary lab results:

There is no interaction term.

The equations are given by Ireland (2010), pp. 130-131.

Reference

Ireland, C.R. 2010. Experimental Statistics for Agriculture and Horticulture. CABI, 352 pp.

79

Two-way repeated measures ANOVA

Three data columns are needed: A group column (set data type to Group with ‘Column attributes’)

with the levels for factor A, a group column with the levels for factor B, a group column with

identifiers for the cases (subjects) and a column of the corresponding measured values.

Each subject must have exactly one entry for each combination of levels. Therefore, if you have M

levels for factor A, N levels for factor B and S subjects, you need exactly MxNxS rows in total.

Other functionality is similar to the two-way ANOVA module described above. Missing data are not

supported!

Example data formatted for two-way repeated measures ANOVA.

80

One-way ANCOVA

ANCOVA (Analysis of Covariance) tests for equality of means for several univariate groups, adjusted

for covariance with another variate. ANCOVA can be compared with ANOVA, but has the added

feature that for each group, variance that can be explained by a specified "nuisance" covariate (x) is

removed. This adjustment can increase the power of the test substantially.

The program expects two or more pairs of columns, where each pair (group) is a set of correlated x-y

data (means are compared for y, while x is the covariate). The example below uses three pairs

(groups) a, b and c.

The Plot tab presents a scatter plot and linear regression lines for all the groups. The ANOVA-like

summary table contains sum-of-squares etc. for the adjusted means (between-groups effect) and

adjusted error (within-groups), together with an F test for the adjusted means. An F test for the

equality of regression slopes (as assumed by the ANCOVA) is also given. In the example, equal

adjusted means in the three groups can be rejected at p<0.05. Equality of slopes can not be rejected

(p=0.74).

The Groups tab gives the summary statistics for each group (mean, adjusted mean and regression

slope).

Assumptions include similar linear regression slopes for all groups, normal distributions, similar

variance and sample sizes.

Missing data: x-y pairs with either x or y missing are disregarded.

81

Correlation table

Two or more columns are required. A matrix is presented with the correlations between all pairs of

columns. In the ‘Statistic \ p(uncorr)’ table format, correlation values are given in the lower triangle

of the matrix, and the two-tailed probabilities that the columns are uncorrelated are given in the

upper. Both parametric and non-parametric coefficients and tests are available.

Missing data: Supported by pairwise deletion, except for partial correlation which uses mean value

imputation.

Linear r (Pearson)

Pearson’s r is the most commonly used parametric correlation coefficient. It is given by

(x_i− x)(y_i− y)

∑

i

.

r =

2

(x_i− x)

(y_i− y)

∑

i

The significance is computed using a two-tailed t test with n-2 degrees of freedom and

n − 2

.

t = r

1− r²

Bayes factor

A Bayes factor for Pearson’s r is computed according to “Jeffrey’s integrated Bayes factor”, eq. (30) in

Ly et al. (2016):

82

ꢀ + 1

2

ꢀ + 2

Γ �

�

휋

2ꢀ − 3 2ꢀ − 1 ꢀ + 2

2ꢎ �

; ꢂ²�

ꢍꢎ^푗,ꢄ(ꢀ, ꢂ) =

;

1

10

2

4

2

Γ �

2

where ₂F₁is the Gaussian hypergeometric function (evaluated as a Taylor series; Section 4.2 in

Pearson et al. 2016). Large values of the Bayes factor (>3) indicate evidence for correlation (positive

or negative).

Spearman’s D and rs

Spearman’s (non-parametric) rank-order correlation coefficient is the linear correlation coefficient

(Pearson’s r) of the ranks. Following Press et al. (1992) it is computed as

6



1



1−

D +

(

f_k³− f_k

)

+

(

g_m³− g_m

)

∑





n³− n

12

k

m





r_s=

.

f_k³− f_k

(

g_m³− g_m

)





)



(

∑







k

m

1−





n³− n













Here, D is the sum squared difference of ranks (midranks for ties):

n

2

D =

(R_i− S_i

)

.

∑

i=1

The f_kare the numbers of ties in the kth group of ties among the R_i’s, and the g_mare the numbers of

ties in the mth group of ties among the S_i’s.

For n>9, the probability of non-zero r_s(two-tailed) is computed using a t test with n-2 degrees of

freedom:

n − 2

t = r_s

.

1− r_s²

For small n this approximation is inaccurate, and for n<=9 the program therefore switches

automatically to an exact test. This test compares the observed r_sto the values obtained from all

possible permutations of the first column.

Kendall’s tau

This non-parametric correlation coefficient is not in very common use. It is computed according to

Press et al. (1992). All possible N(N-1)/2 pairs of bivariate data points are considered. If two pairs

have the same direction in x as in y (x and y both decrease, or both increase), they are called

concordant. If not, they are discordant. A tie in the x’s is called an extra-x, and a tie in the y’s is called

an extra-y. Pairs with ties in both variables are discarded. The number of pairs in the four categories

are counted. Then,

83

concordant − discordant

τ =

.

concordant + discordant + extray concordant + discordant + extrax

The asymptotic test is based on Kendall’s tau being approximately normal, with zero mean and

4N +10

var(τ

)

=

.

9N

(

N −1

)

Polyserial correlation

This correlation is only carried out if the second column consists of integers with a range less than

100. It is designed for correlating a normally distributed continuous/interval variable (first column)

with an ordinal variable (second column) that bins a normally distributed variable. For example, the

second column could contain the numbers 1-3 coding for “small”, “medium” and “large”. There

would typically be more “medium” than “small” or “large” values because of the underlying normal

distribution of sizes.

Past uses the two-step algorithm of Olsson et al. (1982). This is more accurate than their “ad hoc”

estimator, and nearly as accurate as the full multivariate ML algorithm. The two-step algorithm was

chosen because of speed, allowing a permutation test (but only for N<100; not yet in Past 3). For

larger N the given asymptotic test (log-ratio test) is accurate.

Partial linear correlation

Using this option, for each pair of columns, the linear correlation is computed while controlling for all

the remaining columns. For example, with three columns A, B, C the correlation AB is controlled for

C; AC is controlled for B; BC is controlled for A. The partial linear correlation can be defined as the

correlation of the residuals after regression on the controlling variable(s). The significance is

estimated with a t test with n-2-k degrees of freedom, where k is the number of controlling variables:

n − 2 − k

t = r

1− r²

Phi coefficient

The phi coefficient (Lovell et al. 2015) was designed for compositional (relative) data such as

percentages. The usual correlation coefficients can be misleading for such data. The coefficient

measures the degree of proportionality; the smaller the value (close to zero), the more the variables

exhibit a proportional relationship. Pairs of variables can show strong correlations but low

proportionality when they are linearly related, but with a non-zero intercept term.

Each variable is first transformed using the centered logratio transformation

ꢏ

ꢐ

1

( )

clr x = �ln _ꢅ(풙), ⋯ , ln _ꢅ(풙)

�

84

where g(x) is the geometric mean of the data vector. The proportionality between two transformed

variables x and y is then computed by

⁄

var(풙 풚)

휑 =

var(풙)

where var(x) is the variance of the vector x. Significance is computed by testing whether x+y and x-y

are uncorrelated. Significance testing is however not recommended by Lovell et al. (2015).

Tetrachoric correlation

Tetrachoric correlation is appropriate when both variables are binary (0/1), but reflecting underlying

quantities on a continuous scale. Past uses an accurate approximation due to Bonett & Price (2005).

Let f₁₁be the number of rows with the binary pair 1/1, f₁₂the number of 1/0, f₂₁the number of 0/1,

and f₂₂is the number of rows with the binary pair 1/1. An estimator of the odds ratio of this

contingency table is

(

)(

)

푓

₁₁+ 0.5 푓 + 0.5

22

휔� =

)(

₁₂+ 0.5 푓 + 0.5

푓

21

We then calculate the sample proportions p from the frequencies f, after adding 0.5 to each f, i.e. p₁₁

= (f₁₁+0.5)/(N+2) etc. (we add 2 because the total number has been increased by adding 0.5 to each

of the four f; this is not precisely specified by Bonett & Price). Then calculate p₁₊= p₁₁+p₁₂and p₊₁

=

p₁₁+p₂₁. Also, p_minis the smallest marginal proportion (row or column sum in the p table). Then,

|

푝

|

−푝

1+ +1

5

2

(

)

1−

− 0.5−ꢑ

min

ꢃ̂

=

2

and the estimator for the tetrachoric correlation coefficient is

휋

휌�^∗= cos �

�

̂

1 + 휔�^푐

A standard error of this estimate is calculated by eq. (9) in Bonett & Price (2005), and a p value is

then estimated by a simple two-sided Z test. For small sample sizes, the permutation test calculated

by Past is probably better.

Permutation tests

Monte Carlo permutation tests (N=9999) are available for all the correlation coefficients except

polyserial and partial correlation, and the phi coefficient.

Correlation table plots

Plotting of the correlation table includes several options. The “Ellipses” function shows the

correlation coefficients r as ellipses with major axis of unity, and minor axis d according to Schilling

(1984):

85

1 − 푑²

1 + 푑²

ꢂ =

The correlation table plot is not valid for the phi coefficient.

References

Bonett, D.G. & Price, R.M. 2005. Inferential methods for the tetrachoric correlation coefficient.

Journal of Educational and Behavioral Statistics 30:213-225.

Lovell, D., Pawlowsky-Glahn, V., Egozcue, J.J., Marguerat, S. & Bähler, J. 2015. Proportionality: A valid

alternative to correlation for relative data. PLoS Computational Biology 11(3): e1004075

Ly, A., Verhagen, J., Wagenmakers, E.-J. 2016. Harold Jeffreys’s default Bayes factor hypothesis tests:

Explanation, extension, and application in psychology. Journal of Mathematical Psychology 72:19-32.

Olsson, U., F. Drasgow & N.J. Dorans. 1982. The polyserial correlation coefficient. Psychometrika

47:337-347.

Pearson, J.W., Olver, S., Porter, M.A. 2016. Numerical methods for the computation of the confluent

and Gauss hypergeometric functions. Numerical Algorithms DOI 10.1007/s11075-016-0173-0.

Press, W.H., S.A. Teukolsky, W.T. Vetterling & B.P. Flannery. 1992. Numerical Recipes in C. Cambridge

University Press.

Schilling, M.F. 1984. Some remarks on quick estimation of the correlation coefficient. The American

Statistician 38:330.

86

Intraclass correlation

A typical use of the intraclass correlation coefficient (ICC) is to quantify rater reliability, i.e. level of

agreement between several ‘raters’ measuring the same objects. It is a standard tool to assess

measurement error. ICC=1 would indicate perfect reliability. Raters (or ‘judges’) go in columns, while

the objects measured go in rows. In the example below there are four raters A-B, which have

measured 6 objects.

Past follows the standard reference, Shrout and Fleiss (1979), which provides a number of different

coefficients, referred to as ICC(m,k) where m is the model type. If k=1, the coefficient evaluates

individual measurements (by a single rater); otherwise it evaluates the average measurement across

raters. The models are

Model 1: the raters rating different objects are different, and randomly sampled from a larger set of

raters

Model 2: the same raters rate all objects, and the raters are a subset of a larger set of raters.

Model 3: no assumptions about the raters.

The most commonly used ICC is ICC(2,1), which is therefore marked in red in Past.

The analysis is based on a two-way ANOVA without replication, as described elsewhere in this

manual. Confidence intervals are parametric, following the equations of Shrout and Fleiss (1979). The

87

data in the example above are from the Shrout and Fleiss paper, the output from Past reproducing

their results.

Reference

Shrout, P.E., Fleiss, J.L. 1979. Intraclass correlations: Uses in assessing rater reliability. Psychological

Bulletin 86:420-428.

88

Normality tests

Four statistical tests for normal distribution of one or several samples of univariate data, given in one

or more separate columns or with a single data column and a group column. The data used below

were generated by the normal and uniform random number generators in Past (‘Evaluate expression’

module).

For all four tests, the null hypothesis is

H₀: The sample was taken from a population with normal distribution.

If the given p(normal) is less than 0.05, normal distribution can be rejected (marked in pink). Of the

given tests, the Shapiro-Wilk and Anderson-Darling are considered to be the more exact, and the

Lilliefors and Jarque-Bera are given for reference. And even poorer test (four-bin chi-squared) was

included in previous versions of Past. There is a maximum sample size of n=5000, while the minimum

sample size is 3 (the tests will of course have extremely small power for such small n).

Remember the multiple testing issue if you run these tests on several samples – a Bonferroni or

other correction may be appropriate.

Missing data: Supported by deletion.

Shapiro-Wilk test

The Shapiro-Wilk test (Shapiro & Wilk 1965) returns a test statistic W, which is small for non-normal

samples, and a p value. The implementation is based on the standard code “AS R94” (Royston 1995),

correcting an inaccuracy in the previous algorithm “AS 181” for large sample sizes.

Anderson-Darling test

The data X_iare sorted in ascending sequence, and normalized for mean and standard deviation:

89

ˆ

X_i− µ

.

Y_i=

ˆ

σ

With F the normal cumulative distribution function (CDF), the test statistic is

1

n

A²= −n −

(2i −1)[ln F(Y_i)+ ln(1− F(Y_n+1−k))

]

.

∑

i=1

n

Significance is estimated according to Stephens (1986). First, a correction for small sample size is

applied:

0.75

2.25

n²







A^*2= A²1+

+

.



n





The p value is estimated as

2

exp(1.2937 5.709A^*2

−

+

A^*2−

0.0186(A^*2

)

A^*2

≥

0.6





2

p

(

−

1.38

(A^*2

)

0.34

< A^*2<

0.6

0.2 < A^*2≤ 0.6

A^*2≤ 0.2



ex 0.9177 4.279



p =



2

1− exp(− 8.318 + 42.796a^*2− 59.938(A^*2

)



2

1− exp

(

−13.436 +101.14a^*2− 223.73

(

a^*2

)





Lilliefors test

The Lilliefors test is basically the same as the Kolmogorov-Smirnov test, comparing the sample

distribution with a normal distribution with mean and variance estimated from the data. Because of

this parameter estimation, significance must be computed in a different way than for Kolmogorov-

Smirnov. In addition to the Monte Carlo procedure, Past reports a p value using an approximation to

published tables given by Molin & Abdi (1998) and Abdi & Molin (2007):

b₂= 0.08861783849346

b₁= 1.30748185078790

b₀= 0.37872256037043

2

(

)

(

)

(

⁄

)

− 푏₁+ ꢁ + � 푏₁+ ꢁ

− 4푏₂푏₀− 1 퐿

퐴 =

2푏₂

p = -0.37782822932809 + 1.67819837908004A - 3.02959249450445A²

+ 2.80015798142101A³- 1.39874347510845A⁴

+ 0.40466213484419A⁵- 0.06353440854207A⁶

+ 0.00287462087623A⁷+ 0.00069650013110A⁸

+ 0.00011872227037A⁹+ 0.00000575586834A¹⁰

Jarque-Bera test

The Jarque-Bera test (Jarque & Bera 1987) is based on skewness S and kurtosis K. The test statistic is

90

2













n

(K − 3)

2

JB =

S +

6

4

.

In this context, the skewness and kurtosis used are

3

(x_i− x)

1

∑

S =







³





n

1

2

(x_i− x)

∑



n

,

4

(x_i− x)

1

∑

K =



⁴





n

1

2





(x_i− x)

∑



n

.

Note that these equations contain simpler estimators than the G₁and G₂given in the Univariate

summary statistics module, and that the kurtosis here will be 3, not zero, for a normal distribution.

Asymptotically (for large sample sizes), the test statistic has a chi-square distribution with two

degrees of freedom, and this forms the basis for the p value given by Past. It is known that this

approach works well only for large sample sizes, and Past therefore also includes a significance test

based on Monte Carlo simulation, with 10,000 random values taken from a normal distribution.

References

Abdi, H. & Molin, P. 2007. Lilliefors/Van Soest’s test of normality. In: Neil Salkind (Ed.) Encyclopedia

of Measurement and Statistics. Thousand Oaks (CA): Sage.

Jarque, C. M. & Bera, A. K. 1987. A test for normality of observations and regression residuals.

International Statistical Review 55:163–172.

Molin, P. & Abdi, H. 1998. New tables and numerical approximation for the

KolmogorovSmirnov/Lilliefors/Van Soest test of normality. Technical report, University of Bourgogne.

Royston, P. 1995. A remark on AS 181: The W-test for normality. Applied Statistics 44:547-551.

Shapiro, S. S. & Wilk, M. B. 1965. An analysis of variance test for normality (complete samples).

Biometrika 52:591–611.

Stephens, M.A. 1986. Tests based on edf statistics. Pp. 97-194 in D'Agostino, R.B. & Stephens, M.A.

(eds.), Goodness-of-Fit Techniques. New York: Marcel Dekker.

91

Outlier tests

A single column of numbers is required. These tests provide objective procedures for detecting

outliers in normally distributed data. The “single outlier” tests (Grubbs and Dixon) are designed to

detect one outlier only, and should not be repeated for several outliers. The “multiple outlier” test

(generalized ESD) attempts to detect multiple outliers, if present.

Grubbs test

The Grubbs test (also known as the Pearson-Hartley or the Extreme Studentized Deviate) tests for a

single outlier. The basic reference is Grubbs (1950) but the actual reference for the G statistic as

defined by e.g. Wikipedia and NIST is difficult to trace down. In any case, we define G as

|

max 푥_ꢄ− 푥

̅

퐺 =

푠

where 푥

̅

is the sample mean and s is the sample standard deviation. The critical value for G at a two-

sided significance level of α is given by

2

�푡_훼

�

ꢁ − 1

⁄(

)

2푁 ,푁−2

�

퐺 >

2

�

ꢁ

√

ꢁ − 2 + �푡_훼

⁄(

)

2푁 ,푁−2

where 푡_{ꢒ⁄(2ꢁ),ꢁ−2}is the critical value of the t distribution with N-2 degrees of freedom and a

significance level of α/(2N). Past solves this equation for α to give the p value for given G and N. Note

that this is a two-sided test, testing for presence of an outlier at either end of the range (smallest or

largest value). The Grubb’s test is recommended for relatively large sample sizes (N>30). For smaller

sample sizes, the Dixon test is preferable, although the two tests usually give similar results.

Dixon’s test

The Dixon test (Dixon 1950) tests for a single outlier. It can only be used for small sample sizes

(N<=30) but is then considered superior to the Grubbs test. The gap between the smallest (or largest)

value and its adjacent value is compared to the total range, giving a test statistic Q. The calculation of

Q depends on sample size as follows (the sample values x are sorted in ascending order):

Sample size N

Test statistic Q

At data minimum

At data maximum

푥₂− 푥₁

푥_푁− 푥₁

푥_푁− 푥_푁−1

푥_푁− 푥₁

3 ≤ N ≤ 7

r₁₀

푥₂− 푥₁

푥_푁− 푥_푁−1

푥_푁− 푥₂

8 ≤ N ≤ 10

11 ≤ N ≤ 13

14 ≤ N ≤ 30

r₁₁

r₂₁

r₂₂

푥_푁−1− 푥₁

푥₃− 푥₁

푥_푁− 푥_푁−2

푥_푁− 푥₂

푥_푁−1− 푥₁

푥₃− 푥₁

푥_푁− 푥_푁−2

푥_푁− 푥₃

푥_푁−2− 푥₁

92

The p value (two-tailed) is estimated by Monte Carlo simulation with 200,000 random, normally

distributed samples of size N (it will vary slightly between runs).

Generalized ESD (Extreme Studentized Deviate) test

This procedure can detect more than one outlier. Moreover, it can detect outliers even when the

one-outlier tests above do not report significance, because of so-called “masking”.

The procedure starts by testing the most extreme value in the complete sample, giving a test statistic

R₁(= Grubbs G). This most extreme value is then removed from the sample and the procedure is

repeated until 20% of the sample has been tested. The critical value R_crit(for significance at p<0.05) is

adjusted for each iteration (Rosner 1983). Past marks the values for which R>R_critin pink. A p value is

not calculated explicitly in Past.

Important: All of the most extreme values, up to the last value for which R>R_crit, are to be considered

outliers, and are marked as such in Past. Quite often, the initial, most extreme values do not give

R>R_critbut they can still be outliers because of a significant value further down in the list. This is due

to the masking effect. It looks odd, but is not a bug!

References

Dixon, W.J. 1950. Analysis of extreme values. Annals of Mathematical Statistics 21:488-506.

Grubbs, F. 1950. Sample criteria for testing outlying observations. Annals of Mathematical Statistics

21:27-58.

Rosner, B. 1983. Percentage points for a generalized ESD many-outlier procedure. Technometrics

25:165-172.

93

Contingency table (chi²etc.)

This test expects a frequency table with numbers of elements in different categories (rows and

columns). Rows represent the different states of one nominal variable, columns represent the states

of another nominal variable, and cells contain the integer counts of occurrences of that specific state

(row, column) of the two variables. The contingency table analysis then gives information on whether

the two variables are associated. For example, this test can be used to compare two samples

(columns) with the number of individuals in each taxon organized in the rows. You should be

cautious about this test if any of the cells contain less than five individuals (see Fisher’s exact test

below).

The significance of association between the two variables is given, with p values from the chi-squared

distribution and from a Monte Carlo (random) permutation test where all cells are random but with

constant row and column sums.

For 2x2 tables, Yates’s correction is applied.

The "Sample vs. expected" box should be ticked if you have two columns, and your second column

consists of counts from a theoretical distribution (expected values) with zero sampling error, possibly

non-integer. This is not a small-sample correction. In this case, only the chi-squared test is available.

See e.g. Brown & Rothery (1993) or Davis (1986) for details.

The Fisher's exact test is also given (two-tailed). When available, the Fisher's exact test will be

superior to the chi-square. For large tables or large counts, the computation time can be prohibitive,

and this test is not carried out. In such cases the parametric test is probably acceptable in any case.

The procedure is complex and based on the network algorithm of Mehta & Patel (1986).

94

Two further measures of association are given. Both are transformations of chi-squared (Press et al.

1992). With n the total sum of counts, M the number of rows and N the number of columns:

χ²

Cramer’s V:

V =

n min(M −1, N −1)

χ²

Contingency coefficient C:

C =

χ²+ n

Missing data not supported.

Residuals

If you get a significant association (p<0.05) in the chi-squared test, it may be of interest to see which

of the cells contribute most strongly to the departure from the expected values under the null

hypothesis of no association (post-hoc analysis). The table of residuals can show the following values

for each cell:

Raw residuals: O-E, where O is the observed and E the expected value.

Standardized residuals: (O-E)/√E, standardizing for the magnitude of the expected value.

Adjusted residuals:

푂 − 퐸

adj_resid =

(

⁄ )(

⁄ )

�퐸 1 − RowMarginal ꢀ 1 − ColumnMarginal ꢀ

where the RowMarginal is the row sum and ColumnMarginal is the column sum of the cell.

p values: The adjusted residuals are approximately normally distributed, meaning that values outside

the two-sigma interval [-1.96, 1.96] can be considered significant at p<0.05, although the multiple

testing problem applies. It is recommended to use the Bonferroni correction. Significant p values are

marked in pink.

References

Brown, D. & P. Rothery. 1993. Models in biology: mathematics, statistics and computing. John Wiley

& Sons.

Davis, J.C. 1986. Statistics and Data Analysis in Geology. John Wiley & Sons.

Mehta, C.R. & N.R. Patel. 1986. Algorithm 643: FEXACT: a FORTRAN subroutine for Fisher's exact test

on unordered r×c contingency tables. ACM Transactions on Mathematical Software 12:154-161.

Press, W.H., S.A. Teukolsky, W.T. Vetterling & B.P. Flannery. 1992. Numerical Recipes in C. Cambridge

University Press.

95

Cochran-Mantel-Haenszel test

Similar to the chi-squared test but testing several (repeated) 2x2 contingency tables simultaneously,

in order to control for a nuisance factor (Mantel & Haenszel 1959). An example could be controlling

for season when testing for the effect of a medicine that was used in repeated experiments through

a year. The 2x2 tables must be given in consecutive rows in the spreadsheet (first, one 2x2 table,

followed below by the next table, etc.).

Our algorithm follows McDonald (2014). Indexing the individual tables by k, we write one particular

table as

푎_푘

ꢃ_푘

푏_푘

푑_푘

�

With M the number of tables, and n_k= a_k+ b_k+ c_k+ d_k, the chi-squared is calculated as

2

ꢓ

푘=1

∑

(

)(

)⁄

)

��

푎_푘− 푎_푘+ 푏_푘푎_푘+ ꢃ_푘ꢀ_푘� − 0.5�

휒_ꢓ²_ꢔ

=

ꢓ

푘=1

3

2

∑

(

) (

)(

) (

)

푎_푘+ 푏_푘푎_푘+ ꢃ_푘푏_푘+ 푑_푘ꢃ_푘+ 푑_푘/ ꢀ_푘− ꢀ_푘

Note that other, algebraically equivalent, forms are often given in the literature. The subtraction of

0.5 is a continuity correction, not always included in other software. This test statistic is distributed

as χ2 with one degree of freedom.

In addition, the common odds ratio (equal to one for total independence) is calculated using the

Mantel-Haenszel (1959) estimate:

ꢓ

∑

⁄

_푘=1푎_푘푑_푘ꢀ_푘

�

휃_ꢓꢔ

=

ꢓ

∑

⁄

_푘=1푏_푘ꢃ_푘ꢀ_푘

The confidence interval is calculated as [exp(ln θ_MH– 1.96 √V_US), exp(ln θ_MH+ 1.96 √V_US)] where V_USis

the variance of the logarithm of θ_MH(Robins et al. 1986).

Missing data not supported.

References

Mantel, N. & W. Haenszel. 1959. Statistical aspects of the analysis of data from retrospective studies

of disease. Journal of the National Cancer Institute 22:719-748.

McDonald, J.H. 2014. Handbook of Biological Statistics (3rd ed.). Sparky House Publishing, Baltimore,

Maryland.

Robins, J., Breslow, N., Greenland, S. 1986. Estimators of the Mantel-Haenszel variance consistent in

both sparse data and large-strata limiting models. Biometrics 42:311-323.

96

Risk/odds

This module compares the counts of a binary outcome under two different treatments, with statistics

that are in common use in medicine. The data are entered in a 2x2 table, with treatments in rows

and counts of the two different outcomes in columns.

The following example shows the results of a vaccination trial on 460 patients:

Got influenza

Did not get influenza

Vaccine

Placebo

20

80

220

140

In general, the data take the form

Outcome 1

Outcome 2

Treatment 1

Treatment 2

d₁

d₀

h₁

h₀

Let n₁=d₁+h₁, n₀=d₀+h₀and p₁=d₁/n₁, p₀=d₀/n₀. The statistics are then computed as follows:

97

Risk difference:

RD = p₁-p₀

95% confidence interval on risk difference (Pearson’s chi-squared):

p_|1(1− p₁)

p₀(1− p_0|

)

s_e=

+

n₁

n₀

Interval: RD - 1.96 s_eto RD + 1.96 s_e

Z test on risk difference (two-tailed):

RD

s_e

z =

Risk ratio:

RR = p₁/p₀

95% confidence interval on risk ratio (“delta method”):

1

s_e(ln RR)

=

−

+

−

d₁

n₁

d₀

n₀

EF = e^1.96s

e

Interval: RR / EF to RR x EF

Z test on risk ratio (two-tailed):

ln RR

z =

s_e

d₁h₁

d₀h₀

Odds ratio:

OR =

95% confidence interval on odds ratio (“Woolfs’s formula”):

1

s_e(lnOR)

=

+

d₁

h₁

d₀

h₀

EF = e^1.96s

e

Interval: OR / EF to OR x EF

Note there is currently no continuity correction.

Missing data are not allowed and will give an error message.

98

Single proportion

A simple module for calculating the probability of an observed (sample) proportion (in the range 0-1)

against a hypothetical proportion. No input data are required in the spreadsheet.

With p the sample proportion, P the hypothetical proportion, and n the sample size, we calculate the

standard deviation

(

)

푃 1 − 푃

�

휎 =

ꢀ

Further, we calculate the z (normal distribution) statistic

ꢕ − 푃

푧 =

휎

The (two-tailed) significance is calculated directly from z and the normal distribution.

The 95% confidence interval for the proportion is calculated using two different methods. The ’exact’

interval is computed using the Clopper-Pearson method (Clopper and Pearson 1934) as

−1

ꢀ − 푥 + 1

ꢀ − 푥

�1 +

�

< 휃 < �1 +

�

[

⁄

(

)]

(

) [ ⁄ )]

(

)

(

푥ꢎ 1 − ꢒ 2 ; 2푥, 2 ꢀ − 푥 + 1

푥 + 1 ꢎ ꢒ 2 ; 2 푥 + 1 , 2 ꢀ − 푥

where α = 0.05, x is the number of successes computed as round(pn), and F(c; d₁, d₂) is the 1-c

quantile for an F distribution with d₁and d₂degrees of freedom. The normal approximation is

computed as

(

)

ꢕ 1 − ꢕ

�

푏 =

ꢀ

(

)

ꢉ퐼 = ꢕ − 1.96푏, ꢕ + 1.96푏

The normal approximation CI is more commonly used. The exact CI is more accurate for small n. For

large n the two methods will give similar results.

Reference

Clopper, C. & Pearson, E. S. 1934. The use of confidence or fiducial limits illustrated in the case of the

binomial. Biometrika 26:404–413.

99

Multiple proportion confidence intervals

Accepts two columns of data, the first column with proportions given as percentages (0-100) and the

second column containing sample sizes (N). The program will compute 95% confidence intervals for

all the proportions using the Clopper-Pearson method (see above), and plot them.

Missing data are disregarded.

100

Ratio of counts confidence interval

This module calculates confidence intervals for ratios of counts. It is specially designed for

environmental indices based on microfossil counts in paleontology, on the form (a + c)/(b + c). This

general formula covers the common cases a/b, where c=0, and c/(b+c), where a=0. No data are

required in the spreadsheet.

Two methods are provided, as described by Escarguel et al. (2019). The first is a Monte Carlo

procedure based on a transformation of the normal distribution. The second is a straightforward

bootstrap, with two versions (simple, which can produce negative values), and percentile.

ꢘ+ꢙ

ꢘ

ꢙ

+

ꢖ+푐

ꢗ+푐

ꢚ

ꢛ+ꢙ

ꢚ

ꢛ

ꢚ

For the Monte Carlo procedure, first note that ꢂ =

=

_ꢙ, where T = a+b+c.

ꢚ

Let e and g be the arcsine-transformed values 푒 = sin^-1√(a/T) and g = sin^-1√(c/T). e and g should then

be normally distributed variables with a sample standard deviation s = √(1/(4푇)) (Sokal and Rohlf,

2011). Therefore, let e* and g* be normally distributed random variates with mean e and g, and

standard deviation s. Then, e* and g* are back-transformed to proportion values (a/T)*= sin²e* and

∗

ꢘ

ꢙ

� � +� �

ꢚ

∗

(c/T)*= sin²g*. Finally, a Monte-Carlo estimate of r is calculated as ꢂ^∗=

.

ꢘ

1−� �

ꢚ

This procedure is reiterated a large number (say, 10,000) times, leading to a Monte-Carlo distribution

of r, from which the 2.5 and 97.5 percentiles define the 95% confidence interval.

Reference

Suchéras-Marx, B., Escarguel, G., Ferreira, J. & Hammer, Ø. 2019. Statistical confidence intervals for

relative abundances and abundance-based ratios: simple practical solutions for an old overlooked

question. Marine Micropaleontology 151: 101751.

101

Survival analysis (Kaplan-Meier curves, log-rank test etc.)

Survival analysis for two groups (treatments) with provision for right censoring. The module draws

Kaplan-Meier survival curves for the two groups and computes three different tests for equivalence.

The program expects four columns. The first column contains times to failure (death) or censoring

(failure not observed up to and including the given time) for the first group, the second column

indicates failure (1) or censoring (0) for the corresponding individuals. The last two columns contain

data for the second group. Failure times must be larger than zero.

The program also accepts only one treatment (given in two columns), or more than two treatments

in consecutive pairs of columns, plotting one or multiple Kaplan-Meier curves. The statistical tests are

only comparing the first two groups, however.

The Kaplan-Meier curves and the log-rank, Wilcoxon and Tarone-Ware tests are computed according

to Kleinbaum & Klein (2005).

Average time to failure includes the censored data. Average hazard is number of failures divided by

sum of times to failure or censorship.

The log-rank test is by chi-squared on the second group:







²





(

)

m − e

∑

2 j

2



(O₂− E₂

)

j

χ ²=

=

.

(

)(

)

n_{1 j}n_{2 j}m_{1 j}+ m_{2 j}n_{1 j}+ n_{2 j}− m_{1 j}− m_{2 j}

var(O₂− E₂

)

∑

2

(

) (

)

n_{1 j}+ n_{2 j}

n_{1 j}+ n_{2 j}−1

j

102

Here, n_ijis the number of individuals at risk, and m_ijthe number of failures, in group i at distinct

failure time j. The expected number of failures in group 2 at failure time j is

(

)

n_{2 j}m_{1 j}+ m_{2 j}

e_{2 j}=

.

n_{1 j}+ n_{2 j}

The chi-squared has one degree of freedom.

The Wilcoxon and Tarone-Ware tests are weighted versions of the log-rank test, where the terms in

the summation formulas for O₂-E₂and var(O₂-E₂) receive weights of n_jand √n_j, respectively. These

tests therefore give more weight to early failure times. They are not in common use compared with

the log-rank test.

This module is not strictly necessary for survival analysis without right censoring – the Mann-Whitney

test may be sufficient for this simpler case.

Missing data: Data points with missing value in one or both columns are disregarded.

Reference

Kleinbaum, D.G. & Klein, M. 2005. Survival analysis: a self-learning text. Springer.

103

Combine errors

A simple module for producing a weighted mean and its standard deviation from a collection of

measurements with errors (one sigma). Expects two columns: the data x and their one-sigma errors

σ. The sum of the individual gaussian distributions is also plotted.

The weighted mean and its standard deviation are computed as

x σ ²

∑

i

1

i

,

.

µ

=

σ =

1 σ ²

∑

i

∑

i

This is the maximum-likelihood estimator for the mean, assuming all the individual distributions are

normal with the same mean.

Missing data: Rows with missing data in one or both columns are deleted.

Propagate errors

A simple calculator for propagation of errors. Two columns are required, with values and their

standard deviations. The number of rows depends on the function f selected:

•

f = A + B + …: At least two rows (A and B) must be given. The variance is computed as

휎²= 휎_ꢌ²+ 휎_ꢌ²+ ⋯

ꢜ

104

•

f = A * B * …: At least two rows (A and B) must be given:

2

휎_ꢌ

휎_퐵

휎²≈ 푓²�� + � � + ⋯ �

ꢜ

퐴

ꢍ

f = A / B: Exactly two rows (A and B) must be given:

2

휎_ꢌ

휎_퐵

휎²≈ 푓²�� + � � �

ꢜ

퐴

ꢍ

f = A^a: A single row (A) must be given, in addition to the exponent a:

2

푓푎휎_ꢌ

휎²≈ �

�

ꢜ

퐴

•

f = A^B: Exactly two rows (A and B) must be given:

2

ꢍ휎_ꢌ

휎²≈ 푓²��

� + ln 퐴 휎_퐵

�

2

( ( )

)

ꢜ

퐴

The standard deviation of f is computed as the square root of the variance. Note that any covariation

between the given variables would produce additional terms in the equations above, but these are

assumed to be zero.

105

Multivariate menu

Principal components

Principal components analysis (PCA) finds hypothetical variables (components) accounting for as

much as possible of the variance in your multivariate data (Davis 1986, Harper 1999, Legendre &

Legendre 1998). These new variables are linear combinations of the original variables. PCA may be

used for reduction of the data set to only two variables (the two first components), for plotting

purposes. One might also hypothesize that the most important components are correlated with

other underlying variables. For morphometric data, this might be size, while for ecological data it

might be a physical gradient (e.g. temperature or depth).

The input data is a matrix of multivariate data, with items in rows and variates in columns.

The PCA routine finds the eigenvalues and eigenvectors of the variance-covariance matrix or the

correlation matrix, with the SVD algorithm. Use variance-covariance if all variables are measured in

the same units (e.g. centimetres). Use correlation (normalized var-covar) if the variables are

measured in different units; this implies normalizing all variables using division by their standard

deviations. The eigenvalues give a measure of the variance accounted for by the corresponding

eigenvectors (components). The percentages of variance accounted for by these components are

also given. If most of the variance is accounted for by the first one or two components, you have

scored a success, but if the variance is spread more or less evenly among the components, the PCA

has in a sense not been very successful.

In the example below (landmarks from gorilla skulls), component 1 is strong, explaining 45.9% of

variance. The bootstrapped confidence intervals are not shown unless the ‘Bootstrap N’ value is non-

zero.

106

Groups: If groups are specified with a group column, the PCA can optionally be carried out within-

group or between-group. In within-group PCA, the average within each group is subtracted prior to

eigenanalysis, essentially removing the differences between groups. In between-group PCA, the

eigenanalysis is carried out on the group means (i.e. the items analysed are the groups, not the

rows). For both within-group and between-group PCA, the PCA scores are computed using vector

products with the original data.

Supplementary variables: It is possible to include one or more initial columns containing additional

supplementary variables for the analysis. These variables are not included in the ordination. The

correlation coefficients between each supplementary variable and the PCA scores are presented as

vectors from the origin (triplot). The lengths of the vectors are arbitrarily scaled to make a readable

plot, so only their directions and relative lengths should be considered.

Row-wise bootstrapping is carried out if a positive number of bootstrap replicates (e.g. 1000) is given

in the 'Bootstrap N' box. The bootstrapped components are re-ordered and reversed according to

Peres-Neto et al. (2003) to increase correspondence with the original axes. 95% bootstrapped

confidence intervals are given for the eigenvalues.

The 'Scree plot' (simple plot of eigenvalues) may also indicate the number of significant components.

After this curve starts to flatten out, the components may be regarded as insignificant. 95%

confidence intervals are shown if bootstrapping has been carried out. The eigenvalues expected

under a random model (Broken Stick) are optionally plotted - eigenvalues under this curve may

represent non-significant components (Jackson 1993).

In the gorilla example above, the eigenvalues for the 16 components (blue line) lie above the broken

stick values (red dashed line) for the first two components, although the broken stick is inside the

95% confidence interval for the second component.

The scatter plot shows all data points (rows) plotted in the coordinate system given by two of the

components. If you have groups, they will be shown with different symbols and colours. The Minimal

Spanning Tree is the shortest possible set of lines connecting all points. This may be used as a visual

107

aid in grouping close points. The MST is based on an Euclidean distance measure of the original data

points, and is most meaningful when all variables use the same unit. The 'Biplot' option shows a

projection of the original axes (variables) onto the scattergram. This is another visualisation of the

PCA loadings (coefficients) - see below.

1

d_k

If the "Eigenval scale" is ticked, the data points will be scaled by

, and the biplot eigenvectors

d_k

by

- this is the correlation biplot of Legendre & Legendre (1998). If not ticked, the data points

are not scaled, while the biplot eigenvectors are normalized to equal length (but not to unity, for

graphical reasons) - this is the distance biplot.

The loadings plot shows to what degree your different original variables (given in the original order

along the x axis) contribute to the different components (as chosen in the radio button panel). These

component loadings are important when you try to interpret the 'meaning' of the components. The

'Coefficients' option gives the PC coefficients, while 'Correlation' gives the correlation between a

variable and the PC scores. If bootstrapping has been carried out, 95% confidence intervals are

shown (only for the Coefficients option).

Sphericity tests

Bartlett’s sphericity test (Bartlett 1951) tests the null hypothesis that the points are sampled from a

spherical distribution. If so, PCA will not be able to provide a useful reduction of dimensionality. The

p value from this test should ideally be <0.05 (significant departure from sphericity). For PCA on the

correlation matrix, Past uses the chi-squared approximation

1

휒²= − �푀 − 1 −

2ꢁ + 5 � ln det 푹

(

)

6

108

where M is the number of points, N the number of variables, and R the correlation matrix. The

determinant is calculated as the product of eigenvalues. Bartlett (1951) used M instead of M-1, but

later authors have tended to the latter as a bias correction. The χ²has N(N-1)/2 degrees of freedom.

Past also provides a Bartlett’s sphericity test for PCA on the var-covar matrix V, following Bartlett

(1951). However, this test is not in common use, and its properties are not quite clear. It assumes

equal variances in the variables. Use with caution:

1

휒²= − �푀 −

2ꢁ + 1 + 2 ꢁ� ln det 푽

(

)

⁄

6

with (N+2)(N-1)/2 degrees of freedom. Before calculating the determinant, all eigenvalues are

divided by their mean.

The Kaiser-Meyer-Olkin (KMO) measure (Kaiser 1970), also known as Measure of Sampling Adequacy

(MSA), is only supported for PCA on the correlation matrix.

The inverse of the correlation matrix P = R^-1is used to calculate the partial correlation matrix A, with

elements

푃

ꢄ푗

퐴_ꢄ푗= −

푃 푃

�

ꢄꢄ 푗푗

The KMO is then calculated as

2

∑_ꢄ∑

_푗≠ꢄ푅_ꢄ푗

퐾푀푂 =

2

∑_ꢄ∑

∑ ∑

_푗≠ꢄ퐴_ꢄ푗

_푗≠ꢄ푅_ꢄ푗

+

ꢄ

KMO < 0.5 is reported as "unacceptable"; 0.5 ≤ KMO < 0.7 is reported as “mediocre”; 0.7 ≤ KMO <0.8

is “good”; 0.8 ≤ KMO < 1 is “excellent”.

Missing data can be handled by one of two methods:

1. Mean value imputation: Missing values are replaced by their column average. Not

recommended.

2. Iterative imputation: Missing values are inititally replaced by their column average. An initial

PCA run is then used to compute regression values for the missing data. The procedure is

iterated until convergence. This is usually the preferred method, but can cause some

overestimation of the strength of the components (see Ilin & Raiko 2010).

References

Bartlett, M.S. 1951. The effect of standardization on a χ²approximation in factor analysis. Biometrika

38:337-344.

Davis, J.C. 1986. Statistics and Data Analysis in Geology. John Wiley & Sons.

Harper, D.A.T. (ed.). 1999. Numerical Palaeobiology. John Wiley & Sons.

109

Ilin, A. & T. Raiko. 2010. Practical approaches to Principal Component Analysis in the presence of

missing values. Journal of Machine Learning Research 11:1957-2000.

Jackson, D.A. 1993. Stopping rules in principal components analysis: a comparison of heuristical and

statistical approaches. Ecology 74:2204-2214.

Kaiser, H.F. 1970. A second generation little jiffy. Psychometrika 35:401-415.

Legendre, P. & L. Legendre. 1998. Numerical Ecology, 2nd English ed. Elsevier, 853 pp.

Peres-Neto, P.R., D.A. Jackson & K.M. Somers. 2003. Giving meaningful interpretation to ordination

axes: assessing loading significance in principal component analysis. Ecology 84:2347-2363.

110

Principal coordinates

Principal coordinates analysis (PCoA) is another ordination method, also known as Metric

Multidimensional Scaling. The algorithm is from Davis (1986).

The PCoA routine finds the eigenvalues and eigenvectors of a matrix containing the distances or

similarities between all data points. The Gower measure will normally be used instead of Euclidean

distance, which gives results similar to PCA. An additional eleven distance measures are available -

these are explained under Cluster Analysis. The eigenvalues, giving a measure of the variance

accounted for by the corresponding eigenvectors (coordinates) are given for the first four most

important coordinates (or fewer if there are fewer than four data points). The percentages of

variance accounted for by these components are also given.

The similarity/distance values are raised to the power of c (the "Transformation exponent") before

eigenanalysis. The standard value is c=2. Higher values (4 or 6) may decrease the "horseshoe" effect

(Podani & Miklos 2002).

The scatter plot allows you to see all your data points (rows) plotted in the coordinate system given

by the PCoA. If you have colored (grouped) rows, the different groups will be shown using different

symbols and colours. The "Eigenvalue scaling" option scales each axis using the square root of the

eigenvalue (recommended). The minimal spanning tree option is based on the selected similarity or

distance index in the original space.

Missing data is supported by pairwise deletion (not for the Raup-Crick, Rho or user-defined indices).

References

Davis, J.C. 1986. Statistics and Data Analysis in Geology. John Wiley & Sons.

Podani, J. & I. Miklos. 2002. Resemblance coefficients and the horseshoe effect in principal

coordinates analysis. Ecology 83:3331-3343.

111

Non-metric MDS

Non-metric multidimensional scaling is based on a distance matrix computed with any of 21

supported distance measures, as explained under Similarity and Distance Indices below. The

algorithm then attempts to place the data points in a two- or three-dimensional coordinate system

such that the ranked differences are preserved. For example, if the original distance between points

4 and 7 is the ninth largest of all distances between any two points, points 4 and 7 will ideally be

placed such that their euclidean distance in the 2D plane or 3D space is still the ninth largest. Non-

metric multidimensional scaling intentionally does not take absolute distances into account.

The program may converge on a different solution in each run, depending upon the initial conditions.

Each run is actually a sequence of 11 trials, from which the one with smallest stress is chosen. One of

these trials uses PCoA as the initial condition, the others are random. The solution is automatically

rotated to the major axes (2D and 3D).

The algorithm implemented in PAST, which seems to work very well, is based on a new approach

developed by Taguchi and Oono (2005).

The minimal spanning tree option is based on the selected similarity or distance index in the original

space.

Environmental variables: It is possible to include one or more initial columns containing additional

“environmental” variables for the analysis. These variables are not included in the ordination. The

correlation coefficients between each environmental variable and the NMDS scores are presented as

vectors from the origin. The lengths of the vectors are arbitrarily scaled to make a readable biplot, so

only their directions and relative lengths should be considered.

Column scores: The columns can be included in the NMDS plot as weighted averages of the row

scores, as in Correspondence Analysis. The weighting uses the raw data values, and therefore does

112

not honour the choice of similarity index. However, it seems to work well for e.g. ecological data,

allowing the plotting of species together with samples (sites).

Shepard plot: This plot of obtained versus observed (target) ranks indicates the quality of the result.

Ideally, all points should be placed on a straight ascending line (x=y). The R²values are the

coefficients of determination between distances along each ordination axis and the original distances

(perhaps not a very meaningful value but is reported by other NMDS programs so is included for

completeness).

Variance: Lists the Euclidean distances to the group centroids in the NMDS scatter plot. The data

points are arranged in columns according to group. These values can be used e.g. in the study of

disparity.

Missing data is supported by pairwise deletion (not for the Raup-Crick, Rho and user-defined

indices). For environmental variables, missing values are not included in the computation of

correlations.

Reference

Taguchi, Y.-H., Oono, Y. 2005. Relational patterns of gene expression via non-metric

multidimensional scaling analysis. Bioinformatics 21:730-40.

113

Correspondence analysis

Correspondence analysis (CA) is yet another ordination method, somewhat similar to PCA but for

counted data (Legendre & Legendre 1998). For comparing associations (columns) containing counts

of taxa, or counted taxa (rows) across associations, CA is the more appropriate algorithm. Also, CA is

more suitable if you expect that species have unimodal responses to the underlying parameters, that

is they favor a certain range of the parameter, becoming rare for lower and higher values (this is in

contrast to PCA, which assumes a linear response).

The CA routine finds the eigenvalues and eigenvectors of a matrix containing the Chi-squared

distances between all rows (or columns, if that is more efficient – the result is the same). The

algorithm follows Greenacre (2010), with SVD. The eigenvalue, giving a measure of the similarity

accounted for by the corresponding eigenvector, is given for each eigenvector. The percentages of

similarity accounted for by these components are also given.

The scatter plot allows you to see all your data points (rows) plotted in the coordinate system given

by the CA. If you have grouped rows, the different groups can be shown using separate convex hulls

and concentration ellipses.

In addition, the variables (columns, associations) can be plotted in the same coordinate system (Q

mode), optionally including the column labels. If your data are 'well behaved', taxa typical for an

association should plot in the vicinity of that association.

Relay plot: This is a composite diagram with one plot per column. The plots are ordered according to

CA column scores. Each data point is plotted with CA first-axis row scores on the vertical axis, and the

original data point value (abundance) in the given column on the horizontal axis. This may be most

114

useful when samples are in rows and taxa in columns. The relay plot will then show the taxa ordered

according to their positions along the gradients, and for each taxon the corresponding plot should

ideally show a unimodal peak, partly overlapping with the peak of the next taxon along the gradient

(see Hennebert & Lees 1991 for an example from sedimentology).

Scalings: Three different scalings of the row and column scores are available. In the row principal and

column principal scalings, the distances between rows, between columns, and between rows and

columns are all proportional to chi-squared distance. However, this will show rows and columns in

two different scales, which can be difficult to read. The symmetric scaling uses the same scale for

rows and columns, but chi-squared distance between rows and columns is not maintained.

Row and column contributions: These tables report the proportion (0-1) of contribution of each row

or column point to the inertia of each axis. In the literature, these values are referred to as absolute

contributions, or CTR. For row i and axis j, they are calculated as

ꢂ

ꢄ

휆_푗²ꢎ²

휔_ꢄ푗

=

ꢄ푗

where r_iis the row sum divided by total count, λ_jis the eigenvalue for the axis, and F are the row

scores. Absolute column contributions are calculated similarly, using column sums and scores.

Also, the contributions of each axis to the inertia of each point (relative contribution, or COR) are

given:

ꢎ²

ꢄ푗

휎_ꢄ푗

=

_ꢞꢎ²

∑

ꢄꢞ

Missing data is supported by column average substitution.

References

Greenacre, M. 2010. Biplots in practice. Fundación BBVA, 237 pp.

Hennebert, M. & A. Lees. 1991. Environmental gradients in carbonate sediments and rocks detected

by correspondence analysis: examples from the Recent of Norway and the Dinantian of southwest

England. Sedimentology 38:623-642.

Legendre, P. & L. Legendre. 1998. Numerical Ecology, 2nd English ed. Elsevier, 853 pp.

115

Detrended correspondence analysis

The Detrended Correspondence (DCA) module uses the same algorithm as Decorana (Hill & Gauch

1980), with modifications according to Oxanen & Minchin (1997). It is specialized for use on

'ecological' data sets with abundance data; samples in rows, taxa in columns.

Eigenvalues for the four ordination axes are given as in CA, indicating their relative importance in

explaining the spread in the data.

Detrending is a sort of normalization procedure in two steps. The first step involves an attempt to

'straighten out' points lying in an arch, which is a common occurrence. The second step involves

'spreading out' the points to avoid clustering of the points at the edges of the plot. Detrending may

seem an arbitrary procedure, but can be a useful aid in interpretation.

The units on the axes are standard deviations x 100.

Missing data is supported by column average substitution.

References

Hill, M.O. & H.G. Gauch Jr. 1980. Detrended Correspondence analysis: an improved ordination

technique. Vegetatio 42:47-58.

Oxanen, J. & P.R. Minchin. 1997. Instability of ordination results under changes in input data order:

explanations and remedies. Journal of Vegetation Science 8:447-454.

116

Canonical correspondence

Canonical Correspondence Analysis (Legendre & Legendre 1998) is correspondence analysis of a

site/species matrix where each site has given values for one or more environmental variables

(temperature, depth, grain size etc.). The ordination axes are linear combinations of the

environmental variables. CCA is thus an example of direct gradient analysis, where the gradient in

environmental variables is known a priori and the species abundances (or presence/absences) are

considered to be a response to this gradient.

Each site should occupy one row in the spreadsheet. The environmental variables should enter in the

first columns, followed by the abundance data (the program will ask for the number of

environmental variables).

The implementation in PAST follows the eigenanalysis algorithm given in Legendre & Legendre

(1998). The ordinations are given as site scores - fitted site scores are presently not available.

Environmental variables are plotted as correlations with site scores. Both scalings (type 1 and 2) of

Legendre & Legendre (1998) are available. Scaling 2 emphasizes relationships between species.

Missing values are supported by column average substitution.

Mystery rows: Rows can contain missing values (‘?’) for all environmental variables. These rows,

which must be placed at the bottom of the data matrix, are not included in the CCA analysis itself,

but their site scores are estimated using the CCA vectors and included in the biplot. This could be

used, for example, when environmental variables are only known for a modern data set but not for

“fossil” (downcore) samples.

Reference

Legendre, P. & L. Legendre. 1998. Numerical Ecology, 2nd English ed. Elsevier, 853 pp.

117

Seriation

Seriation of an absence-presence (0/1) matrix using the algorithm described by Brower & Kile (1988).

This method is typically applied to an association matrix with taxa (species) in the rows and samples

in the columns. For constrained seriation (see below), columns should be ordered according to some

criterion, normally stratigraphic level or position along a presumed faunal gradient.

The seriation routines attempt to reorganize the data matrix such that the presences are

concentrated along the diagonal. There are two algorithms: Constrained and unconstrained

optimization. In constrained optimization, only the rows (taxa) are free to move. Given an ordering of

the columns, this procedure finds the 'optimal' ordering of rows, that is, the ordering of taxa which

gives the prettiest range plot. Also, in the constrained mode, the program runs a 'Monte Carlo'

simulation, generating and seriating 30 random matrices with the same number of occurences within

each taxon, and compares these to the original matrix to see if it is more informative than a random

one (this procedure is time-consuming for large data sets).

In the unconstrained mode, both rows and columns are free to move.

Missing data are treated as absences.

Reference

Brower, J.C. & K.M. Kile. 1988. Seriation of an original data matrix as applied to palaeoecology.

Lethaia 21:79-93.

118

CABFAC factor analysis

This module implements the classical Imbrie & Kipp (1971) method of factor analysis and

environmental regression (CABFAC and REGRESS, see also Klovan & Imbrie 1971).

The program asks whether the first column contains environmental data. If not, a simple factor

analysis with Varimax rotation will be computed on row-normalized data.

If environmental data are included, the factors will be regressed onto the environmental variable

using the second-order (parabolic) method of Imbrie & Kipp, with cross terms. PAST then reports the

RMA regression of original environmental values against values reconstructed from the transfer

function. Different methods for cross-validation (leave-one-out and k-fold) are available. You can also

save the transfer function as a text file that can later be used for reconstruction of

palaeoenvironment (see below). This file contains:

•

Number of taxa

Number of factors

Factor scores for each taxon

Number of regression coefficients

Regression coefficients (second- and first-order terms, and intercept)

Missing values are supported by column average substitution.

References

Imbrie, J. & N.G. Kipp. 1971. A new micropaleontological method for quantitative paleoclimatology:

Application to a late Pleistocene Caribbean core. In: The Late Cenozoic Glacial Ages, edited by K.K.

Turekian, pp. 71-181, Yale Univ. Press, New Haven, CT.

Klovan, J.E. & J. Imbrie. 1971. An algorithm and FORTRAN-IV program for large scale Q-mode factor

analysis and calculation of factor scores. Mathematical Geology 3:61-77.

119

Discriminant analysis

This module provides discriminant analysis for two or more groups (the latter is sometimes called

Canonical Variates Analysis). The groups must be specified with a group column.

A scatter plot of specimens along the first two canonical axes produces maximal and second to

maximal separation between all groups. The axes are linear combinations of the original variables as

in PCA, and eigenvalues indicate amount of variation explained by these axes. If only two groups are

given, a histogram is plotted instead.

Missing data supported by column average substitution.

Classifier

Classifies the data, assigning each point to the group that gives minimal Mahalanobis distance to the

group mean. The Mahalanobis distance is calculated from the pooled within-group covariance

matrix, giving a linear discriminant classifier. The given and estimated group assignments are listed

for each point. In addition, group assignment is cross-validated by a leave-one-out cross-validation

(jackknifing) procedure.

Mystery specimens: Rows with unknown group, i.e. ‘?’ in the group column, are not included in the

discriminant analysis itself, but will be classified. In this way, it is possible to classify new specimens

that are not part of the training set.

120

Confusion matrix

A table with the numbers of points in each given group (rows) that are assigned to the different

groups (columns) by the classifier. Ideally each point should be assigned to its respective given group,

giving a diagonal confusion matrix. Off-diagonal counts indicate the degree of failure of classification.

Computational details

Different softwares use different versions of CVA. The computations used by Past are given below.

Let B be the given data, with n items in rows and k variates in columns, centered on the grand means

of columns (column averages subtracted). Let g be the number of groups, n_ithe number of items in

group i. Compute the gxk matrix X of weighted means of within group residuals, for group i and

variate j

X_ij= n_iB_ij

,

where

is a column average within group i. Compute B₂from B by centering within groups. Now

B_ij

compute W and the normalized, pooled, within-group covariance matrix W_cov

:

′

W = B₂B₂

1

W_cov

=

W

.

n − g

e and U are the eigenvalues and eigenvectors of W; e_cand U_care the eigenvalues and eigenvectors of

W_cov. Then,

′

′ ′

.

Z Z = diag(1 e)U X XUdiag(1 e)

a and A are the eigenvalues and eigenvectors of Z’Z. We take only the first g-1 eigenvectors (columns

of A), as the rest will be zero. The canonical variates are now

C = Udiag(1 e_c

)

A

.

The CVA scores are then BC. Reification of variables can be done along vectors W_covC.

121

Two-block PLS

Two-block Partial Least squares can be seen as an ordination method comparable with PCA, but with

the objective of maximizing covariance between two sets of variates on the same rows (specimens,

sites). For example, morphometric and environmental data collected on the same specimens can be

ordinated in order to study covariation between the two.

The program will ask for the number of columns belonging to the first block. The remaining columns

will be assigned to the second block. There are options for plotting PLS scores both within and across

blocks, and PLS loadings.

The algorithm follows Rohlf & Corti (2000). Permutation tests and biplots are not yet implemented.

Partition the nxp data matrix Y into Y₁and Y₂(the two blocks), with p₁and p₂columns. The

correlation or covariance matrix R of Y can then be partitioned as

R



₁₂



11

R =

.





R₂₁

R₂₂



The algorithm proceeds by singular value decomposition of the matrix R₁₂of correlations across

blocks:

R₁₂= F₁DF₂^t

.

The matrix D contains the singular values λ_ialong the diagonal. F₁contains the loadings for block 1,

and F₂the loadings for block 2 (cf. PCA).

The "Squared covar %" is a measure of the overall squared covariance between the two sets of

variables, in percent relative to the maximum possible (all correlations equal to 1) (Rohlf & Corti p.

741). The "% covar” of axes are the amounts of covariance explained by each PLS axis, in percents of

λ_i²

the total covariance. They are calculated as 100

.

λ²

∑

i

Missing data supported by column average substitution.

Reference

Rohlf, F.J. & M. Corti. 2000. Use of two-block partial least squares to study covariation in shape.

Systematic Biology 49:740-753.

122

Redundancy analysis (RDA)

Redundancy Analysis (Legendre & Legendre 1998) takes as input a site/data matrix where each site

has given values for one or more environmental/explanatory variables as well as a number of

response (dependent) variables. The ordination axes are linear combinations of the explanatory

(independent) variables. RDA can be thought of as a canonical version of PCA, i.e. with axes

constrained by explanatory variables.

Each site should occupy one row in the spreadsheet. The explanatory variables should enter in the

first columns, followed by the response data (the program will ask for the number of explanatory

variables).

The implementation in PAST follows Legendre & Legendre (1998). The ordinations can be shown as

site scores or fitted site scores. Explanatory variables are plotted as correlations with site scores.

Both scalings (type 1 and 2) of Legendre & Legendre (1998) are available. The scores can be manually

scaled with the “Amplitude” controls for a clearer plot (these factors should be reported together

with the plot).

Missing values are supported by column average substitution.

Mystery rows: Rows can contain missing values (‘?’) for all explanatory variables. These rows, which

must be placed at the bottom of the data matrix, are not included in the RDA analysis itself, but their

site scores are estimated using the RDA vectors and included in the biplot. This could be used, for

example, when explanatory variables are only known for a modern data set but not for “fossil”

(downcore) samples. Mystery rows are only reported for unfitted site scores.

Reference

Legendre, P. & L. Legendre. 1998. Numerical Ecology, 2nd English ed. Elsevier, 853 pp

123

Nonlinear ordination (UMAP)

PAST includes three methods for nonlinear ordination (embedding): spectral embedding, UMAP, and

ISOMAP. The first two are found in the UMAP module, the last in the ISOMAP module. These

methods can be compared with e.g. Principal Coordinates Analysis, and are based on a distance

measure. However, they can pick up groups having complicated shapes in high-dimensional space,

possibly interfingering with other groups. These methods can be very good at identifying groups and

gradients (often “too good”, causing overfitting), but can also behave quite erratically, depending on

analysis parameters.

The input is a multivariate data set with variables in columns, optionally with a group column for

showing given groups (this does not influence the ordination). A distance measure must be chosen

(default Euclidean).

Spectral embedding

The spectral embedding algorithm basically follows Belkin & Niyogi (2003). First, a graph is

constructed where two points i and j are connected if and only if i is among the n nearest neighbours

of j, or j is among the n nearest neighbours of i. The number n can be chosen by the user (“Embed

neighbors”), and it can substantially influence the ordination. Larger n will tend to produce larger,

less concentrated groups.

A matrix W is constructed, with W_ij=1 if i and j are connected (adjacency matrix). If the “Embed

kernel” option is selected, the values in this matrix are additionally scaled by the kernel function

2

�ꢏ −ꢏ �

ꢟ

ꢠ

ꢡ

푊 = 푒

ꢄ푗

where the norm is according to the selected distance measure. The parameter t is fixed at 1/50 of

the largest distance between any pairs of points. Then compute the “graph Laplacian” L = D – W,

where D is a diagonal matrix containing the row (or column, as W is symmetric) sums of W. Past uses

the normalized graph Laplacian,

1

2

1

2

− −

′

L = D

LD

124

The scores on the two ordination axes are given by the two eigenvectors corresponding to the

second smallest and third smallest eigenvalue.

UMAP

UMAP (McInnes et al. 2018) is a modern method for nonlinear ordination. The implementation in

Past also builds on a description by Oskolkov (2019). The procedure is complex, for details see the

references.

The implementation in Past uses a spectral embedding as starting condition, with parameters as

given above. The “UMAP neighbors” (k) and the “Minimum distance” (min_dist) parameters can

change the ordination quite a lot.

References

Belkin, M., Niyogi, P. 2003. Laplacian eigenmaps for dimensionality reduction and data

representation. Neural Computation 15:1373-1396.

McInnes, L., Healy, J., Melville, J. 2018. UMAP: Uniform Manifold Approximation and Projection for

dimension reduction. https://arxiv.org/abs/1802.03426.

Oskolkov, N. 2019. How to program UMAP from scratch. https://towardsdatascience.com/how-to-

program-umap-from-scratch-e6eff67f55fe

125

Nonlinear ordination (ISOMAP)

ISOMAP (Tenenbaum et al. 2000) is a relatively simple but often effective method for nonlinear

dimensionality reduction. The implementation in Past follows these steps:

1. Find the k closest neighbors of each point, using any distance measure. The result will

strongly depend on the value of k, which must be specified by the user.

2. Produce a graph where two vertices are connected if they are among their closest neighbors.

3. Find the shortest paths between all vertices in the graph using the Floyd-Warshall algorithm.

These path lengths represent “geodesic” distances along the manifold.

4. Compute a standard PCoA (metric multidimensional scaling) on the geodesic distances.

Reference

Tenenbaum, J.B., de Silva, V., Langford, J.C. 2000. A global geometric framework for nonlinear

dimensionality reduction. Science 290:2319-2323.

126

Classical cluster analysis

The hierarchical clustering routine produces a 'dendrogram' showing how data points (rows) can be

clustered. For 'R' mode clustering, putting weight on groupings of taxa, taxa should go in rows. It is

also possible to find groupings of variables or associations (Q mode), by entering taxa in columns.

Switching between the two is done by transposing the matrix (in the Edit menu).

Three different algorithms are available:

•

Unweighted pair-group average (UPGMA). Clusters are joined based on the average distance

between all members in the two groups.

Single linkage (nearest neighbor). Clusters are joined based on the smallest distance between

the two groups.

•

Ward's method. Clusters are joined such that increase in within-group variance is minimized.

Complete linkage. Clusters are joined based on the largest distance within groups.

One method is not necessarily better than the other, though single linkage is not recommended by

some. It can be useful to compare the dendrograms given by the different algorithms, to informally

assess the robustness of the clusters.

For Ward's method, a Euclidean distance measure is inherent to the algorithm. For the other

methods, the distance matrix can be computed using 24 different indices, as described under the

‘Similarity and distance indices’ section.

Missing data: The cluster analysis algorithm can handle missing data, coded with question marks (?).

This is done using pairwise deletion, meaning that when distance is calculated between two points,

any variables that are missing are ignored in the calculation. For Raup-Crick, missing values are

treated as absence. Missing data are not supported for Ward's method, nor for the Rho similarity

measure.

127

Two-way clustering: The two-way option allows simultaneous clustering in R-mode and Q-mode.

Stratigraphically constrained clustering: This option will allow only adjacent rows or groups of rows

to be joined during the agglomerative clustering procedure. May produce strange-looking (but

correct) dendrograms.

Group-constrained clustering: This option will only allow joining of clusters within the given groups.

May produce strange-looking (but correct) dendrograms.

Bootstrapping: If a number of bootstrap replicates is given (e.g. 100), the columns are subjected to

resampling. Press Enter after typing to update the value in the “Boot N” number box. The percentage

of replicates where each node is still supported is given on the dendrogram.

Note on Ward’s method: PAST produces Ward’s dendrograms identical to those made by Stata, but

somewhat different from those produced by Statistica. The reason for the discrepancy is unknown.

Constrained clustering does not work.

128

Neighbour joining

Neigbour joining clustering (Saitou & Nei 1987) is an alternative method for hierarchical cluster

analysis. The method was originally developed for phylogenetic analysis, but may be superior to

UPGMA also for ecological data. In contrast with UPGMA, two branches from the same internal node

do not need to have equal branch lengths. A phylogram (unrooted dendrogram with proportional

branch lengths) is given.

Distance indices and bootstrapping are as for other cluster analysis (above). To run the bootstrap

analysis, type in the number of required bootstratp replicates (e.g. 1000, 10000) in the “Boot N” box

and press Enter to update the value.

Negative branch lengths are forced to zero, and transferred to the adjacent branch according to

Kuhner & Felsenstein (1994).

The tree is by default rooted on the last branch added during tree construction (this is not midpoint

rooting). Optionally, the tree can be rooted on any row in the data matrix, as selected in the Root

menu.

Missing data supported by pairwise deletion.

References

Kuhner, M.K. & J. Felsenstein. 1994. A simulation comparison of phylogeny algorithms under equal

and unequal evolutionary rates. Molecular Biology and Evolution 11:459-468

Saitou, N. & M. Nei. 1987. The neighbor-joining method: a new method for reconstructing

phylogenetic trees. Molecular Biology and Evolution 4:406-425.

129

K-means clustering

K-means clustering (e.g. Bow 1984) is a non-hierarchical clustering method. The number of clusters

to use is specified by the user, usually according to some hypothesis such as there being two sexes,

four geographical regions or three species in the data set.

The cluster assignments are initially random, using the Forgy method. In an iterative procedure,

items are then moved to the cluster which has the closest cluster mean, and the cluster means are

updated accordingly. This continues until items are no longer "jumping" to other clusters. The result

of the clustering is to some extent dependent upon the initial, random ordering, and cluster

assignments may therefore differ from run to run. This is not a bug, but normal behaviour in k-means

clustering.

The cluster assignments may be copied and pasted back into the main spreadsheet, and

corresponding colors (symbols) assigned to the items using the 'Numbers to colors' option in the Edit

menu.

Clustering statistics

WGSS: The Within-Group (or Within-Cluster) sum of squares; will decrease with cluster

separationand with the number of clusters.

F: Ratio of Between-Group to Within-Group sum of squares; will increase with cluster separation and

number of clusters.

Variance explained (percent): Ratio of Between-Group to total sum of squares, multiplied with 100.

Average silhouette: The average silhouette value over all objects (see below).

Silhouette plot and table

The silhouette plot (Rousseeuw 1987) gives an indication of how well each object has been classified,

on a scale from -1 to 1, where 1 means a perfectly appropriate assignment to a cluster; -1 means the

object would have been better placed in another cluster; 0 means the object is on the boundary

between two clusters.

Missing data supported by column average substitution.

References

Bow, S.-T. 1984. Pattern recognition. Marcel Dekker, New York.

Rousseeuw, P.J. 1987. Silhouettes: a graphical aid to the interpretation and validation of cluster

analysis". Computational and Applied Mathematics 20:53–65.

130

K-medoids clustering

K-medoids clustering (Kaufman & Rousseeuw 1990) can be compared to k-means clustering, and

requires the user to select the number of clusters. Unlike k-means, the clusters are centered on a

point in the data set, rather than a cluster mean. Also, importantly, k-medoids allows any distance

measure to be used, making it useful for e.g. ecological and genetic data.

The algorithm in Past follows the original PAM method described by Kaufman & Rousseeuw (1990).

Reference

Kaufman, L. & Rousseeuw, P.J. 1990. Partitioning around medoids (program PAM). Ch. 2 in Finding

groups in data: An introduction to cluster analysis. John Wiley & Sons.

DBSCAN clustering

DBSCAN clustering (Ester et al. 1996) is a non-hierarchical clustering method. Unlike K-means or K-

medoids, the number of clusters does not need to be pre-defined. Any distance measure can be

used. DBSCAN is a popular clustering method in the machine learning community. It does not always

perform better than K-medoids but can trace out non-linear cluster boundaries.

The “MinPts” parameter is the minimum number of points required to form a dense region. By

default it is set to the number of data dimensions dim*2, or N/4, whichever is smaller. It may be

increased for large or noisy data sets.

The “Epsilon” parameter is the neighborhood radius. By default, it is set to 0.1 times the maximal

pairwise distance in the data set. The value can be optimized using the K-distance graph.

In the output table, cluster assignment is indicated by a positive integer. The value -1 indicates a

“noise point” that is not be assigned to any group.

K-distance graph: This graph shows the distance from each point to its MinPts-1 nearest neighbor,

sorted from smallest to largest. You should look for a break in the slope of this curve. The

corresponding “K-distance” may be used as a value for the parameter Epsilon.

Reference

Ester, M., Kriegel, H.-P., Sander, J., Xu, X. 1996. A density-based algorithm for discovering clusters in

large spatial databases with noise. Proceedings of the Second International Conference on Knowledge

Discovery and Data Mining (KDD-96), 226–231.

131

K-nearest neighbors classifier

K-nearest neighbours (Fix & Hodges 1951) is a simple but effective supervised machine learning

algorithm. A multivariate data set with a group column is supplied, where one part (the training set)

is assigned to groups, while the other part (the test set) has question marks in the group column. The

rows in the test set are then assigned to groups (classification) according to their similarity with rows

in the training set. Any distance measure can be used.

For each test row, the group is selected by a majority vote among the nearest k neighbors. The value

of k is a parameter selected by the user. Small k may lead to overfitting, while large k may lead to

oversmoothing. In addition, each vote among the nearest neighbors may be weighted according to

distance d from the test point. Four weighting schemes are available in Past:

•

No weighting. This is usually a poor choice (Gou et al. 2012) but is included for comparison.

Inverse weighting, w = 1/d. This is the default weighting in some programs, but it is usually

not optimal.

•

Linear weighting (Dudani 1976) is a standard weighting that usually performs well:

푑_ꢊꢖꢏ− 푑

ꢢ =

푑_ꢊꢖꢏ− 푑_ꢊꢄ푛

•

Weighting due to Gou et al. (2012), a modification of linear weighting that can perform

slightly better:

푑_ꢊꢖꢏ− 푑

푑_ꢊꢖꢏ+ 푑_ꢊꢄ푛

ꢢ =

×

푑_ꢊꢖꢏ− 푑_ꢊꢄ푛

푑_ꢊꢖꢏ+ 푑

Jackknifing

Past automatically performs a jackknifing (cross-validation) procedure on the training set. One row is

removed from the training set at a time and classified using the remaining training data. The

percentage of correctly classified rows is reported. This may be used to indicate good choices for k

and weighting.

References

Dudani, S.A. 1976. The distance-weighted k-nearest-neighbor rule. IEEE Transactions on Systems,

Man, and Cybernetics 6:325-327.

Fix, E. Hodges, J.L. 1951. Discriminatory analysis, nonparametric discrimination: consistency

properties. Technical Report No. 4, USAF School of Aviation Medicine, Randolf Field Texas, 238–247.

Gou, J., Du, L., Zhang, Y., Xiong, T. 2012. A new distance-weighted k-nearest neighbor classifier.

Journal of Information & Computational Science 9:1429–1436

132

Naïve Bayes classifier

Naïve Bayes (Hand & Yu 2001) is a simple supervised machine learning algorithm, like k-nearest

neighbors. A multivariate data set with a group column should be supplied, where one part (the

training set) is assigned to groups, while the other part (the test set) has question marks in the group

column. The rows in the test set are then assigned to groups (classification) according to their

similarity with rows in the training set.

The Gaussian model is used for continuous data. For this model, all within-group variances for all

variables must be non-zero. Therefore, a value is added to all zero-variances. This value is set to 0.1

times the variance of the given variable across all rows in the test and training sets. Classification is

usually not very good when this rule is applied.

The multinomial model is used for count data. Laplacian smoothing is used.

The Bernoulli model is used for binary data, with smoothing.

The multinomial and Bernoulli models are implemented according to Manning et al. (2008).

The “flat prior” option assigns the same value to the prior for all group assignments, instead of being

proportional to sample size.

Jackknifing cross-validation is implemented as for K-nearest neighbors.

Missing values supported by column average substitution.

References

Hand, D.J., Yu, K. 2001. Idiot’s Bayes – not so stupid after all? International Statistical Review 69:385–

398.

Manning, C.D., Raghavan, P., Schütze, H. 2008. Introduction to Information Retrieval. Cambridge

University Press.

133

Decision tree classification

A multivariate data set with a group column should be supplied, where one part (the training set) is

assigned to groups, while the other part (the test set) has question marks in the group column. The

rows in the test set are then assigned to groups (i.e., classified) according to their similarity with rows

in the training set.

The algorithm is similar to CART algorithm (Breiman et al. 1984) but with some modifications.

Initial tree and pre-pruning

Tree construction starts with a root node Q_mcontaining all samples (rows) in the training set. All

possible splitting variables and splitting values are traversed, searching for minimal impurity, defined

as

ꢀ_ꢊ^ꢞꢣꢜꢡ

ꢀ_ꢊ^{푟ꢄꢅℎꢡ}

ꢀ_ꢊ

=

퐻�푄_ꢊ^ꢞꢣꢜꢡ� +

퐻�푄_ꢊ^{푟ꢄꢅℎꢡ}

�

(

퐺 푄_ꢊ

)

ꢀ_ꢊ

where n_mis the number of samples in a node, while left and right refer to the two subnodes in the

tree that would result from the given splitting variable and value. We use Gini impurity for each

subnode:

(

)

(

)

퐻 푄_ꢊ= � ꢕ_ꢊ푘1 − ꢕ_ꢊ푘

푘

where p_mkis the proportion of samples of class k in the node. After the optimal splitting rule has been

found in this way, the sample set is assigned to the new subnodes accordingly, and the procedure is

repeated recursively for each subnode.

The recursion stops when there is a single sample in the node, or all samples in the node are assigned

to the same class. These termination criteria lead to a fully resolved tree, where each sample is

assigned to a class unambiguously. This can lead to overfitting, giving poorer quality classification of

new samples. The user can therefore set a stricter stopping rule (“pre-pruning”) by choosing a value

N>=2 for the minimum number of rows In a node to allow it to split, and/or the maximal depth of the

tree (the large default value of 10 will usually give a fully resolved tree).

Post-pruning

The initial tree computed above can be too detailed, leading to overfitting in the classification. The

tree can therefore be pruned further, using the “minimal cost-complexity” procedure of Breiman et

al. (1984), also formulated by Mingers (1989). For an internal node t, considered for pruning, we

calculate the error cost

( )

푅 푡 = ꢂ 푡 ꢕ(푡)

where r is the proportion of samples in the node that will be incorrectly classified if the node is

pruned (the error rate), and p is the total proportion of samples in the node relative to the complete

training set. Also compute the error cost for the subtree if the node is not pruned, as

(

)

푅 푇 = � 푅(푖)

ꢡ

ꢄ

where i are the leaves of the subtree T_trooted on t. Finally, compute the complexity parameter

134

( )

푅 푡 − 푅(푇 )

ꢡ

ꢒ =

ꢁ_ꢤ− 1

where N_Tis the number of leaves in the subtree. The node in the tree with the smallest α is selected

for pruning, and the procedure continues until α is greater than the “Pruning alpha” value. Setting

this value to zero will disable the post-pruning entirely.

Classification

Classification of a sample follows the decision rules of the tree down to a leaf node. The class is then

decided by a majority vote among the class memberships of the training samples in the node. The

“probability” value (a slight misnomer) is the proportion of samples of the majority class in the node.

The complete decision tree is shown in the “Rules” tab. Each node in the tree has two subnodes; the

first subnode is selected if the splitting rule is true, otherwise the second subnode is selected. You

can click on a node to hide its subtree.

Cross validation

The decision tree process is automatically cross-validated by the leave-one-out (jackknifing) method.

Missing values are treated as follows. During training, for each internal tree node, the optimal

splitting variable and its splitting point are chosen using only non-missing values. Then, the test set

samples in the node are assigned to the left or right subnode by this splitting rule only if they have

the appropriate non-missing values. The remaining samples in the node (those with missing values

for the splitting variable) are assigned to the subnode that acquired the most samples in the first

step. During classification, the decision tree is traversed as usual, except that when a decision cannot

be made because of a missing value, the subnode with the larger number of samples is chosen.

References

Breiman, L., Friedman, J., Olshen, R., Stone, C. 1984. Classification and Regression Trees. Wadsworth,

Belmont, CA.

Mingers, J. 1989. An empirical comparison of pruning methods for decision tree induction. Machine

Learning 4:227–243.

Random forest classifier

Random forest classification (Breiman 2001) in Past uses the same algorithm as decision trees except

there is no pre- or post-pruning – each tree is grown to its tips. To decrease sensitivity to noise in the

training set, a large number of randomized trees are generated, and classification is based on a

majority vote among these trees. There are two levels of randomization:

1. The rows in the training set are sampled randomly with replacement (bootstrapped). This is

also known as bagging.

2. During training, for every node, not all of the N variables are considered when searching for

the optimal splitting rule. Instead, a number M < N of variables is selected at random. Setting

M = N will disable this step. Typically, M is set to N/2 or smaller - √N is sometimes suggested.

Reference

Breiman, L. 2001. Random forests. Machine Learning 45:5–32.

135

Parsimony (cladistic)

Warning: the cladistics package in PAST is fully operational, but lacking in comprehensive

functionality. The PAST cladistics package is adequate for education and initial data exploration, but

for more ‘serious’ work we recommend a specialized program for systematics (Paup, TNT etc.).

The module requires a character matrix with taxa in rows, outgroup in first row.

Algorithms are from Kitching et al. (1998) and Felsenstein (2004).

Character states should be coded using integers in the range 0 to 255, or with the letters c, a, g, t, u

(upper or lower case). The first taxon is treated as the outgroup, and will be placed at the root of the

tree.

Missing values are coded with a question mark (?). Please note that PAST does not collapse zero-

length branches. Because of this, missing values can lead to a proliferation of equally shortest trees

ad nauseam, many of which are in fact equivalent.

Algorithms

There are four algorithms available for finding short trees:

Exhaustive

The exhaustive algorithm evaluates all possible trees. Like the branch-and-bound algorithm it will

necessarily find all shortest trees, but it is very slow. The maximum number of taxa allowed for

exhaustive search is 12, for which more than 600 million trees are evaluated. The only advantage

over branch-and-bound is the plotting of tree length distribution. This histogram may indicate the

‘quality’ of your matrix, in the sense that there should be a tail to the left such that few short trees

are ‘isolated’ from the greater mass of longer trees (but see Kitching et al. 1998 for critical comments

on this).

Branch-and-bound

The branch-and-bound algorithm is guaranteed to find all shortest trees. The total number of

shortest trees is reported, but a maximum of 10000 trees are saved. The branch-and-bound

algorithm can be very time consuming for data sets with more than 16 taxa.

Heuristic, nearest neighbour interchange

This heuristic algorithm adds taxa sequentially in the order they are given in the matrix, to the branch

where they will give least increase in tree length. After each taxon is added, all nearest neighbour

trees are swapped to try to find an even shorter tree.

Like all heuristic searches, this one is much faster than the algorithms above and can be used for

large numbers of taxa, but is not guaranteed to find all or any of the most parsimonious trees. To

decrease the likelihood of ending up on a suboptimal local minimum, a number of reorderings can be

specified. For each reordering, the order of input taxa will be randomly permuted and another

heuristic search attempted.

Note: Because of the random reordering, the trees found by the heuristic searches will normally be

different each time. To reproduce a search exactly, you will have to start the parsimony module

again from the menu, using the same value for “Random seed”. This will reset the random number

generator to the seed value.

136

Heuristic, subtree pruning and regrafting

This algorithm (SPR) is similar to the one above (NNI), but with a more elaborate branch swapping

scheme: A subtree is cut off the tree, and regrafting onto all other branches in the tree is attempted

in order to find a shorter tree. This is done after each taxon has been added, and for all possible

subtrees. While slower than NNI, SPR will often find shorter trees.

Heuristic, tree bisection and reconnection

This algorithm (TBR) is similar to the one above (SPR), but with an even more complete branch

swapping scheme. The tree is divided into two parts, and these are reconnected through every

possible pair of branches in order to find a shorter tree. This is done after each taxon is added, and

for all possible divisions of the tree. TBR will often find shorter trees than SPR and NNI, at the cost of

longer computation time.

Character types

Unordered

Characters are reversible and unordered, meaning that all changes have equal cost. This is the

criterion with fewest assumptions and is therefore generally preferable.

Ordered

Characters are reversible and ordered, meaning that 0->2 costs more than 0->1, but has the same

cost as 2->0.

Bootstrap

Bootstrapping is performed when the ‘Bootstrap replicates’ value is set to non-zero. The specified

number of replicates (typically 100 or even 1000) of your character matrix are made, each with

randomly weighted characters. The bootstrap value for a group is the percentage of replicates

supporting that group. A replicate supports the group if the group exists in the majority rule

consensus tree of the shortest trees made from the replicate.

Warning: Specifying 1000 bootstrap replicates will clearly give a thousand times longer computation

time than no bootstrap! Exhaustive search with bootstrapping is unrealistic and is not allowed.

Cladogram plotting

All shortest (most parsimonious) trees can be viewed, up to a maximum of 10000 trees. If

bootstrapping has been performed, a bootstrap value is given at the root of the subtree specifying

each group.

Character states can also be plotted onto the tree, as selected by the ‘Character’ menu. This

character reconstruction is unique only in the absence of homoplasy. In case of homoplasy, character

changes are placed as close to the root as possible, favouring one-time acquisition and later reversal

of a character state over several independent gains (so-called accelerated transformation).

137

The ‘Phylogram’ option allows plotting of trees where the length of vertical lines (joining clades) is

proportional to branch length.

Consistency index

The per-character consistency index (ci) is defined as m/s, where m is the minimum possible number

of character changes (steps) on any tree, and s is the actual number of steps on the current tree. This

index hence varies from one (no homoplasy) and down towards zero (a lot of homoplasy). The

ensemble consistency index CI is a similar index summed over all characters.

Retention index

The per-character retention index (ri) is defined as (g-s)/(g-m), where m and s are as for the

consistency index, while g is the maximal number of steps for the character on any cladogram (Farris

1989). The retention index measures the synapomorphy on the tree and varies from 0 to 1.

Note that in the present version, the retention index is only correctly calculated when using

unordered characters.

Consensus tree

The consensus tree of all shortest (most parsimonious) trees can also be viewed. Two consensus

rules are implemented: Strict (groups must be supported by all trees) and majority (groups must be

supported by more than 50% of the trees).

Bremer support (decay index)

The Bremer support for a clade is the number of extra steps you need to construct a tree (consistent

with the characters) where that clade is no longer present. There are reasons to prefer this index

rather than the bootstrap value. PAST does not compute Bremer supports directly, but for smaller

data sets it can be done ‘manually’ as follows:

•

Perform parsimony analysis with exhaustive search or branch-and-bound. Take note of the

clades and the length N of the shortest tree(s) (for example 42). If there are more than one

shortest tree, look at the strict consensus tree. Clades which are no longer found in the

consensus tree have a Bremer support value of 0.

•

In the box for ‘Longest tree kept’, enter the number N+1 (43 in our example) and perform a

new search.

Additional clades which are no longer found in the strict consensus tree have a Bremer

support value of 1.

For ‘Longest tree kept’, enter the number N+2 (44) and perform a new search. Clades which

now disappear in the consensus tree have a Bremer support value of 2.

Continue until all clades have disappeared.

References

Farris, J.S. 1989. The retention index and the rescaled consistency index. Cladistics 5:417-419.

Felsenstein, J. 2004. Inferring phylogenies. Sinauer Associates.

Kitching, I.J., P.L. Forey, C.J. Humphries & D.M. Williams. 1998. Cladistics. Oxford University Press.

138

Phylogenetically independent contrasts

Phylogenetically independent contrasts (Felsenstein 1985; Garland et al. 1992) is a method for

removing phylogenetic information from a character dataset, e.g. morphometric data. The purpose

of this is to reduce the phylogenetic interdependence of data points, which can violate the

assumptions of many statistical tests.

The module requires one or more columns of character data. In addition, the “tree collection” in Past

must contain one or more trees. The taxon names in the trees must match the row names in the

spreadsheet.

References

Felsenstein, J. 1985. Phylogenies and the comparative method. American Naturalist 125:1-15.

Garland Jr., T., Harvey, P.H., Ives, A.R. 1992. Procedures for the analysis of comparative data using

phylogenetically independent contrasts. Systematic Biology 41:18-32.

139

Phylogenetic generalized least squares (PGLS)

Bivariate linear regression (for two data columns), taking into account the expected phylogenetic

covariation from a Brownian model for evolution (Grafen 1989; Pagel 1999; Symonds & Blomberg

2014).

The “tree collection” in Past must contain one or more trees with branch lengths (otherwise, branch

lengths of 1 are assumed). The taxon names in the trees must match the row names in the

spreadsheet. IMPORTANT: The first row in the data is not used, as it is assumed to refer to an

outgroup. This (dummy) outgroup must be included in the tree – the branch length below the

outgroup is not used. This is to ensure compatibility with the parsimony trees generated by Past,

which are always rooted on the first row.

Standard errors on regression parameters are calculated according to Smaers & Rohlf (2016).

The PGLS var-covar matrix is transformed according to the parameter lambda (Symonds & Blomberg

2014), which can be selected as 0 (equivalent to OLS regression if the tree is ultrametric; a weighted

OLS otherwise), or 1 (full PGLS) or the maximum-likelihood estimate based on regression residuals

(Revell 2010).

The maximum-likelihood estimates for lambda (and their optimization profiles) are calculated for the

first data column only (x), the second data column only (y), a combination of x and y (Freckleton et al.

2002) or the regression residuals (Revell 2010).

References

Freckleton, R.P., Harvey, P.H., Pagel, M. 2002. Phylogenetic analysis and comparative data: A test and

review of evidence. American Naturalist 160:712-726.

Grafen, A. 1989. The phylogenetic regression. Philosophical Transactions of the Royal Society B

326:119–157.

Pagel, M. 1999. Inferring the historical patterns of biological evolution. Nature 401:877–884.

Revell, L.J. 2010. Phylogenetic signal and linear regression on species data. Methods in Ecology &

Evolution 1:319–329.

Smaers, J.B., Rohlf, F.J. 2016. Testing species' deviation from allometric predictions using the

phylogenetic regression. Evolution 70:1145–1149.

Symonds, M.R.E., Blomberg, S.P. 2014. A primer on phylogenetic least squares. Pp. 105-130 in: L. Z.

Garamszegi (ed.), Modern Phylogenetic Comparative Methods and their Application in Evolutionary

Biology. Springer-Verlag.

140

Multivariate normality

Multivariate normality is assumed by a number of multivariate tests. PAST computes Mardia's

multivariate skewness and kurtosis, with tests based on chi-squared (skewness) and normal (kurtosis)

distributions. A powerful omnibus (overall) test due to Doornik & Hansen (1994) is also given. If at

least one of these tests show departure from normality (small p value), the distribution is

significantly non-normal. Sample size should be reasonably large (>50), although a small-sample

correction is also attempted for the skewness test.

Missing data supported by column average substitution.

References

Doornik, J.A. & H. Hansen. 1994. An omnibus test for univariate and multivariate normality. W4&91

in Nuffield Economics Working Papers.

Mardia, K.V. 1970. Measures of multivariate skewness and kurtosis with applications. Biometrika

36:519-530.

141

Box’s M

Test for the equivalence of the covariance matrices for two multivariate samples marked with a

group column. This is a test for homoscedasticity, as assumed by MANOVA. You can use either two

original multivariate samples from which the covariance matrices are automatically computed, or

two specified variance-covariance matrices. In the latter case, you must also specify the sizes

(number of individuals) of the two samples.

The Box's M statistic is given together with a significance value based on a chi-square approximation.

Note that this test is supposedly very sensitive. This means that a high p value will be a good,

although informal, indicator of equality, while a highly significant result (low p value) may in practical

terms be a somewhat too sensitive indicator of inequality.

The statistic is computed as follows – note this equals the “-2 ln M” of some texts (Rencher 2002).

,

M = (n − 2)ln S − (n₁−1)ln S₁− (n₂−1)ln S₂

where S₁and S₂are the covariance matrices, S is the pooled covariance matrix, n=n₁+n₂and |•|

denotes the determinant.

The Monte Carlo test is based on 999 random permutations.

Missing data supported by column average substitution.

Reference

Rencher, A.C. 2002. Methods of multivariate analysis, 2^nded. Wiley.

142

MANOVA

One-way MANOVA (Multivariate ANalysis Of VAriance) is the multivariate version of the univariate

ANOVA, testing whether two or more groups (specified with a group column) have the same

multivariate mean.

Two statistics are provided: Wilk's lambda with its associated Rao's F and the Pillai trace with it's

approximated F. Wilk's lambda is probably more commonly used, but the Pillai trace may be more

robust.

Number of constraints: For correct calculation of the p values, the number of dependent variables

(constraints) must be specified. It should normally be left at 0, but for Procrustes fitted landmark

data use 4 (for 2D) or 6 (for 3D).

Pairwise comparisons (post-hoc): If the MANOVA shows significant overall difference between

groups, the analysis can proceed by pairwise comparisons. In PAST, the post-hoc analysis is simple, by

pairwise Hotelling's tests. The following values can be displayed in the table:

•

Hotelling's p values, not corrected for multiple testing. Marked in pink if significant (p<0.05).

The same p values, but significance (pink) assessed using the sequential Bonferroni scheme.

Bonferroni corrected p values (multiplied by the number of pairwise comparisons). The

Bonferroni correction gives very little power.

•

Squared Mahalanobis distances.

Note: These pairwise comparisons use the within-group covariance matrix pooled over all groups

participating in the MANOVA. They may therefore give slightly other results than if only two of the

groups are selected for analysis.

Missing data supported by column average substitution.

143

One-way ANOSIM

ANOSIM (ANalysis Of Similarities) is a non-parametric test of significant difference between two or

more groups, based on any distance measure (Clarke 1993). The distances are converted to ranks.

ANOSIM is normally used for taxa-in-samples data, where groups of samples are to be compared.

Items go in rows, variates in columns, and groups should be specified with a group column as usual.

In a rough analogy with ANOVA, the test is based on comparing distances between groups with

distances within groups. Let r_bbe the mean rank of all distances between groups, and r_wthe mean

rank of all distances within groups. The test statistic R is then defined as

r_b− r_w

.

R =

N

(N −1)

4

Large positive R (up to 1) signifies dissimilarity between groups. The one-tailed significance is

computed by permutation of group membership, with 9,999 replicates (can be changed).

Pairwise ANOSIMs between all pairs of groups are provided as a post-hoc test. Significant

comparisons (at p<0.05) are shown in pink. The optional Bonferroni correction multiplies the p values

with the number of comparisons. This correction is very conservative (produces large p values). The

sequential Bonferroni option does not output corrected p values, but significance is decided based

on step-down sequential Bonferroni, which is slightly more powerful than simple Bonferroni.

Missing data supported by pairwise deletion (not for the Raup-Crick, Rho and user-defined indices).

Reference

Clarke, K.R. 1993. Non-parametric multivariate analysis of changes in community structure.

Australian Journal of Ecology 18:117-143.

144

One-way PERMANOVA

PerMANOVA (Permutational MANOVA, also known as NPMANOVA) is a non-parametric test of

significant difference between two or more groups, based on any distance measure (Anderson 2001).

PerMANOVA is normally used for ecological taxa-in-samples data, where groups of samples are to be

compared, but may also be used as a general non-parametric MANOVA.

Items go in rows, variates in columns, and groups should be specified with a group column.

PerMANOVA calculates an F value in analogy with ANOVA. In fact, for univariate data sets and the

Euclidean distance measure, PerMANOVA is equivalent to ANOVA and gives the same F value.

The significance is computed by permutation of group membership, with 9,999 replicates (can be

changed by the user).

Repeated measures

A repeated measures (blocked) ANOVA is carried out with the “Repeated measures” box is ticked. In

this case, each group must have the same number of rows, and all the rows in each group must be

consecutive. The first row in each group is then assumed to belong to the first subject (block), the

second row in each group belongs to the second subject, etc. The calculations are analogous to the

univariate repeated measures ANOVA. First, the within-subjects sum of squares SS_wsis calculated in

the same way as the within-group sum of squares above, but with terms taken only when I and j are

within the same subject. The between-subjects sum of squares is then SS_bs= SS_T– SS_ws. Finally, the

error sum of squares SS_err= SS_wg– SS_bs. The F value is calculated as described for the several- sample

repeated-measures tests in the Univariate menu.

The permutations for the p value are carried out only within subjects.

145

Pairwise tests

Pairwise PerMANOVAs between all pairs of groups are provided as a post-hoc test. Significant

comparisons (at p<0.05) are shown in pink. The Bonferroni correction shown in the upper triangle of

the matrix multiplies the p values with the number of comparisons. This correction is very

conservative (produces large p values).

Missing data supported by pairwise deletion.

Reference

Anderson, M.J. 2001. A new method for non-parametric multivariate analysis of variance. Austral

Ecology 26:32-46.

146

Two-way ANOSIM

The two-way ANOSIM in PAST uses the crossed design (Clarke 1993). For more information see one-

way ANOSIM, but two group columns are required. There must be several rows (replication) for each

combination of group levels.

Reference

Clarke, K.R. 1993. Non-parametric multivariate analysis of changes in community structure.

Australian Journal of Ecology 18:117-143.

Two-way ANOSIM without replication

Input data as for two-way ANOSIM above, i.e. two group columns are required. There must be

exactly one row (no replication) for each combination of group levels.

Reference

Clarke, K.R. & Warwick, R.M. 1994. Similarity-based testing for community pattern: the two-way

layout with no replication. Marine Biology 118:167-176.

Two-way PERMANOVA

The two-way NPMANOVA (Anderson, 2001) in PAST uses the crossed design. The algorithm follows

Anderson (2001) with a modification to allow unbalanced designs (each combination of factors can

contain different numbers of values). For more information see one-way NPMANOVA, but two group

columns are required (as for two-way ANOSIM).

Reference

Anderson, M.J. 2001. A new method for non-parametric multivariate analysis of variance. Austral

Ecology 26:32-46.

147

Test for multivariate dispersion (PERMDISP)

Tests for equal dispersion (“spread”) in two or more groups of multivariate data (PERMDISP;

Anderson 2005).

The procedure starts from a standard Principal Coordinates Analysis of the complete data set (all

groups), using the selected distance measure. In the PCoA space, calculate the Euclidean distance z_i

from each point i to its group centroid. These distances are subjected to a standard one-way,

univariate ANOVA. The significance (p value) is estimated by a permutation test.

PAST also includes pairwise tests across groups. These tests are based on the original PCoA of the

complete data set (not PCoA on the reduced data set with only two groups).

A box plot of dispersions is provided, showing the z values for each group.

Reference

Anderson, M.J. 2005. Distance-based tests for homogeneity of multivariate dispersions. Biometrics

62:245-253.

148

Mantel test and partial Mantel test

The Mantel test (Mantel 1967, Mantel & Valand 1970) is a permutation test for correlation between

two distance or similarity matrices. In PAST, these matrices can also be computed automatically from

two sets of original data. The first matrix must be above the second matrix in the spreadsheet, and

the rows be specified as two groups (with a group column). The two matrices must have the same

number of rows. If they are distance or similarity matrices, they must also have the same number of

columns.

The R value is simply the Pearson’s correlation coefficient between all the entries in the two matrices

(because the matrices are symmetric it is only necessary to correlate the lower triangles). It ranges

from -1 to +1. The permutation test compares the original R to R computed in e.g. 9999 random

permutations. The reported p value is one-tailed.

In the example below, the first matrix (gpa) consists of Procrustes-fitted landmark coordinates from

primate skulls, while the second matrix (seq) contains sequence data from the same primates. The

user has selected the Euclidean measure for the first matrix, and Jukes-Cantor for the second. The

two data sets seem to be negatively correlated (R=-0.19), and there is no significant positive

correlation (the test is one-tailed). In other words, there is no correlation between morphology and

genetics.

149

Partial Mantel test

It is possible to add a third matrix C below the two matrices A and B as described above. This matrix

must be marked as above, and contain the same number of rows as A and B. A separate similarity

measure can then be selected for this matrix. If such a third matrix is included, the program will carry

out a partial Mantel test for the correlation of A and B, controlling for similarities given in C

(Legendre & Legendre 1998). Only matrix A is permutated, and the R value is computed as

R

(AB)− R(AC)

R

(BC)

R

(AB • C)

=

2

1− R(AC)

1− R(BC)

where R(AB) is the correlation coefficient between A and B.

References

Legendre, P. & L. Legendre. 1998. Numerical Ecology, 2nd English ed. Elsevier, 853 pp.

Mantel, N. 1967. The detection of disease clustering and a generalized regression approach. Cancer

Research 27:209-220.

Mantel, N. & R.S. Valand 1970. A technique of nonparametric multivariate analysis. Biometrics

26:547-558.

150

SIMPER

SIMPER (Similarity Percentage) is a simple method for assessing which taxa are primarily responsible

for an observed difference between groups of samples (Clarke 1993). The overall significance of the

difference is often assessed by ANOSIM. The Bray-Curtis similarity measure (multiplied with 100) is

most commonly used with SIMPER, but the Euclidean, cosine and chord measures can also be used.

If more than two groups are selected, you can either compare two groups (pairwise) by choosing

from the lists of groups, or you can pool all samples to perform one overall multi-group SIMPER. In

the latter case, all possible pairs of samples are compared using the Bray-Curtis measure. The overall

average dissimilarity is computed using all the taxa, while the taxon-specific dissimilarities are

computed for each taxon individually.

Samples go in rows, grouped with a group column, and taxa in columns. In the output table, taxa are

sorted in descending order of contribution to group difference. The last three columns show the

mean abundance in each of the groups.

Missing data supported by column average substitution.

Reference

Clarke, K.R. 1993. Non-parametric multivariate analysis of changes in community structure.

Australian Journal of Ecology 18:117-143.

151

Indicator species analysis (IndVal)

An alternative to SIMPER for identifying species indicative of given groups of sites (Dufrene &

Legendre 1997). Requires abundances (counts) with samples (sites) in rows, taxa in columns. Also a

group column with at least two groups specified.

For each species i in group j, define the specificity as

⁄

퐴_ꢄ푗= ꢁ

ꢄ푗

ꢁ

ꢄ

where N_ijis the mean number of individuals of species i across sites in group j, and N_iis the sum of

the mean numbers of individuals of species i over all groups. If N_i= 0, we set A_ijto 0.

Similarly, define the fidelity as

⁄

ꢍ_ꢄ푗= ꢁ푠푖푡푒푠_ꢄ푗ꢁ푠푖푡푒푠

푗

where Nsites_ijis the number of sites in group j where species i is present, and Nsites_jis the total

number of sites in group j.

The indicator value of species i in group j is then a value from 0 to 100 (percentage):

퐼ꢁꢥꢦ퐴퐿_ꢄ푗= 100퐴_ꢄ푗ꢍ_ꢄ푗

The statistical significances (p values) of the indicator values are estimated by 9999 random

reassignments (permutations) of sites across groups. The p values can optionally be Bonferroni

corrected (multiplied with the total number of indicator values, this is highly conservative).

Reference

Dufrene, M. & P. Legendre. 1997. Species assemblages and indicator species: The need for a flexible

asymmetrical approach. Ecological Monographs 67:345-366.

152

Paired Hotelling

The paired Hotelling's test expects two groups of multivariate data, marked with a group column.

Rows within each group must be consecutive. The first row of the first group is paired with the first

row of the second group, the second row is paired with the second, etc.

With n the number of pairs and p the number of variables:

Y_i= X_1i− X_2i

1

y =

Y

i

∑

n

i

1

T

S_y=

(Y − y)(Y_i− y)

∑

i

n −1

T ²= ny^TS⁻_y¹y

i

n − p

F =

T ²

p(n −1)

The F has p and n-p degrees of freedom.

For n<=16, the program also calculates an exact p value based on the T²statistic evaluated for all

possible permutations.

Missing data supported by column average substitution.

153

Modern Analog Technique

The Modern Analog Technique is a calibration method for reconstructing a past environmental

parameter (e.g. temperature) from faunal associations. It works by finding modern sites with faunal

associations close to those in downcore samples. Environmental data from the modern sites are then

used to estimate the environment downcore.

The (single) environmental variable, usually temperature, enters in the first column, and taxa in

consecutive columns. All the modern sites, with known values for the environmental variable, go in

the first rows, followed by all the downcore samples (these should have question marks in the

environmental column).

The plot on the first tab shows all the modern samples, with the observed temperature (for example)

versus the MAT reconstructed temperature using leave-one-out cross-validation (jackknifing).

Parameters to set:

•

Weighting: When several modern analogs are linked to one downcore sample, their

environmental values can be weighted equally, inversely proportional to faunal distance, or

inversely proportional to ranked faunal distance.

•

Distance measure: Several distance measures commonly used in MAT are available. "Squared

chord" has become the standard choice in the literature.

Distance threshold: Only modern analogs closer than this threshold are used. A default value

is given, which is the tenth percentile of distances between all sample pairs in the modern

data. The "Dissimilarity distribution" histogram may be useful when selecting this threshold.

N analogs: This is the maximum number of modern analogs used for each downcore sample.

Jump method (on/off): For each downcore sample, modern samples are sorted by ascending

distance. When the distance increases by more than the selected percentage, the

subsequent modern analogs are discarded.

•

154

Note that one or more of these options can be disabled by entering a large value. For example, a very

large distance threshold will never apply, so the number of analogs is decided only by the "N

analogs" value and optionally the jump method.

Cross validation

The scatter plot and R²value show the results of a leave-one-out (jackknifing) cross-validation within

the modern data. The y=x line is shown in red. This only partly reflects the "quality" of the method, as

it gives little information about the accuracy of downcore estimation.

Dissimilarity distribution

A histogram of all distances in the core-top (modern) data.

Semivariogram

Shows a semivariogram of variance in the environmental variable as a function of faunal difference.

Several semivariogram models can be fitted. This type of plot is familiar from spatial geostatistics, but

is also useful for MAT because it gives a good impression of the degree of “noise” in the faunal data

with respect to environmental prediction.

Reconstructions

Reconstruction of the paleoenvironmental values using MAT.

155

Weighted averaging calibration (WA)

Like the Modern Analogue Technique, WA is a method for reconstructing past environmental

parameters (temperature, pH) from a fossil assemblage, based on a training set of modern samples.

The (single) environmental variable, usually temperature, enters in the first column, and taxa in

consecutive columns. All the modern sites, with known values for the environmental variable, go in

the first rows, followed by all the downcore samples (these should have question marks in the

environmental column).

The plot on the first tab shows all the modern samples, with the observed temperature (for example)

versus the reconstructed temperature. This does not use cross-validation. The RMSE (Root Mean

Square Error) is based on these values.

WA in Past follows Birks et al. (1990), with RMA deshrinking.

Reference

Birks, H.J.B., Line, J.M., Juggins, S., Stevenson, A.C., ter Braak, C.J.F. 1990. Diatoms and pH

reconstruction. Philosophical Transactions of the Royal Society of London B 327:263-278.

156

Weighted averaging partial least squares (WA-PLS)

Like the Modern Analogue Technique and weighted averaging, WA-PLS is a method for

reconstructing past environmental parameters (temperature, pH) from a fossil assemblage, based on

a training set of modern samples. First described by ter Braak & Juggins (1993) and ter Braak et al.

(1993), WA-PLS is considered by some as the overall most accurate calibration method.

The (single) environmental variable, usually temperature, enters in the first column, and taxa in

consecutive columns. All the modern sites, with known values for the environmental variable, go in

the first rows, followed by all the downcore samples (these should have question marks in the

environmental column).

The plot on the first tab shows all the modern samples, with the observed temperature (for example)

versus the reconstructed temperature. This does not use cross-validation. The RMSE (Root Mean

Square Error) is based on these values.

The method is also cross-validated with the leave-one-out procedure (jackknifing), which is basis for

the RMSEP value (Root Mean Square Error of Prediction). The number of PLS components should be

set in order to minimize the RMSEP. Sometimes the minimal value is obtained already with only one

component, when the method is equivalent to two-way weighted averaging (WA).

Individual RMSEP values for each downcore sample are estimated with bootstrapping (2000

replicates). The bootstrapping procedure is somewhat complex and follows Birks et al. (1990).

Approximate 95% confidence intervals can be obtained as the predicted value plus/minus

1.96*RMSEP.

The implementation of WA-PLS in Past is based on the algorithm of ter Braak & Juggins (1993),

described below with some additional details and comments.

For the modern (training) set, we have x_ithe value of the measured environmental parameter in site

i, and the n x m matrix Y with y_ikthe abundance of taxon k in site i. There are n sites and m taxa.

Moreover, a ‘+’ replacing a subscript means summation over that subscript.

Step 0

Subtract the weighted mean from the environmental variable:

⁄

푥_ꢄ= 푥_ꢄ− � 푦_ꢄ+푥_ꢄ푦₊₊

ꢄ

Step 1

Take the centered environmental variable x_ias initial site scores r_i.

Do steps 2 to 7 for each PLS component p:

Step 2

Calculate new species scores u_k* by weighted averaging of the site scores:

∗

⁄

푢_푘= � 푦_ꢄ푘ꢂ_ꢄ푦_+푘

ꢄ

Step 3

157

Calculate new site scores r_iby weighted averaging of the species scores:

ꢂ = � 푦_ꢄ푘푢^∗푦_ꢄ+

⁄

ꢄ

푘

Step 4

For the first PLS component, go to step 5. For second and higher components, make the new site

scores r_iuncorrelated with previous components by orthogonalization, according to ter Braak (1987),

Table 5.2b.

Step 5

Take the site scores r_iand the species scores u_k* as the new PLS component consisting of two vectors

r^pand u^p. Note 1: In the original algorithm, site scores are standardized in Step 5. In the Past

implementation, this standardization is not carried out, in order to facilitate reconstruction for new

samples (ter Braak, pers. comm. 2019). Note 2: The species scores u^pare saved as well as the site

scores, as part of the PLS component.

Step 6

Do a weighted multiple regression of x_ion the components r obtained so far using weights y_i+/ y₊₊.

The regression coefficients are a₀… a_p. Take the fitted values as current estimates 푥_푖(as shown in the

̂

plot and used for calculating RMSE). Go to Step 2 with the residuals of the regression as the new site

scores r_i.

Reconstruction

After Steps 2-6 have been iterated the specified number of times, a full PLS model has been

constructed. Reconstruction of the environmental variable x₀from a new sample y_0kis then

computed as follows (in addition it must be remembered to add back the mean value subtracted in

Step 0).

First calculate updated species optima:

푢�_푘=푎₀+ � 푎_ꢑ푢^ꢑ

푘

ꢑ

Then the reconstructed x₀is calculated as the weighted sum

⁄

푥₀= � 푦_0푘푢�_푘푦₀₊

푘

References

Birks, H.J.B., Line, J.M., Juggins, S., Stevenson, A.C., ter Braak, C.J.F. 1990. Diatoms and pH

reconstruction. Philosophical Transactions of the Royal Society B 327: 263-278.

ter Braak, C.J.F. 1987. Ordination. Pp. 91-173 in: Jongman, R.H.G., ter Braak, C.J.F., van Tongeren,

O.F.R. (eds), Data Analysis in Community and Landscape Ecology, Pudoc.

ter Braak, C.J.F., Juggins, S., Birks, H.J.B., van der Voet, H. 1993. Weighted Averaging Partial Least

Squares regression (WA-PLS): definition and comparison with other methods for species-

158

environment calibration. Pp. 525-560 in: Patil, G.P. & Rao, C.R. (eds), Multivariate Environmental

Statistics, Elsevier.

ter Braak, C.J.F., Juggins, S. 1993. Weighted averaging partial least squares (WA-PLS): an improved

method for reconstructing environmental variables from species assemblages. Hydrobiology

269/270: 485-502.

159

Similarity and distance indices

Computes a number of similarity or distance measures between all pairs of rows. The data can be

univariate or (more commonly) multivariate, with variables in columns. The results are given as a

symmetric similarity/distance matrix. This module is rarely used, because similarity/distance matrices

are usually computed automatically from primary data in modules such as PCO, NMDS, cluster

analysis and ANOSIM in Past.

Euclidean

Basic Euclidean distance (the value is adjusted for missing data).

2

(

)

d _jk

=

x_ji− x_ki

∑

i

.

Gower

A distance measure that averages the difference over all variables, each term normalized for the

range of that variable:

x_ji− x_ki

1

d _jk

=

∑

n

max x_si− min x_si

i

s

.

The Gower measure is similar to Manhattan distance (see below) but with range normalization.

When using mixed data types (see below), this is the default measure for continuous and ordinal

data.

Chord

Euclidean distance between normalized vectors. Commonly used for abundance data. Can be written

as

x x

∑

ji ki

i

d _jk= 2 − 2

x²

ki

∑ ∑

ji

i

.

Manhattan

The sum of differences in each variable:

d _jk

=

x − x

∑

ji

ki

i

.

Bray-Curtis

Bray-Curtis is a popular similarity index for abundance data. Past calculates Bray-Curtis similarity as

follows:

160

x − x

∑

ji

ki

i

.

d _jk= 1−

(

)

x_ji+ x_ki

∑

i

This is algebraically equivalent to the form given originally by Bray and Curtis (1957):

(

)

min x , x

∑

ji

ki

i

d _jk= 2

.

(

)

x_ji+ x_ki

∑

i

Many authors operate with a Bray-Curtis distance, which is simply 1-d.

Cosine

The inner product of abundances each normalised to unit norm, i.e. the cosine of the angle between

the vectors.

x x

∑

ji ki

i

d _jk=

x²

ki

∑ ∑

ji

i

.

Morisita

For abundance data.

(

)

x x −1

∑

ji

i

λ₁=

λ₂=







x

x −1



∑ ∑

ji

^ji



i

x

_ki(x_ki−1)

∑

i





x

x −1





∑ ∑

ki

^ki



i

2

x x

ji ki

∑

i

d _jk

=

(λ₁+ λ₂

)

x

ki

∑ ∑

ji

i

.

Horn

Horn’s overlap index for abundance data (Horn 1966).

N =

x

∑

j

ji

i

N =

∑

k

ki

i

161

[

(x_ji+ x_ki)ln(x_ji+ x_ki

)

]

−

)

x ln x −

x ln x

ki ki

∑

ji

i

d _jk

=

(

) (

N _j+ N_kln N _j+ N_k− N _jln N _j− N_kln N_k

.

Mahalanobis

A distance measure taking into account the covariance structure of the data. With S the variance-

covariance matrix:

T

−1

(

)

(

)

d _jk

=

x_j− x_k

S x_j− x_k

.

Correlation

The complement 1-r of Pearson’s r correlation across the variables:

(

)

(

)

x_ji− x_jx_ki− x_k

∑

i

d _jk= 1−

2

(

)

(

)

x_ji− x_j

x_ki− x_k

∑

i

.

Taking the complement makes this a distance measure. See also the Correlation module, where

Pearson’s r is given directly and with significance tests.

Rho

The complement 1-r_sof Spearman’s rho, which is the correlation coefficient of ranks. See also the

Correlation module, where rho is given directly and with significance tests.

Dice

Also known as the Sorensen coefficient. For binary (absence-presence) data, coded as 0 or 1 (any

positive number is treated as 1). The Dice similarity puts more weight on joint occurences than on

mismatches.

When comparing two rows, a match is counted for all columns with presences in both rows. Using M

for the number of matches and N for the total number of columns with presence in just one row, we

have

d_jk= 2M / (2M+N).

Jaccard

A similarity index for binary data. With the same notation as given for Dice similarity above, we have

d_jk= M / (M+N).

Kulczynski

A similarity index for binary data. With the same notation as given for Dice similarity above (with N₁

and N₂referring to the two rows), we have

162

M

+

M + N₁

M + N₂

d _jk

=

2

.

Ochiai

A similarity index for binary data, comparable to the cosine similarity for other data types:

M

d _jk

=

M + N₁M + N₂

.

Simpson

The Simpson index is defined simply as M / N_min, where N_minis the smaller of the numbers of

presences in the two rows. This index treats two rows as identical if one is a subset of the other,

making it useful for fragmentary data.

Raup-Crick

Raup-Crick index for absence-presence data. This index (Raup & Crick 1979) uses a randomization

(Monte Carlo) procedure, comparing the observed number of species ocurring in both associations

with the distribution of co-occurrences from 1000 random replicates from the pool of samples.

Hamming

Hamming distance for categorical data as coded with integers (or sequence data coded as CAGT). The

Hamming distance is the number of differences (mismatches), so that the distance between (3,5,1,2)

and (3,7,0,2) equals 2. In PAST, this is normalised to the range [0,1], which is known to geneticists as

"p-distance".

Jukes-Cantor

Distance measure for genetic sequence data (CAGT). Similar to p (or Hamming) distance, but takes

into account probability of reversals:

3

4

3







d = − ln 1−

p



4





Kimura

The Kimura 2-parameter distance measure for genetic sequence data (CAGT). Similar to Jukes-Cantor

distance but takes into account different probabilities of nucleotide transitions vs. transversions

(Kimura 1980). With P the observed proportion of transitions and Q the observed number of

transversions, we have

1

d = − ln(1− 2P − Q)− ln(1− 2Q)

2

4

.

Tajima-Nei

Distance measure for genetic sequence data (CAGT). Similar to Jukes-Cantor distance, but does not

assume equal nucleotide frequencies.

163

Tamura

Distance measure for genetic sequence data (CAGT). An extension of the Kimura 2-parameter

distance, handling unequal transition/transversion probability, but also taking into account a possible

bias in the G+C frequency. With P and Q as for Kimura distance, and h = 2θ(1- θ) where θ is the G+C

frequency (0-1):

푃

ℎ

1

2

( ) ( )

1 − ℎ ln 1 − 2푄

푑 = −ℎln �1 −

− 푄� −

Geographical

Distance in meters along a great circle between two points on the Earth’s surface. Exactly two

variables (columns) are required, with latitudes and longitudes in decimal degrees (e.g. 58 degrees 30

minutes North is 58.5). Coordinates are expected in the WGS84 datum, and distance is calculated

with respect to the WGS84 ellipsoid. Use of other datums will result in very slight errors.

The accuracy of the algorithm used (Vincenty 1975) is on the order of 1 mm with respect to WGS84.

User-defined similarity

Expects a symmetric similarity matrix rather than original data. No error checking!

User-defined distance

Expects a symmetric distance matrix rather than original data. No error checking!

Mixed

This option requires that data types have been assigned to columns (see Entering and manipulating

data). A pop-up window will ask for the similarity/distance measure to use for each datatype. These

will be combined using an average weighted by the number of variates of each type. The default

choices correspond to those suggested by Gower, but other combinations may well work better. The

"Gower" option is a range-normalised Manhattan distance.

All-zeros rows: Some similarity measures (Dice, Jaccard, Simpson etc.) are undefined when

comparing two all-zero rows. To avoid errors, especially when bootstrapping sparse data sets, the

similarity is set to zero in such cases.

Missing data: Most of these measures treat missing data (coded as ‘?’) by pairwise deletion, meaning

that if a value is missing in one of the variables in a pair of rows, that variable is omitted from the

computation of the distance between those two rows. The exceptions are rho distance, using

column average substitution, and Raup-Crick, which does not accept missing data.

164

References

Bray, J.R. & J.T. Curtis. 1957. An ordination of the upland forest communities of Southern Wisconsin.

Ecological Monographs 27:325-349.

Horn, H.S. 1966. Measurement of overlap in comparative ecological studies. American Naturalist

100:419-424.

Kimura, M. 1980. A simple model for estimating evolutionary rates of base substitutions through

comparative studies of nucleotide sequences. Journal of Molecular Evolution 16:111-120.

Raup, D. & R.E. Crick. 1979. Measurement of faunal similarity in paleontology. Journal of

Paleontology 53:1213-1227.

Simpson, G.G. 1943. Mammals and the nature of continents. American Journal of Science 241:1-31.

Vincenty, T. 1975. Direct and inverse solutions of geodesics on the ellipsoid with application of

nested equations. Survey Review 176:88-93.

165

Genetic sequence stats

A number of simple statistics on genetic sequence (DNA or RNA) data. The module expects a number

of rows, each with a sequence. The sequences are expected to be aligned and of equal length

including gaps (coded as ‘?’). Some of these statistics are useful for selecting appropriate distance

measures elsewhere in Past.

Total length:

The total sequence length, including gaps, of one sequence

The number of positions with variable states

Polymorphic sites:

Average gap:

The number of gap positions, averaged over all sequences

The average number of positions containing each nucleotide

Average A, T/U, C, G:

Average p distance:

The p distance between two sequences, averaged over all pairs of

sequences. The p (or Hamming) distance is defined as the proportion

of unequal positions

Average Jukes-Cantor d:

The Jukes-Cantor d distance between two sequences, averaged over

all pairs of sequences. d = -3ln(1 - 4p/3)/4, where p is the p distance

Maximal Jukes-Cantor d:

Maximal Jukes-Cantor distance between any two sequences

Average transitions (P):

Average number of transitions (a↔g, c↔t, i.e. within purines,

pyrimidines)

Average transversions (Q):

Average number of transversions (a↔t, a↔c, c↔g, t↔g, i.e.

across purines, pyrimidines)

R=P/Q:

The transition/transversion ratio

Missing data: Treated as gaps.

166

Model menu

Linear, bivariate

If two columns are selected, they represent x and y values, respectively. If one column is selected, it

represents y values, and x values are taken to be the sequence of positive integers (1,2,...). A straight

line y=ax+b is fitted to the data. Several bivariate data sets can be regressed in the same plot, and

their slopes compared, by giving an even number of columns, each pair of columns being one x-y set.

Finally, new values can be predicted by entering the x value but giving a ‘?’ for the y value.

There are five different algorithms available: Ordinary Least Squares (OLS), Reduced Major Axis

(RMA), Major Axis (MA), Robust, and Prais-Winsten. OLS regression assumes the x values are fixed,

and finds the line which minimizes the squared errors in the y values. Use this if your x values have

very little error associated with them. RMA and MA try to minimize both the x and the y errors.

RMA/MA fitting, standard error estimation and slope comparison are according to Warton et al.

(2006).

The “Robust” method is an advanced Model I (fixed x values) regression which is robust to outliers. It

sometimes gives strange results, but can be very successful in the case of “almost” normally

distributed errors but with some far-off values. The algorithm is “Least Trimmed Squares” based on

the “FastLTS” code of Rousseeuw & Driessen (1999). Parametric error estimates are not available,

but Past gives bootstrapped confidence intervals on slope and intercept (beware – this is extremely

slow for large data sets).

Prais-Winsten regression (e.g. Wooldridge 2012, ch. 12) is appropriate for data with serially

correlated residuals, typically time series. The fitted model is a sum of a linear function and an AR(1)

autoregressive process with autocorrelation rho. An iterative procedure is used, with a tolerance on

rho of 0.001 and a maximum of 10 iterations. Bootstrapping is not carried out as it would violate the

serial correlation.

167

Both x and y values can be log-transformed (base 10), in effect fitting your data to the 'allometric'

function y=10^bx^a. An a value around 1 indicates that a straight-line ('isometric') fit may be more

applicable.

The values for a and b, their errors, Pearson's r correlation, and the probability that the columns are

not correlated are given. Note the r²is simply the Pearson’s coefficient squared – it does not adjust

for regression method.

The calculation of standard errors for slope and intercept assumes normal distribution of residuals

and independence between the variables and the variance of residuals. If these assumptions are

strongly violated, it is preferable to use the bootstrapped 95 percent confidence intervals (1999

replicates).

The permutation test on correlation (r²) uses 9,999 replicates.

Confidence band for the regression

In OLS regression (not RMA/MA/Robust/Prais-Winsten), a 95 percent "Working-Hotelling"

confidence band for the fitted line is available. The confidence band is calculated as

2









1

(x − x)

∑

CI = b + ax ± t_{0.05 / 2,n−2}SE²

+

^reg







2

n

(x_i− x)

2

SE_r²_eg

=

(y_i− b − ax_i

)

where the squared sum of residuals

.

∑

When the intercept is forced to zero, the confidence band is calculated as

168

x²

CI = ax ± t_{0.05 / 2,n−1}SE_r²_eg

.

x²

∑

i

Confidence band for the forecast (prediction)

In OLS regression, a 95 percent confidence band for forecasting is also given. The confidence band is

calculated as follows. First calculate the standard error of the estimate and the standard error of the

estimate of the mean:

2

∑(

)

푦_ꢄ− 푏 − 푎푥_ꢄ

�

ꢋ퐸퐸 =

ꢀ − 2

2

(

)

ꢋ퐸퐸

푥 − 푥

̅

( )

ꢋ퐸_ꢊꢣꢖ푛푥 =

1 +

�

1

ꢀ

√

2

∑(

푥_ꢄ− 푥

)

Then,

2

�

( )

ꢉ퐼 = 푏 + 푎푥 ± 푡_0.05

ꢋ퐸퐸 + �ꢋ퐸_ꢊꢣꢖ푛푥 �

,푛−2

2

Zero intercept

Forces the regression line through zero. This has implications also for the calculation of slope and the

standard error of the slope. All five methods handle this option.

Residuals

The Residuals window reports the distances from each data point to the regression line, in the x and

y directions. Only the latter is of interest when using ordinary linear regression rather than RMA or

MA. The residuals can be copied back to the spreadsheet and inspected for normal distribution and

independence between independent variable and residual variance (homoskedasticity).

Durbin-Watson test

The Durbin-Watson test for positive autocorrelation of residuals in y (violating an assumption of OLS

regression) is given in the Residuals window. The test statistic varies from zero (total positive

autocorrelation) through 2 (zero autocorrelation) to 4 (negative autocorrelation). For n<=400, an

exact p value for no positive autocorrelation is calculated using the PAN algorithm (Farebrother 1980,

with later corrections). The test is not accurate when using the Zero intercept option.

169

Breusch-Pagan test

The Breusch-Pagan test for heteroskedasticity, i.e. nonstationary variance of residuals (violating an

assumption of OLS regression) is given in the Residuals window. The test statistic is LM = nr²where r

is the correlation coefficient between the x values and the squared residuals. It is asymptotically

distributed as χ²with one degree of freedom. The null hypothesis of the test is homoskedasticity.

Exponential functions

Your data can be fitted to an exponential function y=e^be^axby first log-transforming just your y column

(in the Transform menu) and then performing a straight-line fit.

Prediction (forecasting)

Rows with a ‘?’ for the y value will be included in the table under the ‘Prediction’ tab. The predicted y

value is calculated for the given x, together with a 95% prediction interval calculated as above

(confidence band for the forecast). If the ‘log-log’ option was selected, the back-transformed

prediction and interval will also be given for convenience. Note that this prediction is only strictly

valid for the OLS model, but will be approximately correct also for the RMA and MA models.

RMA equations

Slope

2

(y − y)

∑

.

a = sign(

r

)

2

(x − x)

1− r²

a = abs(

a

)

Standard error on

.

n − 2

Intercept

.

b = y − ax

s_r²

n

b =

+ x²s_a²

Standard error on

, where s_ris the estimate of standard deviation of residuals and s_ais

the standard error on slope.

x = 0

For zero intercept (b=0), set

and

for the calculation of slope and its standard error

y = 0

(including the calculation of r therein), and use n-1 instead of n-2 for the calculation of standard

error.

170

Missing data: Supported by row deletion.

References

Farebrother, R.W. 1980. Pan's procedure for the tail probabilities of the Durbin-Watson statistic.

Applied Statistics 29:224–227.

Rousseeuw, P.J. & van Driessen, K. 1999. Computing LTS regression for large data sets. Institute of

Mathematical Statistics Bulletin.

Warton, D.I., Wright, I.J., Falster, D.S. & Westoby, M. 2006. Bivariate line-fitting methods for

allometry. Biological Review 81:259-291.

Wooldridge, J.M. 2012. Introductory Econometrics – a Modern Approach (5th ed.). South-Western

Cengage Learning.

171

Linear, multivariate (one independent, n dependent)

When you have one independent variate and several dependent variates, you can fit each dependent

variate separately to the independent variate using simple linear regression. This module makes the

process more convenient by having a scroll button going through each dependent variate.

The module expects two or more columns of measured data, with the independent in the first

column and the dependents in consecutive columns.

In addition, an overall MANOVA test of multivariate regression significance is provided. The Wilks'

lambda test statistic is computed as the ratio of determinants

E

Λ =

,

E + H

where E is the error (residuals) sum of squares and crossproducts, and H is the hypothesis

(predictions) sum of squares and crossproducts. The Rao’s F statistic is computed from the Wilks’

lambda and subjected to a one-tailed F test (see ‘Linear, n independent, n dependent’ below).

Missing data supported by column average substitution.

Regression for geometric morphometrics

For Procrustes-fitted landmarks or Elliptic Fourier coefficients as the dependent variables, see the

Geometry menu for regression with visualization of shape change.

172

Linear, multiple (one dependent, n independent)

Requires two or more columns of measured data, with the dependent in the first column and the

independents in consecutive columns.

The program will present the multiple correlation coefficient R and R², together with the "adjusted"

R²and an overall ANOVA-type significance test.

With SSR the regression sum of squares, SSE the error (residuals) sum of squares, n the number of

points and k the number of independent variates, we have R²=SSR/SST,

(

1− R²

)

(n −1)

R_a²_dj= 1−

,

n − k −1

SSR k

F =

.

SSE (n − k −1)

The coefficients (intercept, and slope for each independent variate) are presented with their

estimated standard errors and t tests.

Missing data supported by column average substitution.

173

Linear, multivariate multiple (m independent, n dependent)

Requires two or more columns of measured data, with the dependent variables in the first column(s)

and the independents in consecutive columns. The program will ask for the number of dependent

variables. The output consists of four main parts.

Overall MANOVA

An overall test of multivariate regression significance. The Wilks' lambda test statistic is computed as

the ratio of determinants

E

Λ =

,

E + H

where E is the error (residuals) sum of squares and crossproducts, and H is the hypothesis

(predictions) sum of squares and crossproducts.

The Rao’s F statistic is computed from the Wilks’ lambda. With n the number of rows, p the number

of dependent variables and q the number of independent variables, we have:

1

m = n − q −1− (p − q +1)

2

p²q²− 4



if p²+ q²− 5 > 0



p²+ q²− 5

τ =





1

otherwise



1− Λ^{1 τ}mτ +1− pq 2

F =

⋅

Λ^{1 τ}

pq

The F test has pq and mτ + 1- pq/2 degrees of freedom.

Tests on independent variables

The test for the overall effect of each independent variable (on all dependent variables) is based on a

similar design as the overall MANOVA above, but comparing the residuals of regression with and

without the independent variable in question.

Tests on dependent variables

See ‘Linear, n independent, one dependent’ above for details of the ANOVA tests for the overall

effect of all independent variables on each dependent.

Regression coefficients and statistics

The complete set of coefficients and their significances for all combinations of independent and

dependent variables.

Missing data supported by column average substitution.

174

Generalized Linear Model

This module computes a basic version of the Generalized Linear Model, for a single explanatory

variable. It requires two columns of data (independent and dependent variables). For proportion

data, a third column with total counts can be included.

GLM allows non-normal distributions, and also “transformation” of the model through a link

function. Some particularly useful combinations of distribution and link function are:

Normal distribution and the identity link: This is equivalent to ordinary least squares linear regression.

Normal distribution and the reciprocal link: Fit to the function y=1/(ax+b).

Normal or gamma distribution and the log link: Fit to the function y=exp(ax+b).

Poisson distribution and the log link: Fit to the function y=exp(ax+b), for counts

Binomial distribution and the logit link: Logistic regression for a binary response variable (see figure

above). A third data column with sample sizes (total count) can be included, if so then the binomial

distribution can be used for proportion data (0-1).

Technical details

The program uses the Iteratively Reweighted Least Squares (IRLS) algorithm for maximum likelihood

estimation.

175

The dispersion parameter φ, which is used only for the inference, not the parameter estimation, is

fixed at φ =1 for the Poisson and binomial distributions, and estimated using Pearson’s chi-square for

the normal and gamma distributions.

D

LL = −

The log-likelihood LL is computed from the deviance D by

.

2ϕ

The deviance is computed as follows:

2

Normal:

D =

(y_i− µ_i

)

∑

i





y_i

y_i− µ

ⁱ



Gamma:

Binomial:

Poisson:

D = 2

− ln

+

∑





µ_i

i





y_i

1− y

(the first term defined as zero if y_i=0)





ⁱ

y_iln

+

(

1− y_i)ln

∑

µ_i

1− µ_i

i







y_i

y_iln

−

(y_i− µ_i

)

∑









µ_i

i

The G statistic is the difference in D between the full model and an additional GLM run where only

the intercept is fitted. G is approximately chi-squared with one degree of freedom, giving a

significance for the slope.

176

Polynomial regression

Two columns must be selected (x and y values). A polynomial of up to the fifth order is fitted to the

data. The algorithm is based on a least-squares criterion and singular value decomposition (Press et

al. 1992), with mean and variance standardization for improved numerical stability.

The polynomial is given by

y = a₅x⁵+ a₄x⁴+ a₃x³+ a₂x²+ a₁x + a₀

.

The chi-squared value is a measure of fitting error - larger values mean poorer fit. The Akaike

Information Criterion has a penalty for the number of terms. The AIC should be as low as possible to

maximize fit but avoid overfitting.

R²is the coefficient of determination, or proportion of variance explained by the model. Finally, a p

value, based on an F test, gives the significance of the fit.

Reference

Press, W.H., S.A. Teukolsky, W.T. Vetterling & B.P. Flannery. 1992. Numerical Recipes in C. Cambridge

University Press.

177

Nonlinear

Attempts to fit two columns of x-y data to a number of nonlinear equations, using least squares.

Select a function name from the list. To see more functions, grab a function name and drag up and

down to scroll.

The 95% confidence intervals are based on 1999 bootstrap replicates.

Fitting to a nonlinear function can be a bit tricky. For most of the functions, Past uses an educated

guess for the parameters, followed by Levenberg-Marquardt optimization.

The Akaike Information Criterion (AIC) may aid in the selection of model. Lower values for the AIC

imply a better fit, adjusted for the number of parameters.

Linear

y = ax + b

Included for comparison with the nonlinear functions. Fitting by ordinary least squares regression.

The “Zero constant” option will set b=0.

Quadratic

y = ax²+ bx + c

Included for reference. Fitting by least-squares and SVD (the equation is linear in its coefficients). The

“Zero constant” option will set c=0. See also the Polynomial Model module.

Power

y = ax^b+ c

178

The usual power law equation. Initial guess by log-log transformation and linear regression (i.e. c =

0), followed by nonlinear optimization. The “Zero constant” option will set c=0.

Exponential

y = ae^bx+ c

Initial guess by linearization (log-transforming y), followed by nonlinear optimization. The “Zero

constant” option will set c=0. See also the Generalized Linear Model module.

Von Bertalanffy

y = a

(

1− be^−cx

)

This equation is used for modelling growth of multi-celled animals (Brown & Rothery 1993). It is

sometimes given in a slightly different form:

(x−t₀

)

y = L_∞

(

1− e^−K

)

t₀= (lnb)

c

It is easy to see that

, K = c and

.

L_∞= a

The value of a is first estimated by the maximal value of y, and b and c using a straight-line fit to a

linearized model. Finally nonlinear optimization.

Michaelis-Menten

ax

y =

b + x

The Michaelis-Menten curve can make accurate fits to rarefaction curves, and may therefore

(somewhat controversially) be used for extrapolating these curves to estimate biodiversity (Colwell &

Coddington 1994). It is also an important model equation for chemical kinetics.

The algorithm uses maximum-likelihood estimators for the so-called Eadie-Hofstee transformation

(Raaijmakers 1987; Colwell & Coddington 1994), followed by nonlinear optimization.

Logistic

a

y =

1+ be^−cx

A sigmoidal (S-shaped) curve. The logistic equation can model growth with saturation (Brown &

Rothery 1993), and was used by Sepkoski (1984) to describe the proposed stabilization of marine

diversity in the late Palaeozoic.

The value of a is first estimated by the maximal value of y, and b and c using a straight-line fit to a

linearized model. Finally nonlinear optimization. See also the Generalized Linear Model module.

Gompertz

179

cx

y = ae^be

Initial estimate is computed using regression on a linearized model , followed by nonlinear

optimization.

Gaussian

2

(x−b)

y = ae⁻

2c²

The ‘bell curve’ with mean b and standard deviation c. Initial guess of a by maximal value of y, b by

weighted mean, and c=1, followed by nonlinear optimization.

Hill’s equation (4-parameter logistic)

푎 − 푑

푦 = 푑 +

푐

푥

1 + � �

푏

This sigmoidal function is often used to model dosage-response data. d is the minimum and a the

maximum asymptote. b is the dosage at which 50% of subjects show the response (the IC₅₀value),

while c is the “Hill slope”. The “Zero constant” option will set d=0.

(In previous versions of Past, a slightly different form was used, with b/x instead of x/b and

consequently c had opposite sign).

References

Brown, D. & P. Rothery. 1993. Models in biology: mathematics, statistics and computing. John Wiley

& Sons.

Colwell, R.K. & J.A. Coddington. 1994. Estimating terrestrial biodiversity through extrapolation.

Philosophical Transactions of the Royal Society of London B 345:101-118.

Raaijmakers, J.G.W. 1987. Statistical analysis of the Michaelis-Menten equation. Biometrics 43:793-

803.

Sepkoski, J.J. 1984. A kinetic model of Phanerozoic taxonomic diversity. Paleobiology 10:246-267.

180

Sinusoidal regression

Two columns must be selected (x and y values). A sum of up to eight sinusoids with periods specified

by the user, but with unknown amplitudes and phases, is fitted to the data. This can be useful for

modeling periodicities in time series, such as annual growth cycles or climatic cycles, usually in

combination with spectral analysis. The algorithm is based on a least-squares criterion and singular

value decomposition. By default, the periods are set to the range of the x values, and harmonics (1/2,

1/3, 1/4, 1/5, 1/6, 1/7 and 1/8 of the fundamental period). These values can be changed, and need

not be in harmonic proportion.

The “Fit periods” option will sequentially optimize the period of each sinusoid (over the full

meaningful range from one period to the Nyquist frequency), after subtracting all previously fitted

sinusoids. This is a simple example of the “Matching pursuit” algorithm. The algorithm is slow but

robust and will fairly reliably find the global optimum.

The chi-squared value is a measure of fitting error - larger values mean poorer fit. The Akaike

Information Criterion has a penalty for the number of sinusoids (the equation used assumes that the

periods are estimated from the data). The AIC should be as low as possible to maximize fit but avoid

overfitting.

R²is the coefficient of determination, or proportion of variance explained by the model. Finally, a p

value, based on an F test, gives the significance of the fit.

181

It is not meaningful to specify periodicities that are smaller than two times the typical spacing of data

points.

Each sinusoid is given by y=a*cos(2*pi*(x-x₀) / T - p), where a is the amplitude, T is the period and p is

the phase. x₀is the first (smallest) x value. An overall constant offset (mean) is also given.

There are also options to enforce a pure sine or cosine series, i.e. with fixed phases.

182

Smoothing spline

Two columns must be selected (x and y values). The data are fitted to a smoothing spline, which is a

sequence of third-order polynomials continuous up to the second derivative. A typical application is

the construction of a smooth curve going through a noisy data set. The algorithm follows de Boor

(2001). Sharp jumps in your data can give rise to oscillations in the curve, and you can also get large

excursions in regions with few data points. Multiple data points at the same X value are collapsed to

a single point by weighted averaging and calculation of a combined standard deviation.

An optional third column specifies standard deviations on the data points. These are used for

weighting the data. If unspecified, they are all set to 10% of the standard deviation of the Y values.

The smoothing value set by the user is a normalized version of the smoothing factor of de Boor

(default 1). Larger values give smoother curves. A value of 0 will start a spline segment at every point.

Clicking "Optimize smoothing" will calculate an "optimal" smoothing by a cross validation procedure.

"View given points" gives a table of the given data points X, Y and stdev(Y), the corresponding Y

values on the spline curve (ys) and the residuals. The chi-squared test for each point may be used to

identify outliers. The final column suggests an stdev(Y) value to use if forcing the p value to 0.5.

An optional fourth input column (if used then the third column must also be filled with stdev values)

may contain a different number of values from the previous columns. It contains X values to be used

for interpolation between the data points. Optional columns 5-7 contain lower and upper limits for X

values (rectangular distribution) and standard deviation for Y values (normal distribution), to be used

by bootstrapping (Monte Carlo) simulation providing error bars for the interpolated values. These

functions are included mainly for computing boundary ages for the geological time scale.

Reference

de Boor, Carl. 2001. A practical guide to splines. Springer.

183

LOESS smoothing

Two columns must be selected (x and y values). The algorithm used is “LOWESS” (LOcally WEighted

Scatterplot Smoothing; Cleveland 1979, 1981), with its recommended default parameters (including

two robustness iterations). Given a number of points n and a smoothing parameter q specified by the

user, the program fits the nq points around each given point to a straight line, with a weighting

function decreasing with distance. The new smoothed point is the value of the fitted linear function

at the original x position.

The Bootstrap option will estimate a 95% confidence band for the curve based on 999 random

replicates. In order to retain the structure of the interpolation, the procedure uses resampling of

residuals rather than resampling of original data points.

The Optimize smoothing option will suggest an “optimal” smoothing factor by minimizing the sum of

squared prediction errors through a 5-fold cross validation. The smoothing factor is constrained to

the range 0.15 to 0.95. The resulting value will tend towards the lower limit (0.15) for smooth

(serially autocorrelated) data. As LOESS is primarily a technique for improving visual interpretation,

the purpose of such automatic selection of the smoothing parameter is debatable.

184

LOESS or smoothing spline?

This is almost a matter of taste. Compare the curves above, for the same dataset. The spline often

gives a more aesthetically pleasing curve because of its continuous derivatives but can suffer from

overshooting near sharp bends in the data.

References

Cleveland, W.S. 1979. Robust locally weighted fitting and smoothing scatterplots. Journal of the

American Statistical Association 74:829-836.

Cleveland, W.S. 1981. A program for smoothing scatterplots by robust locally weighted fitting. The

American Statistician 35:54.

185

Mixture analysis

Mixture analysis is a maximum-likelihood method for estimating the parameters (mean, standard

deviation and proportion) of two or more univariate normal distributions, based on a pooled sample.

The program can also estimate mean and proportion of exponential, Poisson, and lognormal

distributions. For example, the method can be used to study differences between sexes (two groups),

or several species, or size classes, when no independent information about group membership is

available.

The program expects one column of univariate data, assumed to be taken from a mixture of normally

distributed populations (or exponential or Poisson). In the example below, sizes of two brachiopod

samples have been pooled into one sample. The means, standard deviations and proportions of the

two original samples have been almost perfectly recovered.

PAST uses the EM algorithm (Dempster et al. 1977), which can get stuck on a local optimum. The

procedure is therefore automatically run 20 times, each time with new, random starting positions for

the means. The starting values for standard deviation are set to s/G, where s is the pooled standard

deviation and G is the number of groups. The starting values for proportions are set to 1/G. The user

is still recommended to run the program a few times to check for stability of the solution ("better"

solutions have less negative log likelihood values).

The Akaike Information Criterion (AIC; Akaike 1974) is calculated with a small-sample correction:

2k(k +1)

AICc = 2k − 2ln L +

n − k −1

186

where k is the number of parameters, n the number of data points and L the likelihood of the model

given the data. A minimal value for AIC indicates that you have chosen the number of groups that

produces the best fit without overfitting.

It is possible to assign each of the data points to one of the groups with a maximum likelihood

approach. This can be used as a non-hierarchical clustering method for univariate data. The

“Assignments” button will open a window where the value of each probability density function is

given for each data point. The data point can be assigned to the group that shows the largest value.

Missing data: Supported by deletion.

References

Akaike, H. 1974. A new look at the statistical model identification. IEEE Transactions on Automatic

Control 19: 716-723.

Dempster, A.P., Laird, N.M. & Rubin, D.B. 1977. Maximum likelihood from incomplete data via the

EM algorithm". Journal of the Royal Statistical Society, Series B 39:1-38.

187

Abundance models

This module can be used for plotting taxon abundances in descending rank order on a linear or

logarithmic (Whittaker plot) scale, or number of species in abundance octave classes (as shown when

fitting to log-normal distribution). Taxa go in rows, samples (often only one) in columns. It can also fit

the data to one of four different standard abundance models:

•

Geometric, where the 2nd most abundant species should have a taxon count of k<1 times

the most abundant, the 3rd most abundant a taxon count of k times the 2nd most abundant

etc. for a constant k. With n_ithe count of the ith most abundant taxon, we have n_i= n₁kⁱ⁻¹

.

This will give a straight descending line in the Whittaker plot. Fitting is by simple linear

regression of the log abundances.

•

Log-series, with two parameters alpha and x. The fitting algorithm is from Krebs (1989). The

number of species with n individuals (this equation does not translate directly to the

Whittaker plot representation):

αxⁿ

S_n=

n

•

Broken stick (MacArthur 1957). There are no free parameters to be fitted in this model. With

S

_totthe total number of species and n_totthe total number of individuals:

S_tot−i

n_tot

1

n_i=

.

∑

S_tot_j=0S_tot− j

188

•

Log-normal. The fitting algorithm is from Krebs (1989). The logarithm (base 10) of the fitted

mean and variance are given. The octaves refer to power-of-2 abundance classes:

Octave Abundance

1

2

3

4

5

6

7

...

1

2-3

4-7

8-15

16-31

32-63

64-127

...

A significance value based on chi-squared is given for each of these models, but the power of the test

is not the same for the four models and the significance values should therefore not be compared. It

is important, as always, to remember that a high p value can not be taken to imply a good fit. A low

value does however imply a bad fit. Also note that the chi-squared tests in Past do not seem to

correspond with some other software, possibly because Past use counts rather than the log-

transformed values in the Whittaker plots.

References

Krebs, C.J. 1989. Ecological Methodology. Harper & Row, New York.

MacArthur, R.H. 1957. On the relative abundance of bird species. Proceedings of the National

Academy of Sciences, USA 43:293-295.

189

Species packing (Gaussian)

This module fits Gaussian response models to species abundances along a gradient, for one or more

species. The fitted parameters are optimum (average), tolerance (standard deviation) and maximum.

The module requires one first column of environmental measurements in samples (e.g.

temperature), and one or more additional columns of abundance data (taxa in columns).

The algorithm is the same as for the Gaussian function in the nonlinear regression module: Initial

estimation of optimum and tolerance based on the weighted average, followed by a nonlinear

optimization by the Levenberg-Marquardt method.

Note that the R²value (indicating goodness of fit) can become negative – this is not a bug but means

that the Gaussian fits worse than the sample mean.

190

Fit spiral

Fits a set of points in the plane to a logarithmic or Archimedes spiral. Useful for characterizing e.g.

mollusc shells, teeth, claws and horns. Requires two columns of coordinates (x and y). The points

must be given in sequence, either inwards or outwards. Left-handed and right-handed spirals are

both acceptable.

r = ae^b

The fitted logarithmic spiral in polar coordinates:

^θ. The scale a and the exponent b are given,

together with the estimated center point, marked with a red cross. The whorl expansion rate W

W = e^2πb

(factor increase in radius per whorl) is calculated from b as

.

The center position is estimated by nonlinear optimization and the spiral itself by linearization and

regression.

191

Changepoint modeling

This module suggests positions of abrupt change (changepoints) in a time series, with constant values

between the changepoints. The input data should be a single column with a series of numbers, or

multiple columns with multivariate data collected at the same points in time or stratigraphy. An

example application is the detection of breaks in multivariate geochemical data through a sediment

core. The module implements the method described by Gallagher et al. (2011).

The algorithm is Bayesian, “transdimensional” Markov chain Monte Carlo (MCMC). It produces not a

single set of model parameters, but a large number as samples (“simulations”) from the probability

distribution.

For multiple-column data sets, note that each column is weighted equally, as mean and standard

deviation are automatically normalized away prior to modeling.

Max chpoints: The maximal number of changepoints. This can often be left at the default, 10, unless

you want to allow a larger or enforce a smaller number of changepoints. After analysis, the actual

average number of changepoints (across simulations) is reported.

Simulations: The number of MCMC iterations, default 100,000. This includes the so-called burn-in,

which is the initial number of simulations before the algorithm hopefully converges and data start to

be collected. The number of burn-in iterations is fixed at 20,000. The “History” curve (see below)

should be inspected to see if the number of simulations should be increased. For noisy data, it may

be necessary to increase the number of simulations to a million or more, giving long computation

times.

Changepoints plot: Shows a histogram of the positions of changepoints across all simulations.

192

Model plot: Shows the average changepoint model as a red curve superimposed on the given data

points. If most simulations agree on the changepoint positions, this will be a step-like curve (i.e.

constant between changepoints). Variance (i.e. uncertainty) in changepoint positions will give a more

rounded appearance. For multivariate data, you can select the plotted variable in the drop-down

menu.

History: Shows the model log likelihood as a function of iteration number. Ideally, this curve should

start at some large negative value, and quickly increase to a relatively stable value, varying as

unstructured noise around a mean. The end of the burn-in is shown as a vertical line. If the log

likelihood does not seem to stabilize, the number of simulations may have to be increased.

Missing values are treated by linear interpolation before the analysis.

Reference

Gallagher, K., Bodin, T., Sambridge, M., Weiss, D., Kylander, M., Large, D. 2011. Inference of abrupt

changes in noisy geochemical records using transdimensional changepoint models. Earth and

Planetary Science Letters 311:182-194.

193

Diversity menu

Alpha diversity indices

These statistics apply to association data, where number of individuals are tabulated in rows (taxa)

and possibly several columns (samples). The available statistics are as follows, for each sample:

•

Number of taxa (S)

Total number of individuals (n)

Dominance = 1-Simpson index. Ranges from 0 (all taxa are equally present) to 1 (one taxon

dominates the community completely).





²

n

i

D =

where n_iis number of individuals of taxon i.



∑

n





i

If the “Unbiased” option is selected, an alternative form of D is computed:

(

)

ꢀ_ꢄꢀ_ꢄ− 1

ꢥ = �

(

)

ꢀ ꢀ − 1

ꢄ

•

Simpson index 1-D. Measures 'evenness' of the community from 0 to 1. Note the confusion in

the literature: Dominance and Simpson indices are often interchanged!

Shannon index (entropy). A diversity index taking into account the number of individuals as

well as number of taxa. Varies from 0 for communities with only a single taxon to high values

for communities with many taxa, each with few individuals.

194

n_i

n

n_i

n

H = −

ln

∑

i

If the “Unbiased” option is selected, H is computed with a bias correction

H_u= H + (S-1)/(2n).

If the “Unbiased” and “Use ACE for S” options are selected, the ACE species richness

estimator is used instead of S in the bias correction. This corresponds to the Bias-corrected

MLE (MLE_bc) estimator for Shannon’s index given by Chao & Shen (2003).

If the “log2” option is selected, the Shannon index is reported with the logarithm to base 2.

•

Buzas and Gibson's evenness: e^H/S

Brillouin’s index:

ln(n!)

−

ln(n_i!)

∑

n

i

HB =

S

•

Menhinick's richness index:

n

•

Margalef's richness index: (S-1) / ln(n)

Equitability. Shannon diversity H divided by the logarithm of number of taxa. This measures

the evenness with which individuals are divided among the taxa present. This index is

computed without bias correction for H (because S is not corrected correspondingly, this

could produce values larger than 1 which is outside the standard range for equitability).

•

Fisher's alpha - a diversity index, defined implicitly by the formula S=a*ln(1+n/a) where S is

number of taxa, n is number of individuals and a is the Fisher's alpha.

Berger-Parker dominance: simply the number of individuals in the dominant taxon relative to

n.

Chao1: An estimate of total species richness (Chao 1984). Version without bias correction:

ꢀ − 1 ꢎ²

1

ꢋ_{퐶ℎꢖꢧ1}= ꢋ + �

�

ꢀ

2ꢎ₂

where F₁is the number of singleton species and F₂the number of doubleton species. If F₂=0,

we use

(

)

ꢀ − 1 ꢎ ꢎ − 1

1

ꢋ_{퐶ℎꢖꢧ1}= ꢋ + �

�

ꢀ

2

When the “Unbiased” option is selected, the equation is

195

(

)

ꢀ − 1 ꢎ ꢎ − 1

1

ꢋ_{퐶ℎꢖꢧ1}= ꢋ + �

�

(

)

ꢀ

2 ꢎ₂+ 1

unless F₁= 1, in which case we revert to the version without bias correction.

•

iChao1: “Improved” Chao1 estimator (Chiu et al. 2014), taking into account also the numbers

F₃and F₄of species observed 3 and 4 times:

ꢀ − 3 ꢎ₃

ꢀ − 3 ꢎ₂ꢎ₃

ꢋ_{ꢄ퐶ℎꢖꢧ1}= ꢋ_{퐶ℎꢖꢧ1}

+

× max �ꢎ −

, 0�

1

(

)

4ꢀ ꢎ

2 ꢀ − 1

ꢎ

4

If F₄= 0, we use F₄=1 to avoid division by zero (Chiu et al. 2014).

ACE: Abundance-based Coverage Estimator (Chao & Lee 1992):

S

_rareis the number of species with 10 or less individuals. S_abundis the number of species with

more than 10 individuals (S = S_rare+ S_abund). n_rareis the number of individuals among the rare

species.

퐹

1

ꢉ_ꢌ퐶ꢨ= 1 − _푛

(sample cover estimate)

(coefficient of variation)

ꢩꢘꢩꢪ

10

ꢮ=1

∑

(

)

푘 푘−1 퐹

훾_ꢌ²_퐶ꢨ= max �^푆_퐶

ꢩꢘꢩꢪ

₎, 0�

−1

ꢫꢬꢭ ꢩꢘꢩꢪ ꢩꢘꢩꢪ

ꢮ

(

푛

푆

퐹

ꢩꢘꢩꢪ

1

ꢋ_ꢌ퐶ꢨ= ꢋ_{ꢖꢗꢯ푛ꢰ}

+

훾_ꢌ²_퐶ꢨ

퐶

ꢫꢬꢭ

•

Squares: Richness estimator (Alroy, 2018), designed to be more accurate than Chao-1 when

abundance distributions are even:

푆

ꢎ²

1

ꢀ²− ꢎ ꢋ

ꢋ_ꢱꢲ= ꢋ +

� ꢀ_ꢄ²

1

ꢄ=1

Some of these indices are explained in Harper (1999).

Confidence intervals

Approximate confidence intervals for all these indices can be computed with a bootstrap procedure.

The given number of random samples (default 9999) is produced, each with the same total number

of individuals as in the original sample. For each individual in the random sample, the taxon is chosen

with probabilities proportional to the original abundances. A 95 percent confidence interval is then

calculated. Note that the diversity in the replicates will often be less than, and never larger than, the

pooled diversity in the total data set – this bias can optionally be “fixed” by centering the confidence

interval on the original value.

Alternatively, analytical estimates of the 95% confidence interval are available for some of the

indices:

196

•

Dominance and Simpson index: Confidence interval estimated as [D - 1.96√(var D), D +

1.96√(var D)], with the variance var D computed as in the “Diversity t test” module (see

below).

Shannon index: Confidence interval estimated as [H - 1.96√(var H), H + 1.96√(var H)], with

the variance var H computed as in the “Diversity t test” module (see below).

These analytical confidence intervals for the Dominance/Simpson and the Shannon indices

correspond well to the centred bootstrap intervals, at least for large n.

•

Fisher alpha: Following Fisher et al. (1943), we use

2ꢀ + ꢒ

3

ꢒ ꢁ + ꢒ ²ln

− ꢒꢀ

(

)

ꢀ + ꢒ

var ꢒ =

2

(

)

ꢋꢀ + ꢋꢒ − ꢁꢒ

The 95% confidence interval is then [α - 1.96√(var α), α + 1.96√(var α)]. This analytical

confidence interval tends to be considerably wider than the bootstrapped CI. The reason for

the discrepancy is unknown.

•

Chao1: The computations are a bit complex. Past uses equations given in Colwell (2013). For

F₁>0 and F₂>0, and for the version without bias correction, we use

2

3

2

4

1 ꢀ − 1 ꢎ

ꢀ − 1

ꢎ

1 ꢀ − 1

ꢎ

(

)

var ꢋ_{퐶ℎꢖꢧ1}= ꢎ₂�

�

¹� + �

� � ¹� + �

� � ¹� �

2

ꢀ

ꢎ₂

ꢀ

ꢎ₂

4

ꢀ

ꢎ₂

For the bias-corrected version:

2

(

)

(

)

2

(

)

4

ꢀ − 1 ꢎ ꢎ − 1

ꢀ − 1

ꢎ 2ꢎ − 1

ꢀ − 1

ꢎ ꢎ₂ꢎ − 1

1

(

)

var ꢋ_{퐶ℎꢖꢧ1}= �

�

+ �

�

+ �

�

(

)

(

)

(

)

ꢀ

2 ꢎ₂+ 1

ꢀ

4 ꢎ₂+ 1

ꢀ

4 ꢎ₂+ 1

If F₂=0 but F₁> 1, we use for both versions:

2

ꢎ⁴

1

2

(

)

(

)

ꢀ − 1 ꢎ ꢎ − 1

ꢀ − 1

ꢎ 2ꢎ − 1

ꢀ − 1

1

(

)

var ꢋ_{퐶ℎꢖꢧ1}= �

�

+ �

�

− �

�

ꢀ

2

ꢀ

4

ꢀ

4ꢋ_{퐶ℎꢖꢧ1}

We then calculate T = S_Chao1– S and

(

)

var ꢋ_{퐶ℎꢖꢧ1}

�

퐾 = exp �1.96 ln �1 +

��

푇²

The 95% confidence interval is [S + T/K, S + TK].

Otherwise (F₁=0, or F₂=0 and F₁=1), we use

197

2

1

var ꢋ_{퐶ℎꢖꢧ1}≈ � ꢎ �푒^−ꢄ− 푒^−2ꢄ� − �� 푖푒^−ꢄꢎ �

(

)

ꢄ

ꢀ

ꢄ≥1

_ꢄ≥1ꢎ 푒^−ꢄ

ꢄ

∑

푃 =

ꢋ

95% confidence interval: [S/(1-P) – 1.96√(var(S_Chao1))/(1-P), S/(1-P) – 1.96√(var(S_Chao1))/(1-P)].

If the lower bound is smaller than S, it is set to S.

Bootstrapped comparison of diversity indices in two samples is provided in the Compare diversities

module.

References

Alroy, J. 2018. Limits to species richness in terrestrial communities. Ecology Letters 21:1781–1789.

Chao, A. 1984. Nonparametric estimation of the number of classes in a population. Scandinavian

Journal of Statistics 11:265-270.

Chao, A., Lee, S.-M. 1992. Estimating the number of classes via sample coverage. Journal of the

American Statistical Association 87:210–217.

Chao, A., Shen, T.-J. 2003. Nonparametric estimation of Shannon's diversity index when there are

unseen species in sample. Environmental and Ecological Statistics 10:429-443.

Chiu, C.-H., Wang, Y.-T., Walther, B.A., Chao, A. 2014. An improved nonparametric lower bound of

species richness via a modified Good–Turing frequency formula. Biometrics 70:671-682.

Colwell, R. K. 2013. EstimateS: Statistical estimation of species richness and shared species from

samples. Version 9. User's Guide and application published at: http://purl.oclc.org/estimates.

Fisher, R.A., Corbet, A.S., Williams, C.B. 1943. The relation between the number of species and the

number of individuals in a random sample of an animal population. The Journal of Animal Ecology

12:42-58.

Harper, D.A.T. (ed.). 1999. Numerical Palaeobiology. John Wiley & Sons.

198

Quadrat richness

Requires two or more columns, each containing presence/absence (1/0) of different taxa down the

rows (positive abundance is treated as presence).

Four non-parametric species richness estimators are included in PAST: Chao 2, first- and second-

order jackknife, and bootstrap. All of these require presence-absence data in two or more sampled

quadrats of equal size. Colwell & Coddington (1994) reviewed these estimators and found that the

Chao2 and the second-order jackknife performed best.

The output from Past is divided into two panels. First, the richness estimators and their analytical

standard deviations (only for Chao2 and Jackknife1) are computed from the given set of samples.

Then the estimators are computed from 1000 random resamplings of the samples with replacement

(bootstrapping), and their means and standard deviations are reported. In other words, the standard

deviations reported here are bootstrap estimates, and not based on the analytical equations.

Chao2

The Chao2 estimator (Chao 1987) is calculated as in EstimateS version 8.2.0 (Colwell 2009), with bias

correction:

Q

₁(Q₁−1)

m −1









ˆ

S_Chao2= S_obs

+

m

2

(Q₂+1)





where S_obsis the total observed number of species, m the number of samples, Q₁the number of

uniques (species that occur in precisely one sample) and Q₂the number of duplicates (species that

occur in precisely two samples).

If Q₁>0 and Q₂>0, variance is estimated as

199

2

Q

₁(Q₁−1)

Q₁(2Q₁−1)

Q₁²Q₂(Q₁−1)

m −1

























ˆ

.

var(S_Chao2

)

=

+

−

ˆ

2

4

m

2

(Q₂+1)

m

4

(Q₂+1)

4

(Q₂+1)













If Q₁>0 but Q₂=0:

2

Q

₁(Q₁−1)

Q₁(2Q₁−1)

Q₁⁴

m −1

























ˆ

var(S_Chao2

)

.

ˆ

m

2

m

4

m

4S_Chao2













If Q₁=0:

= S_obse^{−M S}

(

1− e^{−M S}

)

,

obs

ˆ

var

(

S_Chao2

ˆ

where M is the total number of occurrences of all species in all samples.

Jackknife 1

First-order jackknife (Burnham & Overton 1978, 1979; Heltshe & Forrester 1983):

m −1









ˆ

S _jack1= S_obs

+

Q

1

.

m





Q₁²

m

S









m −1







ˆ





(

var S

ˆ

)

=

j²f −

,



∑

jack1

j

m





j=0

where f_jis the number of samples containing j unique species.

Jackknife 2

Second-order jackknife (Smith & van Belle 1984):

2

Q₁(2m − 3)

Q₂(m − 2)

ˆ

S _{jack 2}= S_obs

+

−

.

m

(

m −1

)

No analytical estimate of variance is available.

Bootstrap

Bootstrap estimator (Smith & van Belle 1984):

S_obs

m

ˆ

,

S_boot= S_obs

+

(

1− p

)

∑

k

k=1

where p_kis the proportion of samples containing species k. No analytical estimate of variance is

available.

200

References

Burnham, K.P. & W.S. Overton. 1978. Estimation of the size of a closed population when capture

probabilities vary among animals. Biometrika 65:623-633.

Burnham, K.P. & W.S. Overton. 1979. Robust estimation of population size when capture

probabilities vary among animals. Ecology 60:927-936.

Chao, A. 1987. Estimating the population size for capture-recapture data with unequal catchability.

Biometrics 43, 783-791.

Colwell, R.K. & J.A. Coddington. 1994. Estimating terrestrial biodiversity through extrapolation.

Philosophical Transactions of the Royal Society (Series B) 345:101-118.

Heltshe, J. & N.E. Forrester. 1983. Estimating species richness using the jackknife procedure.

Biometrics 39:1-11.

Smith, E.P. & G. van Belle. 1984. Nonparametric estimation of species richness. Biometrics 40:119-

129.

201

Beta diversity

Two or more rows (samples) of presence-absence (0/1) data, with taxa in columns.

The beta diversity module in Past can be used for any number of samples (not limited to only two

samples). The eight measures available are described in Koleff et al. (2003):

Past

Koleff et al.

b_w

Equation

Ref.

Whittaker

(1960)

S

−1

Whittaker

α

Harrison

et al.

(1992)

S

−1

Harrison

b_-1

α

N −1

Cody

(1975)

g

(

H

)+ l(

H

)

Cody

b_c

b_I

2

Routledge

(1977)

 1

  1



log₁₀(

T

)

−

e log₁₀(e_i

)

−

α log

(

α_i

)

Routledge

∑

i

10



 



T



i

 

i



Wilson &

Shmida

(1984)

g

(

H

)+ l(

H

)

Wilson-Shmida b_t

2α

Mourelle

& Ezcurra

(1997)

g

(

H

)+ l(

H

)

Mourelle

b_me

2α

(

N −1

)

Harrison

et al.

(1992)

S

−1

Harrison 2

Williams

b_-2

b_-3

α_max

N −1

Williams

(1996)

α_max

1−

S

202

S: total number of species;

: average number of species; N: number of samples; g(H): total gain of

α

species along gradient (samples ordered along columns); l(H): total loss of species; e_i: number of

samples containing species i; T: total number of occurrences.

References

Harrison, S., S.J. Ross & J.H. Lawton. 1992. Beta diversity on geographic gradients in Britain. Journal

of Animal Ecology 61:151-158.

Koleff, P., K.J. Gaston & J.J. Lennon. 2003. Measuring beta diversity for presence-absence data.

Journal of Animal Ecology 72:367-382.

Routledge, R.D. 1977. On Whittaker’s components of diversity. Ecology 58:1120-1127.

Whittaker, R.H. 1960. Vegetation of the Siskiyou mountains, Oregon and California. Ecological

Monographs 30:279-338.

203

Taxonomic

Requires one or more columns (samples), each containing counts of individuals of different

taxa down the rows. In addition, one or more group columns with names of genera/families

etc. (see below).

Taxonomic diversity and taxonomic distinctness as defined by Clarke & Warwick (1998),

including confidence intervals computed from 10,000 random replicates taken from the

pooled data set (all samples). Note that the "global list" of Clarke & Warwick is not entered

directly, but is calculated internally by pooling (summing) the given samples.

These indices depend on taxonomic information also above the species level, which has to

be entered for each species as follows. Species names go in the name column (leftmost; in

the Row attributes), genus names in the first group column, family in second group column

etc., up to four group columns. Of course you can substitute for other taxonomic levels as

long as they are in ascending order. Species counts for the samples follow in the columns

thereafter.

Taxonomic diversity in one sample is given by (note other, equivalent forms exist):

w x x

∑∑

ij

i

j

i< j

∆ =

,

x x +

x

_i(x_i−1)

2

∑∑

∑

i

j

i< j

i

204

where the w_ijare weights such that w_ij= 0 if i and j are the same species, w_ij= 1 if they are

the same genus, etc. The x are the abundances.

Taxonomic distinctness:

w x x

∑∑

ij

i

j

i< j

∆^*=

.

x x

∑∑

i

j

i< j

For presence-absence data, taxonomic diversity and distinctness will be valid but equal to

each other.

Reference

Clarke, K.R. & Warwick, R.M. 1998. A taxonomic distinctness index and its statistical properties.

Journal of Applied Ecology 35:523-531.

205

Individual rarefaction

For comparing taxonomical diversity in samples of different sizes. Requires one or more columns of

counts of individuals of different taxa (each column must have the same number of values). When

comparing samples: Samples should be taxonomically similar, obtained using standardised sampling

and taken from similar 'habitat'.

Given one or more columns of abundance data for a number of taxa, this module estimates how

many taxa you would expect to find in a sample with a smaller total number of individuals. With this

method, you can compare the number of taxa in samples of different size. Using rarefaction analysis

on a large sample, you can read out the number of expected taxa for any smaller sample size

(including that of the smallest sample). The algorithm is from Krebs (1989), using a log Gamma

function for computing combinatorial terms. An example application in paleontology can be found in

Adrain et al. (2000).

Let N be the total number of individuals in the sample, s the total number of species, and N_ithe

number of individuals of species number i. The expected number of species E(S_n) in a sample of size n

and the variance V(S_n) are then given by





N − N 







ⁱ





s

n





E

(S_n

)

=

1−

∑



N

 

ⁱ⁼¹

 

n









 

206









N − N





N − N 



_i



_i











s

n











V

(S_n

)

=

1

−

∑

N

 

i

=

¹









 

n









 

 N − N − N

N − N















i



^j





ⁱ

^j









j

−1

s

n











+ 2

−

∑∑_

N

 

  

j

=2 i=

¹

 

  

n









 

  

Standard errors (square roots of resampling variances) are given by the program. In the graphical

plot, these standard errors are converted to 95 percent confidence intervals.

Unconditional variance

The classical rarefaction variance given above is called conditional variance. It is conditional on the

reference sample, and will reduce to zero for S_n= s. In contrast, Colwell et al. (2012) described an

unconditional rarefaction variance estimate that will not reduce to zero at the end of the rarefaction

curve. This method is also available in Past.

There are two models for individual rarefaction described by Colwell et al. (2012), the multinomial

model (classical rarefaction) and the Poisson model (Coleman rarefaction). The two methods give

quite similar results. The “industry standard” rarefaction software, EstimateS, somewhat

incongruously computes E(S_n) according to the multinomial equation (eq. (4) in Colwell et al.,

equivalent to the equation given above), while V(S_n) uses the Poisson formulation (eq. 7 in Colwell et

al.), according to the EstimateS manual. This approach is followed in Past for compatibility with

EstimateS. The computation also requires an estimate for total (sampled and unsampled) species

richness. The Chao1 estimator is used for this (cf. Colwell et al. 2012).

Rarefaction of Simpson and Shannon indices

In addition to rarefaction of species richness, Past also includes rarefaction of the Simpson D and

Shannon H diversity indices, following Chao et al. (2014). In the forms 1/D and e^H, these are special

cases of so-called Hill numbers, which is a family of diversity indices. Past reports the rarefaction

curves in these forms, for consistency with Chao et al. (2014), but for convenience the Shannon index

can also be reported as the conventional H. Past does not yet compute confidence intervals for these

rarefaction curves (which would require bootstrapping).

Let Xi be the number of individuals of the ith species that are observed in the sample, i = 1, 2, …ꢀ, S.

Let f_kbe the number of species represented by exactly k individuals in the sample, k = 0, 1, …ꢀ, n. For

smaller sample sizes (m<n), and k>0, the estimator for f_kis (eq. 7 in Chao et al. 2014)

푋_ꢄ

ꢀ − 푋_ꢄ

� � �

ꢳ

�

푚 − ꢳ

ꢀ

̂

푓 = �

푘

� �

ꢴ ≥푘

ꢟ

푚

The estimator for the Hill number with q=1, which is equivalent to the exponential e^Hof the Shannon

index, is (eq. 10a in Chao et al. 2014)

207

ꢊ

ꢳ

1

̂

�

(

)

(

)

ꢥ 푚 = exp �� −

ln � 푓 푚 �

푘

푚

푘=1

The estimator for the Hill number with q=2, which is equivalent to the inverse 1/D of the Simpson

index, is (eq. 11b in Chao et al. 2014)

1

2

�

(

)

ꢥ 푚 =

2

ꢳ

ꢊ

푘=1

̂

∑

(

)

� � 푓 푚

푘

푚

Missing values: Treated as zero.

References

Adrain, J.M., S.R. Westrop & D.E. Chatterton. 2000. Silurian trilobite alpha diversity and the end-

Ordovician mass extinction. Paleobiology 26:625-646.

Chao, A., Gotelli, N.J., Hsieh, T.C., Sander, E.L., Ma, K.H., Colwell, R.K. & Ellison, A.M. 2014.

Rarefaction and extrapolation with Hill numbers: a framework for sampling and estimation in species

diversity studies. Ecological Monographs 84:45-67.

Colwell, R.K., Chao, A., Gotelli, N.J., Lin, S.-L., Mao, C.X., Chazdon, R.L. & Longino, J.T. 2012. Models

and estimators linking individual-based and sample-based rarefaction, extrapolation and comparison

of assemblages. Journal of Plant Ecology 5:3-21.

Krebs, C.J. 1989. Ecological Methodology. Harper & Row, New York.

208

Shareholder Quorum Subsampling (SQS)

Shareholder Quorum Subsampling was introduced by John Alroy (2010). Chao and Jost (2012)

provided analytical solutions for an almost equivalent method, which they called “Coverage-based

rarefaction”. Past follows the algorithm of Alroy (2010).

Like rarefaction, SQS can be used to standardize species counts across samples of different size, but

standardizing on a fixed “coverage” rather than a fixed sample size. Coverage is defined as the

proportion of individuals in the population that are represented by the species recovered in the

sample. If all species are recovered in the sample, coverage is 1. If only one species is recovered, but

50% of the individuals in the population belong to this species, coverage is 0.5. SQS gives a more fair

sampling of communities with different evenness than classical rarefaction, which suffers from a

“compression effect” where differences in richness are artificially dampened. SQS is therefore

gradually replacing classical rarefaction in the literature.

The module expects one or more columns of count data. The following parameters can be selected:

Quorum: The desired subsampled coverage level (0-1). This value should be larger than 0.4. A value

of e.g. 0.9 might be better, but for small samples this coverage may not be achieved, giving an error

message. The quorum value is not critical: Larger values will give higher SQS species richness, but this

is of little consequence for the comparison of several samples.

Trials: The number of subsampling runs. Higher values give more exact results, but takes longer.

Ignore singletons and Ignore dominant: Disregard species with abundance 1; disregard the single

most common species. The possible advantage of this is slightly unclear.

The module outputs the original sample size (N), the number of observed species (S_obs), and

Good’s u which is a simple estimator of sample coverage: u = 1 - f₁/N, where f₁is the number of

singletons (Chao & Jost 2012 give an improved estimator), in addition to the SQS richness.

References

Alroy, J. 2010. Geographical, environmental and intrinsic biotic controls on Phanerozoic marine

diversification. Palaeontology 53:1211–1235.

Chao, A., Jost, L. 2012. Coverage-based rarefaction and extrapolation: standardizing samples by

completeness rather than size. Ecology 93:2533-2547.

209

Sample rarefaction (Mao’s tau)

Sample rarefaction requires a matrix of presence-absence data (abundances treated as presences),

with taxa in rows and samples in columns. Sample-based rarefaction (also known as the species

accumulation curve) is applicable when a number of samples are available, from which species

richness is to be estimated as a function of number of samples. PAST implements the analytical

solution known as "Mao’s tau", with standard deviation. In the graphical plot, the standard errors are

converted to 95 percent confidence intervals. See Colwell et al. (2004) for details.

With H samples and S_obsthe total number of observed species, let s_jbe the number of species found

in j samples, such that s₁is the number of species found in exactly one sample, etc. The total number

of species expected in h≤H samples is then

H

~

.

τ

(

h

)

= S_obs

−

α s

∑

jh

j

j=1

The combinatorial coefficients α are

(H − h)

!

(H − j)

!





for j + h ≤ H

.

α _jh

=

(H − h − j)!H!







0

for j + h > H

These coefficients are computed via a log Gamma function. The variance estimator is

2

H

~

τ

(

h

)

σ =

1−α

²s _j−

2

~

(

)

,

∑

~

jh

S

j=1

~

where

is an estimator for the unknown total species richness. Following Colwell et al. (2004), a

S

Chao2-type estimator is used. For s₂>0,

(H −1)s₁²

~

S = S_obs

+

.

2Hs₂

For s₂=0,

(H −1)s₁(s₁−1)

~

.

S = S_obs

+

2H(s₂+1)

For modeling and extrapolating the curve using the Michaelis-Menten equation, use the Copy Data

button, paste to a new Past spreadsheet, and use the nonlinear fitting module in the Model menu.

Past also includes extrapolation of the sample rarefaction curve, up to 4H samples, using equation 18

in Colwell et al. (2012). Confidence intervals for the extrapolation is not yet implemented.

210

References

Colwell, R.K., C.X. Mao & J. Chang. 2004. Interpolating, extrapolating, and comparing incidence-based

species accumulation curves. Ecology 85:2717-2727.

Colwell, R.K., Chao, A., Gotelli, N.J., Lin, S.-Y., Mao, C.X., Chazdon, R.L., Longino, J.T. 2012. Models and

estimators linking individual-based and sample-based rarefaction, extrapolation and comparison of

assemblages. Journal of Plant Ecology 5:3–21

211

SHE analysis

SHE analysis (Hayek & Buzas 1997, Buzas & Hayek 1998) requires a matrix of integer abundance data

(counts), with taxa in rows and samples in columns. The program calculates log species abundance

(ln S), Shannon index (H) and log evenness (ln E = H – ln S) for the first sample. Then the second

sample is added to the first, and the process continues. The resulting cumulative SHE profiles can be

interpreted ecologically. If the samples are taken not from one homogenous population but across a

gradient or up a stratigraphic section, breaks in the curve may be used to infer discontinuities (e.g.

biozone boundaries).

References

Buzas, M.A. & L.-A. C. Hayek. 1998. SHE analysis for biofacies identification. The Journal of

Foraminiferal Research 28:233-239.

Hayek, L.-A. C. & M.A. Buzas. 1997. Surveying natural populations. Columbia University Press.

212

Diversity permutation test

Expects two columns of abundance data with taxa down the rows.This module computes a number

of diversity indices for two samples, and then compares the diversities using random permutations.

999 random matrices with two columns (samples) are generated, each with the same row and

column totals as in the original data matrix.

213

Diversity t test

Comparison of the Shannon and Simpson diversities in two samples. Then Shannon t test is described

by e.g. Hutcheson (1970), Poole (1974), Magurran (1988). This is an alternative to the randomization

test available in the Diversity permutation test module. Requires two columns of abundance data

with taxa down the rows.

The Shannon index here include a bias correction and may diverge slightly from the uncorrected

estimates calculated elsewhere in PAST, at least for small samples. With p_ithe proportion (0-1) of

taxon i, S the number of taxa and N the number of individuals, the estimator of the index is

S

S −1

H'= −

p ln p −

(note that the second term is incorrect in Magurran 1988).

∑

i

2N

i=1

The variance of the estimator is

2

p

_i(ln p_i

)

−

[

(p_iln p_i

)

]

S −1

∑

Var H'=

+

.

2N ²

N

The t test statistic is given by

′

H₁− H₂

t =

.

′

Var H₁+ Var H₂

The degrees of freedom for the t test is

2

′

(Var H₁+ Var H₂

)

df =

.

2

′

(Var H₁)

(Var H₂

)

+

N₁

N₂

The Simpson index (dominance) has estimated variance (Brower et al. 1998):

2

4N(N −1)(N − 2)

p³+ 2N(N −1)

p²− 2N(N −1)(2N − 3)(

p²

i

)

∑

i

Var D =

.

2

N ²

(

N −1

)

References

Brower, J.E., Zar, J.H., von Ende, C.N. 1998. Field and Laboratory Methods for General Ecology.

McGraw-Hill, Boston.

Hutcheson, K. 1970. A test for comparing diversities based on the Shannon formula. Journal of

Theoretical Biology 29:151-154.

Magurran, A. 1988. Ecological Diversity and Its Measurement. Princeton University Press.

Poole, R.W. 1974. An introduction to quantitative ecology. McGraw-Hill, New York.

214

Diversity profiles

This module requires one or more columns of abundance data with taxa down the rows. The main

purpose is to compare diversities in several samples.

The validity of comparing diversities across samples can be criticized because of arbitrary choice of

diversity index. One sample may for example contain a larger number of taxa, while the other has a

larger Shannon index. A number of diversity indices may be compared to make sure that the diversity

ordering is robust. A formal way of doing this is to define a family of diversity indices, dependent

upon a single continuous parameter (Tothmeresz 1995).

PAST uses the exponential of the so-called Renyi index, which depends upon a parameter α. For α=0,

this function gives the total species number. α=1 (in the limit) gives an index proportional to the

Shannon index, while α=2 gives an index which behaves like the Simpson index.

S

 1





exp(H_α

)

= exp

ln

p^α

i



∑

1−α

i=1





The program can plot several such diversity profiles together. If the profiles cross, the diversities are

non-comparable. The bootstrapping option (giving a 95% confidence interval) is based on 2000

replicates.

Reference

Tothmeresz, B. 1995. Comparison of different methods for diversity ordering. Journal of

Vegetation Science 6:283-290.

215

Time series menu

Simple periodogram

Since palaeontological data are often unevenly sampled, Fourier-based methods can be difficult to

use. PAST therefore includes the Lomb periodogram for unevenly sampled data (Press et al. 1992),

with time values given in the first column and dependent values in the second column. If only one

column is selected, even spacing of one unit between data points is assumed. The Lomb periodogram

should then give similar results as the FFT. The data are automatically detrended prior to analysis.

The frequency axis is in units of 1/(x unit). If for example, your x values are given in millions of years,

a frequency of 0.1 corresponds to a period of 10 million years. The power axis is in units proportional

to the square of the amplitudes of the sinusoids present in the data. Also note that the frequency

axis extends to very high values. If your data are evenly sampled, the upper half of the spectrum is a

mirror image of the lower half, and is of little use. If some of your regions are closely sampled, the

algorithm may be able to find useful information even above the half-point (Nyquist frequency).

The highest peak in the spectrum is presented with its frequency and power value, together with a

probability that the peak could occur from random data. The 0.01 and 0.05 significance levels ('white

noise lines') are shown as red dashed lines.

The example above shows a spectral analysis of a foram oxygen isotope record from 1 Ma to Recent,

with an even spacing of 0.003 Ma (3 ka). There are periodicities at frequencies of about 9 (split peak),

25 and 43 Ma^-1, corresponding to periods of 111 ka, 40 ka and 23 ka – clearly orbital forcing.

Reference

Press, W.H., S.A. Teukolsky, W.T. Vetterling & B.P. Flannery. 1992. Numerical Recipes in C. Cambridge

University Press.

216

REDFIT spectral analysis

This module is an implementation of the REDFIT procedure of Schulz and Mudelsee (2002). It is a

more advanced version of the simple Lomb periodogram described above. REDFIT includes an option

for “Welch overlapped segment averaging”, which implies splitting the time series into a number of

segments, overlapping by 50%, and averaging their spectra. This reduces noise but also reduces

spectral resolution. In addition, the time series is fitted to an AR(1) red noise model which is usually a

more appropriate null hypothesis than the white noise model described above. The given “false-

alarm lines” are based on both parametric approximations (chi2) and Monte Carlo (using 1000

random realizations of an AR(1) process).

The input must be in the form of two columns with time and data values, or one column of equally-

spaced data values. The data are automatically detrended. The fitting to AR(1) implies that the data

must have the correct time direction (in contrast with the simple spectrogram above where the time

direction is arbitrary). The time values are expected to be ages before present. If not, it will be

necessary to give them negative signs.

The frequency oversampling value controls the number of points along the frequency axis (but

having many points does not increase frequency resolution!). Increasing the number of segments will

reduce noise, but also decrease the resolution. The window function influences the trade-off

between spectral resolution and attenuation of side lobes.

The (average) tau value is the characteristic time scale (the parameter of the AR model). The

bandwidth is the spectral resolution given as the width between the -6dB points.

The fit to an AR(1) model can be assessed using the runs value and its 5% acceptance interval. This

test is only available with Monte Carlo on, oversampling=1, segments=1, window=rectangular.

217

In addition to a fixed set of false-alarm levels (80%, 90%, 95% and 99%), the program also reports a

“critical” false-alarm level (False-al) that depends on the segment length (Thomson 1990).

Important: Because of long computation time, the Monte Carlo simulation is not run by default, and

the Monte Carlo false-alarm levels are therefore not available. When the Monte Carlo option is

enabled, the given spectrum may change slightly because the Monte Carlo results are then used to

compute a “bias-corrected” version (see Schulz and Mudelsee 2002).

Missing values supported.

References

Schulz, M. & M. Mudelsee. 2002. REDFIT: estimating red-noise spectra directly from unevenly spaced

paleoclimatic time series. Computers & Geosciences 28:421-426.

Thomson, D.J. 1990. Time series analysis of Holocene climate data. Philosophical Transactions of the

Royal Society of London, Series A 330:601-616.

218

Multitaper spectral analysis

In traditional spectral estimation, the data are often “windowed” (multiplied with a bell-shaped

function) to reduce spectral leakage. In the multitaper method, several orthogonal window functions

are applied, and the results combined. The resulting spectrum has low leakage, low variance, and

retains information from the beginning and end of the time series. Also, statistical testing can take

advantage of the multiple spectral estimates. A possible disadvantage is reduced spectral resolution.

Requires evenly spaced data, given in one column. The data are not detrended.

The implementation is based on Mann & Lees (1996) and Fortran code by Michael Mann. It includes

a red-noise model based on a “reshaped” spectrum, i.e. after removing and interpolating peaks.

The number of tapers can be set to 3 or 5 for different trade-offs between variance and resolution.

Spectrum type: The two algorithms give very similar results, but the “adaptive” option is

recommended by Mann & Lees (1996).

Sample interval: This value only affects the scaling of the frequency axis.

Reshape threshold: This value affects how strong a spectral peak must be to count as a harmonic

component, to be removed by the reshaping procedure.

Note: Past reproduces examples in Mann & Lees (1996), but some other implementations seem to

give higher levels for the confidence (significance) lines. The reason for this is unknown.

Reference

Mann, M.E. & Lees. J. 1996. Robust estimation of background noise and signal detection in climatic

time series. Climatic Change 33:409-445.

219

Walsh transform

The Walsh transform is a type of spectral analysis (for finding periodicities) of binary or ordinal data.

It assumes even spacing of data points, and expects one column of binary (0/1) or ordinal (integer)

data.

The normal methods for spectral analysis are perhaps not optimal for binary data, because they

decompose the time series into sinusoids rather than "square waves". The Walsh transform may

then be a better choice, using basis functions that flip between -1 and +1. These basis functions have

varying "frequencies" (number of transitions divided by two), known as sequencies. In PAST, each

pair of even ("cal") and odd ("sal") basis functions is combined into a power value using cal²+sal²,

producing a "power spectrum" that is comparable to the Lomb periodogram.

In the example above, compare the Walsh periodogram (top) to the Lomb periodogram (bottom).

The data set has 0.125 periods per sample. Both analyses show harmonics.

The Walsh transform is slightly exotic compared with the Fourier transform, and the results must be

interpretated cautiously. For example, the effects of the duty cycle (percentage of ones versus zeros)

are somewhat difficult to understand.

In PAST, the data values are pre-processed by multiplying with two and subtracting one, bringing 0/1

binary values into the -1/+1 range optimal for the Walsh transform. The data are zero-padded to the

next power of 2 if necessary, as required by the method.

220

Evolutive Fourier transform

Spectral analysis using the Fourier transform (FFT), but dividing the signal into a sequence of

overlapping windows, which are analysed individually. This allows development of the spectrum in

time, in contrast with the global analysis provided by the other spectral analysis modules. Sample

position is shown on the x axis, frequency (in periods per sample) on the y axis, and power on a

logarithmic scale as colour or grey scale.

The Short-time Fourier Transform (STFT) can be compared with wavelet analysis, but with a linear

frequency scale and with constant time resolution independent of frequency.

The window size controls the trade-off between resolution in time and frequency; small windows

give good time resolution but poor frequency resolution. Windows are zero-padded by a factor eight

to give a smoother appearance of the diagram along the frequency axis. The window functions

(Rectangle, Welch, Hanning, Hamming, Blackman-Harris, multitaper with 3, 4 or 5 tapers) give

different trade-offs between frequency resolution and sideband rejection.

Missing values are treated using linear interpolation before analysis.

221

Wavelet transform

Inspection of time series at different scales. Requires one column of ordinal or continuous data with

even spacing of points.

The continuous wavelet transform (CWT) is an analysis method where a data set can be inspected at

small, intermediate and large scales simultaneously. It can be useful for detecting periodicities at

different wavelengths, self-similarity and other features. The vertical axis in the plot is a logarithmic

size scale (base 2), with the signal observed at a scale of only two consecutive data points at the top,

and at a scale of one fourth of the whole sequence at the bottom. One unit on this axis corresponds

to a doubling of the size scale. The top of the figure thus represents a detailed, fine-grained view,

while the bottom represents a smoothed overview of longer trends. Signal power (or more correctly

squared correlation strength with the scaled mother wavelet) is shown with a grayscale or in colour.

The shape of the mother wavelet can be set to Morlet (wavenumber 6), Paul (4^thorder) or DOG

(Derivative Of Gaussian, 2^ndor 6^thderivative). The Morlet wavelet usually performs best.

The example above is based on a foram oxygen isotope record from 1 Ma to Recent, with an even

spacing of 0.003 Ma (3 ka). A band can be seen at a scale of about 2⁵=32 samples, or about 100 ka. A

weaker band around 2^3.7=13 samples corresponds to a scale of about 40 ka. These are orbital

periodicities. In contrast with the “bulk” spectral analysis, the scalogram makes visible changes in

strength and frequency over time.

The so-called “cone of influence” can be plotted to show the region where boundary effects are

present.

The ‘Sample interval’ value can be set to a value other than 1. This will only influence the scaling of

the labels on the x and y axes.

222

The algorithm is based on fast convolution of the signal with the wavelet at different scales, using the

FFT.

Significance test: The significance level corresponding to p=0.05 can be plotted as a contour (chi-

squared test according to Torrence & Compo 1998). The “Lag” value, as given by the user, specifies

the null hypothesis. Lag=0 specifies a white-noise model. Values 0<Lag<1 specifies a red-noise model

with the given MA(1) autocorrelation coefficient. It can be estimated using the ARMA module in the

Time menu (specify zero AR terms and one MA term and note the MA value in the Coefficients tab).

If the “Power” option is deselected, the program will show only the real part of the scalogram (not

squared). This shows the signal in the time domain, filtered at different scales.

In the ‘View numbers’ window, each row shows one scale, with sample number (position) along the

columns.

The ‘Filter’ tab shows the time series at one scale value, as power values if the ‘Power’ option is

selected in the main tab, or real parts if not. This, in effect, works as a bandpass filter.

The wavelet transform was used by Prokoph et al. (2000) for illustrating cycles in diversity curves for

planktic foraminifera. The code in Past is based on Torrence & Compo (1998).

Missing values are treated using linear interpolation before analysis.

References

Prokoph, A., A.D. Fowler & R.T. Patterson. 2000. Evidence for periodicity and nonlinearity in a high-

resolution fossil record of long-term evolution. Geology 28:867-870.

Torrence, C. & G.P. Compo. 1998. A practical guide to wavelet analysis. Bulletin of the American

Meteorological Society 79:61-78.

223

Wavelets for unequal spacing

Inspection of time series at different scales. Requires two input columns, containing time and data

values.

This module is similar to the Wavelet Transform module, but accepts unequally spaced data. It does

not (yet) provide a significance test, and because it does not use the FFT it is considerably slower.

Also, the frequency axis is linear, not logarithmic. The algorithm is based on the Weighted Wavelet z-

transform (WWZ) of Foster (1996) and the Fortran implementation of Templeton (2004). Please note

that this module does not do magic – in intervals with little data the wavelet analysis will not be

informative, especially at high frequencies.

The c parameter of Foster (1996) is fixed at c = 1/72, slightly higher than the recommended c = 1/8π.

This value is chosen because it corresponds to a wavenumber of 6 as used by the Wavelet Transform

module.

References

Foster, G. 1996. Wavelets for period analysis of unevenly sampled time series. The Astronomical

Journal 112:1709-1729.

Templeton, M. 2004. Time-series analysis of variable star data. The Journal of the American

Association of Variable Star Observers 32:41-54.

224

Point events spectrum

This module, using the “circular spectral analysis” method (e.g. Lutz 1985) is used to search for

periodicity in point event series such as earthquakes, volcanic eruptions, and mass extinctions (e.g.

Rampino & Caldeira 2015). A single column of event times (e.g. dates of eruptions in millions of

years) is required. The event times do not need to be in sequential order.

The method works by wrapping the time line around a circle with a circumference corresponding to a

trial period P. If points are P-periodic, they will cluster at a certain angle corresponding to the phase.

The ages of events t_iare converted to angles a_i:

2휋푡_ꢄ

푎_ꢄ=

mod 2휋

푃

As in directional statistics, the mean sines and cosines are computed and converted to a mean vector

magnitude (Rayleigh statistic) R and a phase t₀:

1

ꢋ =

� sin 푎_ꢄ

ꢁ

1

ꢉ =

� cos 푎_ꢄ

ꢁ

2

�

푅 = ꢋ + ꢉ

푃

ꢋ

푡₀=

tan⁻¹

2휋

ꢉ

(taken to the correct quadrant)

R and t₀are computed for P ranging from the average waiting time up to 1/3 of the total duration of

the series, giving a full spectrum.

225

A 95% significance line for R (in red) is computed by a Monte Carlo procedure with 1,000 replicates.

Random event times are computed by a gamma distribution for waiting times. The shape parameter

should be set to k=1 (i.e. exponential distribution) for a null model with no interactions between

events (Poisson process). If closely spaced points are expected to be rare, you can set k=2 or k=3.

Wrapping correction: Lutz (1985) described a correction for non-integer number of wraps causing

some points to be over-represented. This correction, optional in Past, gives a jagged appearance of

the spectral curve and seems to work best for relatively large numbers of points (N>20).

Harmonics: This method is as plagued by harmonics as traditional Fourier analysis. A spectral peak for

a period P will be accompanied by strong peaks also on harmonics, i.e. at P/2, P/3 etc. It is important

to take this into account when interpreting the spectrum.

References

Lutz, T.M. 1985. The magnetic reversal record is not periodic. Nature 317:404-407.

Rampino, M.R. & K. Caldeira. 2015. Periodic impact cratering and extinction levels over the last 260

million years. Monthly Notices of the Royal Astronomical Society 454:3480-3484.

226

Autocorrelation

Autocorrelation (Davis 1986) is carried out on separate column(s) of evenly sampled

temporal/stratigraphic data. Lag times τ up to n/2, where n is the number of values in the vector, are

shown along the x axis (positive lag times only - the autocorrelation function is symmetrical around

zero). A predominantly zero autocorrelation signifies random data - periodicities turn up as peaks.

The "95 percent confidence interval" option will draw lines at

1

±1.76

n −τ + 3

after Davis (1986). This is the confidence interval for random, independent points (white noise).

There are two issues: White noise is an unrealistic null model, and the confidence interval is only

strictly valid at each individual lag (multiple testing problem).

Missing data supported.

Reference

Davis, J.C. 1986. Statistics and Data Analysis in Geology. John Wiley & Sons.

227

Autoassociation

Autoassociation is analogous to autocorrelation, but for a sequence of binary or nominal data coded

as integer numbers.

For each lag, the autoassociation value is simply the ratio of matching positions to total number of

positions compared. The expected autoassociation value (0.335 in the example above) for a random

sequence is (Davis 1986)

m

X ²− n

∑

k

k=1

P =

n²− n

where n is the total number of positions, m is the number of distinct states (3 in the example above)

and X_kis the number of observations in state k.

For non-zero lags, a P value is computed from the overlapping positions only, and the expected

number of matches is then given by E=nP. This is compared with the observed number of matches O

to produce a χ²with 1 degree of freedom:

2

′

(O − E −1 2)

χ ²=

+

′

E

with O’=n-O and E’=n(1-P) the observed and expected number of mismatches. Note the Yates’

correction. The resulting p values (two-tailed) can be shown as a function of lag.

The multiple testing issue arises for the set of p values.

228

The test above is not strictly valid for “transition” sequences where repetitions are not allowed (the

sequence in the example above is of this type). In this case, select the “No repetitions” option. The p

values will then be computed by an exact test, where all possible permutations without repeats are

computed and the autoassociation compared with the original values (on-tailed). This test will take a

long time to run for n>30, and the option is not available for n>40.

Missing data supported.

Reference

Davis, J.C. 1986. Statistics and Data Analysis in Geology. John Wiley & Sons.

229

Cross-correlation

Cross-correlation (Davis 1986)is carried out on two column(s) of evenly sampled

temporal/stratigraphic data. The x axis shows the displacement of the second column with respect to

the first, the y axis the correlation between the two time series for a given displacement. The "p

values" option will draw the significance of the correlation, after Davis (1986).

For two time series x and y, the cross-correlation value at lag time m is

(x_i− x)(y_i−m− y)

∑

.

r_m=

2

(x_i− x)

(y_i−m− y)

∑

The summations and the mean values are only taken over the parts where the sequences overlap for

a given lag time.

The equation shows that for positive lags, x is compared with a y that has been delayed by m

samples. A high correlation value at positive lags thus means that features in y are leading, while x

lags behind. For negative lags, features in x are leading. A reminder of this is given by the program.

The p value for a given m is given by a t test with n-2 degrees of freedom, with n the number of

samples that overlap:

n − 2

t = r_m

.

1− r_m²

It is important to note that this test concerns one particular m. Plotting p as a function of all m raises

the issue of multiple testing – p values smaller than 0.05 are expected for 5% of lag times even for

completely random (uncorrelated) data sets.

In the example above, the “earthquakes” data seem to lag behind the “injection” data with a delay of

0-2 samples (months in this case), where the correlation values are highest. The p values (red curve)

indicates significance at these lags. Curiously, there also seems to be significance for negative

correlation at large positive and negative lags.

Missing data supported.

Reference

Davis, J.C. 1986. Statistics and Data Analysis in Geology. John Wiley & Sons.

230

Mantel correlogram (and periodogram)

This module expects several rows of multivariate data, one row for each sample. Samples are

assumed to be evenly spaced in time.

The Mantel correlogram (e.g. Legendre & Legendre 1998) is a multivariate extension to

autocorrelation, based on any similarity or distance measure. The Mantel correlogram in PAST shows

the average similarity between the time series and a time lagged copy, for different lags.

The Mantel periodogram is a power spectrum of the multivariate time series, computed from the

Mantel correlogram (Hammer 2007).

The Mantel scalogram is an experimental plotting of similarities between all pairs of points along the

time series. The apex of the triangle is the similarity between the first and last point. The base of the

triangle shows similarities between pairs of consecutive points.

231

References

Hammer, Ø. 2007. Spectral analysis of a Plio-Pleistocene multispecies time series using the Mantel

periodogram. Palaeogeography, Palaeoclimatology, Palaeoecology 243:373-377.

Legendre, P. & L. Legendre. 1998. Numerical Ecology, 2nd English ed. Elsevier, 853 pp.

232

Runs test

The runs test is a non-parametric test for randomness in a sequence of values such as a time series.

Non-randomness may include such effects as autocorrelation, trend and periodicity. The module

requires one column of data, which are internally converted to 0 (x≤0) or 1 (x>0).

The test is based on a dichotomy between two values (x≤0 or x>0). It counts the number of runs

(groups of consecutive equal values) and compares this to a theoretical value. The runs test can

therefore be used directly for sequences of binary data. There are also options for “runs about the

mean” (the mean value subtracted from the data prior to testing), or “runs up and down” (the

differences from one value to the next taken before testing).

With n the total number of data points, n₁the number of points ≤0 and n₂the number of points >0,

the expected number of runs in a random sequence, and the variance, are

n + 2n₁n₂

.

E

(

R

)

=

n

2n₁n₂(2n₁n₂− n)

.

Var(

R

)

=

n²(n −1)

With the observed number of runs R, a z statistic can be written as

R − E(

R

)

z =

.

Var(R)

The resulting two-tailed p value is not accurate for n<20. A Monte Carlo procedure is therefore also

included, based on 10,000 random replicates using the observed n, n₁and n₂.

233

Mann-Kendall trend test

A non-parametric test for trend. Requires a single column of data. Missing values are deleted, and n

adjusted accordingly. The procedure follows Gilbert (1987).

Data x₁, ... x_nare assumed to be ordered in sequence of collection time, or in spatial sequence. Define

the indicator function

1, if 푥 > 0

sgn 푥 = �0, if 푥 = 0

1, if 푥 > 0

The S statistic is calculated by summing over all pairs of values:

푛−1

푛

ꢋ = � � 푠푔ꢀ�푥_푗− 푥_ꢄ�

ꢄ=1 푗=ꢄ+1

S will be negative for a negative trend, zero for no trend, and positive for an increasing trend.

For n≤10, the p value is taken from a table of exact values (Gilbert 1987). For n>10, a normal

approximation is used, as follows.

Determine the total number of groups of ties g and the number of tied values t_jwithin each group, in

the sorted sequence. Then estimate the standard deviation of S by

ꢅ

1

(

)(

)

�

ꢋꢥ =

�ꢀ ꢀ − 1 2ꢀ + 5 − � 푡 �푡 − 1��2푡 + 5��

푗

18

푗=1

The Z statistic is then

| |

ꢋ − 1

푍 =

sgn ꢋ

ꢋꢥ

which is used to calculate p from the cumulative normal distribution as usual. The subtraction of 1 is

a continuity correction.

Reference

Gilbert, R.O. 1987. Statistical methods for environmental pollution monitoring. Van Nostrand

Reinhold, New York.

234

Point events

Expects one column containing times of events (e.g. earthquakes or clade divergences) or positions

along a line (e.g. a transect). The times do not have to be in increasing order.

Density trend (Laplace test)

The “Laplace” test for a trend in density (intensity) is described by Cox & Lewis (1978). It is based on

the test statistic

L

t −

2

U =

1

L

12n

t

where is the mean event time, n the number of events and L the length of the interval. L is

estimated as the time from the first to the last event, plus the mean waiting time. U is approximately

normally distributed with zero mean and unit variance under the null hypothesis of constant

intensity. This is the basis for the given p value.

If p<0.05, a positive U indicates an increasing trend in intensity (decreasing waiting times), while a

negative U indicates a decreasing trend. Note that if a trend is detected by this test, the sequence is

not stationary and the assumptions of the exp test below are violated.

Exp test for Poisson process

The exp test (Prahl 1999) for a stationary Poisson process (random, independent events) is based on

the set of n waiting times Δt_ibetween successive events in the sorted sequence . The test statistic is:

1

∆t







ⁱ

M =

1−

∑

n _{∆t <T}

T





i

235

where T is the mean waiting time. M will tend to zero for a regularly spaced (overdispersed)

sequence, and to 1 for a highly clustered sequence. For the null hypothesis of a Poisson process, M is

asymptotically normally distributed with mean 1/e - α/n and standard deviation β/√n, where

α=0.189 and β=0.2427. This is the basis for the given z test.

In summary, if p<0.05 the sequence is not Poisson. You can then inspect the M statistic; if smaller

than the expected value this indicates regularity, if higher it indicates clustering.

For both tests, p values are also estimated by Monte Carlo simulation with 9999 random data sets.

References

Cox, D. R. & P. A. W. Lewis. 1978. The Statistical Analysis of Series of Events. Chapman and Hall,

London.

Prahl, J. 1999. A fast unbinned test on event clustering in Poisson processes. Arxiv, Astronomy and

Astrophysics September 1999.

236

Markov chain

This module requires a single column containing a sequence of nominal data coded as integer

numbers. For example, a stratigraphic sequence where 1 means limestone, 2 means shale and 3

means sand. A transition matrix containing counts or proportions (probabilities) of state transitions is

displayed. The “from”-states are in rows, the “to”-states in columns.

It is also possible to specify several columns, each containing one or more state transitions (two

numbers for one transition, n numbers for a sequence giving n-1 transitions).

The chi-squared test reports the probability that the data were taken from a system with random

proportions of transitions (i.e. no preferred transitions). The transitions with anomalous frequencies

can be identified by comparing the observed and expected transition matrices.

The “Embedded (no repeats)” option should be selected if the data have been collected in such a

way that no transitions to the same state are possible (data points are only collected when there is a

change). The transition matrix will then have zeroes on the diagonal.

The algorithms, including an iterative algorithm for embedded Markov chains, are according to Davis

(1986).

Reference

Davis, J.C. 1986. Statistics and Data Analysis in Geology. John Wiley & Sons.

237

ARMA (and intervention analysis)

Analysis and removal of serial correlations in time series, and analysis of the impact of an external

disturbance ("intervention") at a particular point in time. Assumes stationary time series, except for a

single intervention. Requires one column of equally spaced data.

This powerful but somewhat complicated module implements maximum-likelihood ARMA analysis,

and a minimal version of Box-Jenkins intervention analysis (e.g. for investigating how a climate

change might impact biodiversity).

By default, a simple ARMA analysis without interventions is computed. The user selects the number

of AR (autoregressive) and MA (moving-average) terms to include in the ARMA difference equation.

The log-likelihood and Akaike information criterion are given. Select the numbers of terms that

minimize the Akaike criterion, but be aware that AR terms are more "powerful" than MA terms. Two

AR terms can model a periodicity, for example.

The main aim of ARMA analysis is to remove serial correlations, which otherwise cause problems for

model fitting and statistics. The residual should be inspected for signs of autocorrelation, e.g. by

copying the residual from the numerical output window back to the spreadsheet and using the

autocorrelation module. Note that for many paleontological data sets with sparse data and

confounding effects, proper ARMA analysis (and therefore intervention analysis) will be impossible.

The program is based on the likelihood algorithm of Melard (1984), combined with nonlinear

multivariate optimization using simplex search.

Intervention analysis proceeds as follows. First, carry out ARMA analysis on only the samples

preceding the intervention, by typing the last pre-intervention sample number in the "last samp"

box. It is also possible to run the ARMA analysis only on the samples following the intervention, by

typing the first post-intervention sample in the "first samp" box, but this is not recommended

because of the post-intervention disturbance. Also tick the "Intervention" box to see the optimized

intervention model.

The analysis follows Box and Tiao (1975) in assuming an "indicator function" u(i) that is either a unit

step or a unit pulse, as selected by the user. The indicator function is transformed by an AR(1)

process with a parameter delta, and then scaled by a magnitude (note that the magnitude given by

PAST is the coefficient on the transformed indicator function: first do y(i) = delta*y(i-1)+u(i), then

scale y by the magnitude). The algorithm is based on ARMA transformation of the complete

sequence, then a corresponding ARMA transformation of y, and finally linear regression to find the

magnitude. The parameter delta is optimized by exhaustive search over [0,1].

For small impacts in noisy data, delta may end up on a sub-optimum. Try both the step and pulse

options, and see what gives smallest standard error on the magnitude. Also, inspect the "delta

optimization" data, where standard error of the estimate is plotted as a function of delta, to see if

the optimized value may be unstable.

The Box-Jenkins model can model changes that are abrupt and permanent (step function with

delta=0, or pulse with delta=1), abrupt and non-permanent (pulse with delta<1), or gradual and

permanent (step with delta<0).

238

Be careful with the standard error on the magnitude - it will often be underestimated, especially if

the ARMA model does not fit well. For this reason, a p value is deliberately not computed (Murtaugh

2002).

The example data set (blue curve) is Sepkoski’s curve for percent extinction rate on genus level,

interpolated to even spacing at ca. 5 million years. The largest peak is the Permian-Triassic boundary

extinction. The user has specified an ARMA(2,0) model. The residual is plotted in red. The user has

specified that the ARMA parameters should be computed for the points before the P-T extinction at

time slot 37, and a pulse-type intervention. The analysis seems to indicate a large time constant

(delta) for the intervention, with an effect lasting into the Jurassic.

References

Box, G.E.P. & G.C. Tiao. 1975. Intervention analysis with applications to economic and environental

problems. Journal of the American Statistical Association 70:70-79.

Melard, G. 1984. A fast algorithm for the exact likelihood of autoregressive-moving average models.

Applied Statistics 33:104-114.

Murtaugh, P.A. 2002. On rejection rates of paired intervention analysis. Ecology 83:1752-1761.

239

Simple smoothers

A set of simple smoothers for a single column of evenly spaced data. See also the spline and LOESS

smoothers in the Model menu. Missing data are supported.

Moving average

Simple N-point, centered moving average (N must be odd). Commonly used, but has unfortunate

properties such as a non-monotonic frequency response.

Gaussian

Weighted moving average using a Gaussian kernel with standard deviation set to 1/5 of the window

size (of N points). This is an overall good method.

Moving median

Similar to moving average but takes the median instead of the mean. This method is more robust to

outliers but can produce a “blocky” appearance.

AR 1 (exponential)

Recursive (autoregressive) filter, y_i= αy_i-1+ (1-α)x_iwith α a smoothing coefficient from 0 to 1. This

corresponds to weighted averaging with exponentially decaying weights. Gives a phase delay and

also a transient in the beginning of the series. Included for completeness.

Savitzky-Golay

The Savitzky-Golay method (Savitzky & Golay 1964) implements least-squares fit to a polynomial

inside a moving window of size n points. Second-order (m = 2) and fourth-order (m = 4) polynomials

240

are included. These are “optimal” smoothers in the sense that they preserve all moments up to m.

This tends to give better preservation of peak values and peak widths than other smoothers. The

implementation in Past is based on Press et al. (1992).

Non-local means

Non-local means is a relatively new, powerful smoothing method, mostly used for image denoising

but also effective for time series (Tracey & Miller 2012). It is an averaging method like the moving

average and Gaussian methods, but the average is taken not over neighboring points but over points

in similar regions, which can be far away. This tends to preserve peaks and transitions better than

local averaging. In Past, the size of the local regions (patch size) can be selected; it could be set to e.g.

N = 7 or N = 13. The search radius is fixed at half the length of the time series. The value of lambda

controls the degree of smoothing. Tracey & Miller (2012) suggest a value of about 0.6 times the

standard deviation of the noise (which is usually unknown but can be estimated by eye).

Hampel

The Hampel filter (Hampel 1974) identifies possible outliers and replaces them with a local median.

The length of the sliding window (“points”) should be at least N = 5. Let k = (N-1)/2. For input sample

x_ithe local median is given by

(

)

푚_ꢄ= median 푥_ꢄ−푘, 푥_ꢄ−푘+1, … , 푥_ꢄ, … , 푥_ꢄ+푘−1, 푥_ꢄ+푘

The local median absolute deviation (MAD) is

푀퐴ꢥ_ꢄ= median(푥_ꢄ−푘− 푚_ꢄ, … , 푥_ꢄ+푘− 푚_ꢄ)

A local standard deviation is estimated as

1

휎_ꢄ=

푀퐴ꢥ_ꢄ≈ 1.4826 푀퐴ꢥ_ꢄ

1

2

2erf⁻¹� �

√

|

Now, if 푥_ꢄ− 푚_ꢄ> 3휎_ꢄthen x_iis considered an outlier, and is replaced by m_i.

References

Hampel, F.R. 1974. The influence curve and its role in robust estimation. Journal of the American

Statistical Association 69, 382-393.

Press, W.H., Teukolsky, S.A., Vetterling, W.T. & Flannery, B.P. 1992. Numerical Recipes in C. 2^nd

edition. Cambridge University Press.

Savitzky, A. & Golay, M.J.E. 1964. Smoothing and differentiation of data by simplified least squares

procedures. Analytical Chemistry 36, 1627-1639.

Tracey, B. & Miller, E. 2012. Nonlocal means denoising of ECG signals. IEEE Transactions on

Biomedical Engineering 59, 2383-2386.

241

FIR filter

Filtering out certain frequency bands in a time series can be useful to smooth a curve, remove slow

variation, or emphasize certain periodicities (e.g. Milankovitch cycles). One column of evenly spaced

data is expected. For most applications in data analysis, it is crucial that the filter has linear phase

response. Past therefore uses FIR (Finite Impulse Response) filters, which are designed using the

Parks-McClellan algorithm. The following filter types are available: Lowpass, highpass, bandpass and

bandstop.

Filter parameters

To design an optimal filter takes a little effort. Frequencies are specified in the range 0-0.5, i.e. T₀/T

where T₀is the sampling interval (not specified to the computer) and T is the required period. For

example, if your real sampling interval is 1,000 years, a frequency corresponding to a period of

23,000 years is specified as 1,000/23,000=0.043.

After setting the filter type, you should select a transition width (or leave the default of 0.02).

Decreasing the transition width will make a sharper filter, at the cost of larger ripple (“waves” in the

frequency response).

Note that the values in text fields are not updated until you press Enter. Also, if an invalid

combination is entered (e.g. a transition band crossing 0 or 0.5, or upper limit less than lower limit)

the program will reset some value to avoid errors. It is therefore required to enter the numbers in an

order so that the filter is always valid.

242

The filter types are as follows:

1. Lowpass. The From frequency is forced to zero. Frequencies up to the To frequency pass the

filter. Frequencies from To+Transition to 0.5 are blocked.

2. Highpass. The To frequency is forced to 0.5. Frequencies above the From frequency pass the

filter. Frequencies from 0 to From-Transition are blocked.

3. Bandpass. Frequencies from From to To pass the filter. Frequencies below From-Transition

and above To+Transition are blocked.

4. Bandstop. Frequencies from From to To are blocked. Frequencies from 0 to From-Transition

and from To+Transition to 0.5 pass the filter.

Filter order

The filter order should be large enough to give an acceptably sharp filter with low ripple. However, a

filter of length n will give less accurate results in the first and last n/2 samples of the time series,

which puts a practical limit on filter order for short series.

The Parks-McClellan algorithm will not always converge. This gives an obviously incorrect frequency

response, and attempting to apply such a filter to the data will give a warning message. Try to change

the filter order (usually increase it) to fix the problem.

Missing values are treated using linear interpolation before analysis.

243

Insolation (solar forcing) model

This module computes solar insolation at any latitude and any time from 100 Ma to the Recent (the

results are less accurate before 50 Ma). The calculation can be done for a "true" orbital longitude,

"mean" orbital longitude (corresponding to a certain date in the year), averaged over a certain

month in each year, or integrated over a whole year.

The implementation in PAST is ported from the code by Laskar et al. (2004), by courtesy of these

authors. Please reference Laskar et al. (2004) in any publications.

It is necessary to specify a data file containing orbital parameters. Download the file

INSOLN.LA2004.BTL.250.ASC from http://vo.imcce.fr/insola/earth/online/earth/earth.html and put

in anywhere on your computer. The first time you run the calculation, PAST will ask for the position

of the file.

The amount of data can become excessive for long time spans and short step sizes!

Reference

Laskar, J., P. Robutel, F. Joutel, M. Gastineau, A.C.M. Correia & B. Levrard. 2004. A long-term

numerical solution for the insolation quantities of the Earth. Astronomy & Astrophysics 428:261-285.

244

Date/time conversion

Utility to convert dates and/or times to a continuous time unit for analysis. The program expects one

or two columns, each containing dates or times. If both are given, then time is added to date to give

the final time value.

Dates can be given in the formats Year/Month/Day or Day/Month/Year. Years need all digits (a year

given as 11 will mean 11 AD, not 2011). Only Gregorian calendar dates are supported. Leap years are

taken into account.

Time can be given as Hours:Minutes or Hours:Minutes:Seconds (seconds can include decimals).

The output units can be years (using the Gregorian mean year of 365.2425 days), days (of 86400

seconds), hours, minutes or seconds.

The starting time (time zero) can be the smallest given time, the beginning of the first day, the

beginning of the first year, year 0 (note the “astronomical” convention where the year before year 1

is year 0), or the beginning of the first Julian day (noon, year -4712).

The program operates with simple (UT) time, defined with respect to the Earth’s rotation and with a

fixed number of seconds (86400) per day.

If your input data consists of space-separated date-time values, such as “2011/12/24 18:00:00.00”,

then you may have to use the “Import text file” function to read the data such that dates and times

are split into separate columns.

The calculation of Julian day (which is used to find number of days between two dates) follows

Meeus (1991):

if month <= 2 begin year := year - 1; month := month + 12; end;

A = floor(year/100);

B = 2 – A + floor(A/4);

JD = floor(365.25(year + 4716)) + floor(30.6001(month+1)) + day + B – 1524.5;

Reference

Meeus, J. 1991. Astronomical algorithms. Willmann-Bell, Richmond.

245

Geometrical menu

Circular (one sample)

The module plots a rose diagram (polar histogram) of directions. Used for plotting current-oriented

specimens, orientations of trackways, fault lines, etc. Also appropriate for time-of day data (0-24

hours).

One column of directional (0-360) or orientational (0-180) data in degrees is expected. Directional or

periodic data in other forms (radians, hours, etc.) must be converted to degrees using e.g. the

Evaluate Expression module (Transform menu).

By default, the 'mathematical' angle convention of anticlockwise from east is chosen. If you use the

'geographical' convention of clockwise from north, tick the box.

You can also choose whether to have the abundances proportional to radius in the rose diagram, or

proportional to area (equal area).

The "Kernel density" option plots a circular kernel density estimate.

Descriptive statistics

The mean angle takes circularity into account:

sinθ

∑

i

θ = tan⁻¹

(taken to the correct quadrant).

cosθ

∑

i

246

The 95 percent confidence interval on the mean is estimated according to Fisher (1983). It assumes

circular normal distribution, and is not accurate for very large variances (confidence interval larger

than 45 degrees) or small sample sizes. The bootstrapped 95% confidence interval on the mean uses

5000 bootstrap replicates. The graphic uses the bootstrapped confidence interval.

The concentration parameter κ is estimated by iterative approximation to the solution to the

equation

I₁(

κ

)

= R

I₀(κ

where I₀and I₁are imaginary Bessel functions of orders 0 and 1, estimated according to Press et al.

(1992), and R defined below (see e.g. Mardia 1972).

Rayleigh’s test for uniform distribution

The R value (mean resultant length) is given by:





²







²



n

R =

cosθ

+

sinθ

n

.

∑



ⁱ



ⁱ

i=1

R is further tested against a random distribution using Rayleigh's test for directional data (Davis

1986). Note that this procedure assumes evenly or unimodally (von Mises) distributed data - the test

is not appropriate for e.g. bimodal data. The p values are computed using an approximation given by

Mardia (1972):

K = nR ²

2K − K ²

24K −132K ²+ 76K ³− 9K ⁴





p = e^−K1+

−





288n²



4n





Rao’s spacing test for uniform distribution

The Rao's spacing test (Batschelet 1981) for uniform distribution has test statistic

n

1

U =

T − λ

,

∑

i

2

i=1

λ = 360^on T = θ

where

.

for i<n, T_n= 360^o−θ_n+θ₁. This test is nonparametric and does

−θ_i

i

i+1

not assume e.g. von Mises distribution. The p value is estimated by linear interpolation from the

probability tables published by Russell & Levitin (1995).

A Chi-square test for uniform distribution is also available, with the number of bins set to 4 if n<40,

otherwise 8 bins.

247

The Watson’s U²goodness-of-fit test for von Mises distribution

Let f be the von Mises distribution for estimated parameters of mean angle and concentration:

(

θ −θ

)

e^{κ cos}

(

)

f θ;θ ,κ =

.

2π I₀(

κ

)

The test statistic (e.g. Lockhart & Stevens 1985) is

2

2i −1

1

2

1















U ²=

z −

− n z −

+



∑

i

2n

12n









where

θ_i

(

)

z_i=

f θ;θ ,κ dθ

,

∫

0

estimated by numerical integration. Critical values for the test statistic are obtained by linear

interpolation into Table 1 of Lockhart & Stevens (1985). They are acceptably accurate for n>=20.

Axial data

The 'Orientations' option allows analysis of linear (axial) orientations (0-180 degrees). The Rayleigh

and Watson tests are then carried out on doubled angles (this trick is described by Davis 1986); the

Chi-square uses four bins from 0-180 degrees; the rose diagram mirrors the histogram around the

origin.

References

Batschelet, E. 1981. Circular statistics in biology. Academic Press.

Davis, J.C. 1986. Statistics and Data Analysis in Geology. John Wiley & Sons.

Fisher, N.I. 1983. Comment on "A Method for Estimating the Standard Deviation of Wind Directions".

Journal of Applied Meteorology 22:1971.

Lockhart, R.A. & M.A. Stephens 1985. Tests of fit for the von Mises distribution. Biometrika 72:647-

652.

Mardia, K.V. 1972. Statistics of directional data. Academic Press, London.

Russell, G. S. & D.J. Levitin 1995. An expanded table of probability values for Rao's spacing test.

Communications in Statistics: Simulation and Computation 24:879-888.

248

Circular (two samples)

The module expects two columns of directional (0-360) or orientational (0-180) data in degrees.

Watson-Williams test

The Watson-Williams test for equal mean angle in two samples is a parametric test, assuming von

Mises distribution, but is fairly robust. The concentration parameter κ should be larger than 1.0 for

accurate testing. In addition, the test assumes similar angular variances (R values).

The two samples φ and θ have n₁and n₂values. The resultant length R is calculated for each sample

and for the combined sample:



²



²

n₁



_i



_i

R₁=

R₂=

R =

cosφ

+

sinφ

∑









i=1



n₂



²



²



_i





_i

cosθ

+

sinθ

∑





i=1







n₁

n₂

n₁

n₂



²



²









cosφ +

cos

θ_i

+

sin

φ_i

+

sin

θ_i

.

∑

i

i=1









249

The test statistic U is computed as

R₁+ R₂− R

.

U = (n − 2)

n − (R₁+ R₂

)

The significance is computed by first correcting U according to Mardia (1972a):

U





R n < 0.45

κ ²

1



1−

+

nκ ²

U =

,



8



3





1+

U

R n < 0.95











8κ



where n = n₁+ n₂. The p value is then given by the F distribution with 1 and n-2 degrees of freedom.

The combined concentration parameter κ is maximum-likelihood, computed as described under

“Directions (one sample)” above.

Mardia-Watson-Wheeler test

This non-parametric test for equal distribution is computed according to Mardia (1972b).

2





C₁+ S₁

C₂²+ S₂²





W = 2

+

n₁

n₂





where, for the first sample,

n₁

,

C =

cos(2π r N)

S =

sin (2π r N)

∑

1

1i

1

1i

i=1

and similarly for the second sample (N=n₁+n₂). The r_1iare the ranks of the values of the first sample

within the pooled sample.

For N>14, W is approximately chi-squared with 2 degrees of freedom.

References

Mardia, K.V. 1972a. Statistics of directional data. Academic Press, London.

Mardia, K.V. 1972b. A multi-sample uniform scores test on a circle and its parametric competitor.

Journal of the Royal Statistical Society Series B 34:102-113.

250

Circular correlation

Testing for correlation between two directional or orientational variates. Assumes “large” number of

observations. Requires two columns of directional (0-360) or orientational (0-180) data in degrees.

This module uses the circular correlation procedure and parametric significance test of

Jammalamadaka & Sengupta (2001).

The circular correlation coefficient r between vectors of angles α and β is

n

)

sin(α_i−α )sin(β_i− β

∑

i=1

,

r =

n

sin²

(

α_i−α

)

sin²(β_i− β

)

∑

i=1

where the angular means are calculated as described previously. The test statistic T is computed as

n

sin²

(

α_k−α

)

sin²(β_k− β

)

∑

k=1

.

T = r

n

sin²

(

α_k−α

)

sin²(β_k− β

)

∑

k=1

For large n, this statistic has asymptotically normal distribution with mean 0 and variance 1 under the

null hypothesis of zero correlation, which is the basis for the calculation of p.

Reference

Jammalamadaka, S.R. & A. Sengupta. 2001. Topics in circular statistics. World Scientific.

251

Spherical (one sample)

This module makes stereo plots of axial, spherical data (e.g. strike-dip measurements in structural

geology), and performs the Bingham test for uniform distribution.

Three data formats can be used, all using the geographic angle convention (degrees, clockwise from

north):

•

Trend (azimuth) and plunge (angle down from the horizontal) for axial data

Dip azimuth and dip angle (down from the horizontal) for planes. The pole (normal vector) of

the plane is plotted.

•

Strike and dip for planes, using the right-hand rule convention with the dip down to the right

from the strike. The pole to the plane is plotted.

Density contouring is based on a modified Kamb method algorithm by Vollmer (1995). Both equal

area (Schmidt) and equal angle (Wulff) projections are available. Projections are to the lower

hemisphere. Density estimates can use an inverse area, inverse area squared or exponential law,

giving progressively higher smoothing.

The Bingham test for uniform distribution of axial data can be used to test for preferred direction

(Bingham 1974; Mardia & Jupp 2000, p. 232-233). Past computes the S statistic as follows.

The sample scatter matrix is computed as

푛

1

푻 = � 풙_풊풙^푻

�

풊

ꢀ

ꢄ=1

252

where x_iis the 3-vector of direction cosines for sample i. Then,

15

2

1

− �

3

2

�

ꢀ �tr 푻

(

)

ꢋ =

An adjusted S is computed according to Jupp (2001):

1

(

2

ꢋ^∗= ꢋ �1 −

ꢍ₀+ ꢍ₁ꢋ + ꢍ₂ꢋ �

)

ꢀ

where

31

41

5

ꢍ₀=

,

ꢍ₁= −

,

ꢍ₂=

42

294

1323

When n<=50 and S<=30, the reported p value for uniformity is estimated from the S* value using the

chi-squared distribution with 5 degrees of freedom. Otherwise, the S value is used (the polynomial

equation for S* turns the wrong way for large S). A small p (e.g. p<0.05) means significantly non-

uniform distribution.

References

Bingham, C. 1974. An antipodally symmetric distribution on the sphere. Annals of Statistics 2:1201-

1225.

Jupp, P.E. 2001. Modifications of the Rayleigh and Bingham tests for uniformity of directions. Journal

of Multivariate Analysis 77:1-20.

Mardia, K.V. & Jupp, P.E. 2000. Directional Statistics. John Wiley & Sons.

Vollmer, F.W. 1995. C program for automatic contouring of spherical orientation data using a

modified Kamb method. Computers & Geosciences 21:31-49.

253

Point pattern analysis - nearest neighbours

This module tests for clustering or overdispersion of points given as two-dimensional coordinate

values. The procedure assumes that elements are small compared to their distances, that the domain

is predominantly convex, and n>50. Two columns of x/y positions are required. Applications of this

module include spatial ecology (are in-situ brachiopods clustered), morphology (are trilobite

tubercles overdispersed), and geology (distribution of e.g. volcanoes, earthquakes, springs).

The calculation of point distribution statistics using nearest neighbour analysis follows Davis (1986)

with modifications. The area is estimated either by the smallest enclosing rectangle or using the

convex hull, which is the smallest convex polygon enclosing the points. Both are inappropriate for

points in very concave domains. Two different edge effect adjustment methods are available: wrap-

around ("torus") and Donnelly's correction. Wrap-around edge detection is only appropriate for

rectangular domains.

The null hypothesis is a random Poisson process, giving a modified exponential nearest neighbour

distribution (see below) with mean

A n

µ =

2

where A is the area and n the number of points.

The probability that the distribution is Poisson is presented, together with the R value:

254

d

2d

R =

=

µ

A n

where

is the observed mean distance between nearest neighbours. Clustered points give R<1,

d

Poisson patterns give R~1, while overdispersed points give R>1.

The expected (theoretical) distribution under the null hypothesis is plotted as a continuous curve

together with the histogram of observed distances. The expected probability density function as a

function of distance r is

g

(

r

)

= 2ρπr exp

(

− ρπr²

)

where ρ = n/A is the point density (Clark & Evans 1954).

The orientations (0-180 degrees) and lengths of lines between nearest neighbours, are also included.

The orientations can be subjected to directional analysis to test whether the points are organized

along lineaments (see Hammer 2009 for more advanced methods).

References

Clark, P.J. & Evans, F.C. 1954. Distance to nearest neighbor as a measure of spatial relationships in

populations. Ecology 35:445-453.

Davis, J.C. 1986. Statistics and Data Analysis in Geology. John Wiley & Sons.

Hammer, Ø. 2009. New methods for the statistical analysis of point alignments. Computers &

Geosciences 35:659-666.

255

Ripley’s K point pattern analysis

Ripley's K (Ripley 1979) is the average point density as a function of distance from every point. It is

useful when point pattern characteristics change with scale, e.g. overdispersion over small distances

but clustering over large distances. Two columns of x/y coordinates in a rectangular domain are

expected.

Define the estimated intensity of the point pattern, with n points in an area A, as

. The

λ = n A

distance between points i and j is d_ijThe estimate of Ripley’s K, as a function of distance, is then

computed as

n

1

,

K

(

d

)

=

I

(d_ij≤ d)

_λn∑∑

i=1 j≠i

where the indicator function I is one if the argument is true, zero otherwise.

The normalization of K is such that for complete spatial randomness (CSR), K(d) is expected to

K

(

d

)

= πd ²

increase as the area of circles, i.e.

. The L(d) function is a corresponding transformation

of K(d):

K

(

d

)

L

(

d

)

=

π

For CSR, L(d)=d, and L(d)-d=0. A 95% confidence interval for CSR is estimated using 1000 Monte Carlo

simulations within the bounding rectangle (previous versions used the approximation 1.42 A n ).

256

Ripley's edge correction is included, giving weights to counts depending on the proportion of the test

circle that is inside the rectangular domain.

The example above shows locations of volcanic pipes. L(d)-d is below the 95% confidence interval of

CSR, indicating lateral inhibition, up to a distance of ca. 70 m. For larger distances, the curve flattens

in the manner expected from CSR.

Area

For the correct calculation of Ripley's K, the area must be known. In the first run, the area is

computed using the smallest bounding rectangle, but this can both over- and underestimate the real

area. The area can therefore be adjusted by the user. An overestimated area will typically show up as

a strong overall linear trend with positive slope for L(d)-d.

Fractal dimension

The fractal dimension (if any) can be estimated as the asymptotic linear slope in a log-log plot of K(d).

For CSR, the log-log slope should be 2.0. Fractals should have slopes less than 2.

References

Ripley, B.D. 1979. Tests of 'randomness' for spatial point patterns. Journal of the Royal Statistical

Society, ser. B 41:368-374.

257

Correlation length analysis

Correlation length analysis (Cartwright & Whitworth 2004; Cartwright et al. 2011) investigates the

spatial distribution of a point pattern at different scales, and is an alternative to Ripley’s K. Two

columns of x/y coordinates in a rectangular domain are expected. CLA is simply a histogram of all

pairwise distances between points, i.e. a total of N(N-1)/2 distances (black curve).

The expected curve from a random point pattern (blue curve) and its 95% confidence interval (red

curves) are computed from 1000 Monte Carlo simulations of complete spatial randomness (CSR) in a

rectangle of the same dimensions as the bounding rectangle of the original data. Thus, distances

where the CLA curve from the data (black curve) exceeds the upper red curve, have significantly

higher frequencies than expected from a random pattern.

An overall significance test is based on the observed total mean distance compared with the

expected mean distance from the Monte Carlo simulations.

The number of bins can be set by the user and should be small to reduce noise, but large enough to

capture details.

The “Residual” option flattens the curves on the expected mean (blue curve), i.e. the expected mean

is subtracted from the curves at all distances. This can make the figure clearer especially when the

confidence interval is narrow.

A comparison between Ripley’s K and correlation length analysis for a geological data set is given by

Cartwright et al. (2011).

References

Cartwright, A. & Whitworth, A.P. 2004. The statistical analysis of star clusters. Monthly Notices of the

Royal Astronomical Society 348:589-597.

Cartwright, A., Moss, J. & Cartwright, J. 2011. New statistical methods for investigating submarine

pockmarks. Computers & Geosciences 37:1595-1601.

258

Minimal spanning tree analysis

Minimal spanning tree analysis (Cartwright & Whitworth 2004; Cartwright et al. 2011) investigates

the spatial distribution of a point pattern with focus on small scales, comparable to nearest neighbor

analysis but with somewhat different properties. Two columns of x/y coordinates in a rectangular

domain are expected. The method is based on a histogram of all the lengths of line segments in the

minimal spanning tree (MST). The MST itself can be plotted in the XY graph module (Plot menu).

The expected curve from a random point pattern (blue curve) and its 95% confidence interval (red

curves) are computed from 1000 Monte Carlo simulations of complete spatial randomness (CSR) in a

rectangle of the same dimensions as the bounding rectangle of the original data. Thus, segment

lengths where the histogram from the data (black curve) exceeds the upper red curve, have

significantly higher frequencies than expected from a random pattern.

An overall significance test is based on the observed total mean length compared with the expected

mean length from the Monte Carlo simulations.

The number of bins can be set by the user and should be small to reduce noise, but large enough to

capture details.

The “Residual” option flattens the curves on the expected mean (blue curve), i.e. the expected mean

is subtracted from the curves at all distances. This can make the figure clearer especially when the

confidence interval is narrow.

A comparison between nearest neighbour and minimal spanning tree analysis for a geological data

set is given by Cartwright et al. (2011).

References

Cartwright, A. & Whitworth, A.P. 2004. The statistical analysis of star clusters. Monthly Notices of the

Royal Astronomical Society 348:589-597

Cartwright, A., Moss, J. & Cartwright, J. 2011. New statistical methods for investigating submarine

pockmarks. Computers & Geosciences 37:1595-1601.

259

Kernel density

Makes a smooth map of point density in 2D. Two columns of x/y coordinates in a rectangular domain

are expected. The user can specify the size of the grid (number of rows and columns). The “Radius”

value sets the scale r of the kernel. There is currently no automatic selection of “optimal” radius, so

this value must be set by the user depending on the scale of interest.

The density estimate is based on one of four kernel functions, with radius parameter r. With

2

:

d_i= (x − x_i

)

+

(y − y_i

)

d_i²









1





Gaussian (default):

Paraboloid:

f

(x, y)

=

exp −

∑

πr²

2r²

i



d_i²

r²



3

2πr²



1

−

d_i

≤

>

r

(x, y)



∑

i





0

d_i

r



2



1−

d_i≤ r

f

(x, y)

=

Triangular:

Uniform:



∑

πr²

i





0

d_i> r

1 d ≤ r







1

πr²

i

(x, y)

∑

0 d_i> r

i

260

The scaling gives an estimate of the number of points per area, not a probability density. The

gaussian and paraboloid (quadratic) kernels usually perform best. The uniform kernel gives very low

smoothness.

Point alignments

Detection of linear alignments in a 2D point pattern, using the continuous sector method (Hammer

2009). Typical applications are in geology and geography, to study the distribution of earthquakes,

volcanoes, springs etc. associated with faults and other linear structures.

The Radius parameter sets the scale of analysis. In the example above, lineaments of length 1200 m

(twice the radius) are detected.

Alpha sets the significance level for the Rayleigh test used by the procedure. Note that this is a

pointwise significance, not corrected for the multiple testing of all the points.

The Dispersion filter disables alignments with uneven distribution of points along the lineament.

View numbers lists the alignment positions and their orientations, which can be subjected to circular

statistics if required (Directions module).

Reference

Hammer, Ø. 2009. New methods for the statistical detection of point alignments. Computers &

Geosciences 35:659-666.

261

Quadrat counts

This module provides statistics on the distribution of points in quadrats. The input data consist of a

single column of counts of points in equal-sized quadrats (the order is arbitrary). For a random point

pattern, the data are expected to follow a Poisson distribution.

The Morisita index (Morisita 1959) is expected to have value Id=1 for a random pattern, Id<1 for an

overdispersed (spaced) pattern, and Id>1 (up to Id=n) for a clustered pattern. It is computed as

2

∑

푥 −

푥

퐼_휕= ꢀ

(∑ )²

∑

푥

−

푥

where n is the number of quadrats, and x are the counts. The significance test follows Morisita

(1959), with F ratio

(∑

)

∑

퐼_휕

푥 − 1 + ꢀ −

푥

ꢎ₀=

ꢀ − 1

The degrees of freedom are n-1 and ∞. In addition, a Monte Carlo test is carried out with 9999

replicates, each with random distribution of points on quadrats.

The 95% confidence limits (lower and upper) around Id=1 (random pattern) are called the uniform

and the clumped indices, respectively (Krebs 1999):

휒₀²_.975− ꢀ +

푥

∑

푀_ꢯ=

∑

푥 − 1

휒₀²_.025− ꢀ +

∑

푀_푐=

∑

푥 − 1

Where χ²_0.975is the 97.5 percentile point of the chi-squared distribution with n-1 degrees of freedom.

The Standardized Morisita Index, MIS, was suggested by Smith-Gill (1975). It ranges from -1 to 1, with

MIS=0 for a random pattern and with 95% confidence limits [-0.5, 0.5]. It is calculated as follows:

퐼_ꢵ− 푀_푐

퐼_ꢵ≥ 푀_푐> 1: 푀퐼ꢋ = 0.5 + 0.5

ꢀ − 푀_푐

퐼_ꢵ− 1

푀_푐> 퐼_ꢵ≥ 1: 푀퐼ꢋ = 0.5

푀_푐− 1

퐼_ꢵ− 1

1 > 퐼_ꢵ> 푀_ꢯ: 푀퐼ꢋ = −0.5

푀_ꢯ− 1

퐼_ꢵ− 푀_ꢯ

1 > 푀_ꢯ> 퐼_ꢵ: 푀퐼ꢋ = −0.5 + 0.5

푀_ꢯ

262

References

Krebs, C.J. 1999. Ecological Methodology, 2nd ed. Benjamin Cummings Publishers.

Morisita, M. 1959. Measuring of the dispersion of individuals and analysis of the distributional

patterns. Memoirs of the Faculty of Science, Kyushu University, Series E (biology) 2:215-235.

Smith-Gill, S. J. 1975. Cytophysiological basis of disruptive pigmentary patterns in the leopard frog,

Rana pipiens. II. Wild type and mutant cell specific patterns. Journal of Morphology 146:35–54.

263

Spatial autocorrelation (Moran’s I)

Spatial autocorrelation in Past requires three columns, containing x and y coordinates and

corresponding data values z for a number of points. The Moran’s I correlation statistic is then

computed within each of a number of distance classes (bins), ranging from small to large distances.

The one-tailed critical value for p<0.05 can be plotted for each bin. Moran’s I values exceeding the

critical value may be considered significant, but Bonferroni or other adjustment for multiple testing

should be considered because of the several bins.

The calculation follows Legendre & Legendre (1998). For each distance class d, compute

n

1

w

_hi(z_h− z)(z_i− z)

∑∑

W

h=1 i=1

.

I

(

d

)

=

n

1

2

(z_i− z)

∑

n

i=1

Here, n is the total number of points, W is the number of pairs of points having distances within the

distance class, and w_hia weight function such that w_hi=1 if points h and i are within the distance class

and w_hi=0 otherwise (Kronecker delta). Note that this equation is incorrect in some publications.

For the one-tailed critical level I_0.05, compute

264

n

1

2

S₁=

(w_hi+ w_ih

)

₂∑∑

h=1 i=1

n

2

S =

(w_i++ w_+i

)

∑

2

i=1

n

4

n

(z_i− z)

∑

i=1

b₂=





²

n

²

(z_i− z)

∑

i=1





n

[

(

n²− 3n + 3

)

S₁− nS₂+ 3W ²]− b₂[

(

n²− n

)

S₁− 2nS₂+ 6W ²

]

1

Var(

I

)

=

−

2

(n −1)(n − 2)(n − 3)W ²

(n −1)

−1

I_0.05= 1.6452 Var(

I

)

− k_0.05(n −1)

Here the w_i+and w_+iare the row and column sums. The correction factor k_0.05is set to

10⋅0.05 = 0.707 if

(

)

(

)

, otherwise k_0.05=1.

4 n − n < W ≤ 4 2n − 3 n +1

Reference

Legendre, P. & Legendre, L. 1998. Numerical Ecology, 2nd English ed. Elsevier, 853 pp.

265

Gridding (spatial interpolation)

“Gridding” is the operation of spatial interpolation of scattered 2D data points onto a regular grid.

Three columns with position (x,y) and corresponding data values are required.

Gridding produces a map showing a continuous spatial estimate of some variate such as fossil

abundance or thickness of a rock unit, based on scattered data points. The user can specify the size

of the grid (number of rows and columns). The spatial coverage of the map is set to a square covering

the data points. When plotting, this can be reduced to the convex hull of the points.

A least-squares linear surface (trend) is automatically fitted to the data, removed prior to gridding

and finally added back in. This is primarily useful for the semivariogram modelling and the kriging

method.

Cross validation: This option will remove each data point in turn and re-compute the surface based

on the remaining points (“jackknife”). The differences between the original data values and the cross-

validated values indicate the prediction accuracy of the surface model. These differences are

reported for each point, together with the mean squared error (MSE) over all points.

Four interpolation algorithms are available:

Inverse distance weighting

The value at a grid node is simply the average of the N closest data points, as specified by the user

(the default is to use all data points). The points are weighted in inverse proportion to distance. This

algorithm is fast but will not always give good (smooth) results. A typical artefact is “bull’s eyes”

around data points. One advantage is that the interpolated values will never exceed the range of the

data points. By setting N=1, this algorithm reduces to the nearest-neighbour method, which sets the

value at a grid node to the value of the nearest data point.

266

Thin-plate spline

Maximally smooth interpolator. Can overshoot in the presence of sharp bends in the surface. This is a

radial basis method with radial basis function φ = r ln r.

Multiquadric

Radial basis function φ = r. Popular for terrain modelling.

Kriging

The user is required to specify a model for the semivariogram, by choosing one of four common

models and corresponding parameters to fit the empirical semivariances (the residual sum of squares

should be as small as possible). The semivariogram is computed within each of a number of bins.

Using the histogram option, choose a number of bins so that each bin (except possibly the rightmost

ones) contains at least 30 distances.

The nugget is a constant added to the model. It implies non-zero variance at zero distance and will

therefore allow the surface to not pass exactly through the given data points. The range controls the

extent of the curve along the distance axis. In the equations below, the normalized distance value h

represents distance/range. The scale controls the extent of the curve along the variance axis.



3h

1









nugget + scale

− h³

h < 1

Spherical:

γ

(

h

)

=



2









nugget + scale

h ≥ 1

γ

(

h

)

= nugget + scale

(

1− e^−h

= nugget + scale

(

1− e^−h

)

Exponential:

2

Gaussian:

(

h

)

nugget + scale(7h²− 8.75h³+ 3.5h⁵− 0.75h⁷

)

h < 1



Cubic:

γ

(

h

)

=





nugget + scale

h ≥ 1

The “Optimize all” button will select the model and parameters giving the smallest residual sum of

squares in the semivariogram. This may not be what you want: You may wish to use a specific model

or to have zero nugget to ensure exact interpolation. You must then set the values manually.

The kriging procedure also provides an estimate of standard errors across the map (this depends on

an accurate semivariogram model). Kriging in PAST does not provide for anisotropic semivariance.

Warning: Kriging is slow, do not attempt it for more than ca. 1000 data points on a 100x100 grid.

See e.g. Davis (1986) or de Smith et al. (2009) for more information on gridding.

References

Davis, J.C. 1986. Statistics and Data Analysis in Geology. John Wiley & Sons.

de Smith, M.J., M.F. Goodchild & P.A. Longley. 2009. Geospatial Analysis, 3^rded. Matador.

267

Multivariate allometry

This module is used for investigating allometry in a multivariate morphometric data set. It expects a

multivariate data set with variables (distance measurements) in columns, specimens in rows.

This method for investigating allometry in a multivariate data set is based on Jolicoeur (1963) with

extensions by Kowalewski et al. (1997). The data are (automatically) log-transformed and subjected

to PCA. The first principal component (PC1) is then regarded as a size axis (this is only valid if the

variation accounted for by PC1 is large, say more than 80%). The allometric coefficient for each

original variable is estimated by dividing the PC1 loading for that variable by the mean PC1 loading

over all variables.

95% confidence intervals for the allometric coefficients are estimated by bootstrapping specimens.

2000 bootstrap replicates are made.

Missing data is supported by column average substitution.

References

Jolicoeur, P. 1963. The multivariate generalization of the allometry equation. Biometrics 19:497-499.

Kowalewski, M., E. Dyreson, J.D. Marcot, J.A. Vargas, K.W. Flessa & D.P. Hallmann. 1997. Phenetic

discrimination of biometric simpletons: paleobiological implications of morphospecies in the lingulide

brachiopod Glottidia. Paleobiology 23:444-469.

268

PCA of 2D landmarks (relative warps)

This module is very similar to the standard PCA module, but with some added functionality for

analyzing 2D landmark configurations. The expected data are specimens in rows, alternating x and y

coordinates in columns. Procrustes standardization recommended.

The relative warps (principal components) are ordered according to importance, and the first and

second warps are usually the most informative. Note that this module does a straightforward PCA of

the landmarks, meaning that the affine component is included in the analysis.

The relative warps are visualized with vectors and/or thin-plate spline transformation grids. When

you increase or decrease the score factor away from zero, the original landmark configuration and

grid will be progressively deformed according to the selected relative warp. Vectors are drawn from

the mean to the deformed (dot) landmark position.

269

Thin-plate splines for 2D landmarks

This module shows a shape deformation from one landmark configuration to another. The expected

data are specimens in rows, alternating x and y coordinates in columns. Procrustes standardization

recommended.

Any shape selected in the “From shape” menu, is taken as a reference, with an associated square

grid. The warps from this to all other specimens can be viewed. You can also choose the mean shape

as the reference.

The ‘Expansion factors’ option will display the area expansion (or contraction) factor around each

landmark in yellow numbers, indicating the degree of local growth. This is computed using the

Jacobian of the warp. Also, the expansions are colour-coded for all grid elements, with green for

expansion and purple for contraction.

At each landmark, the principal strains can also be shown, with the major strain in black and minor

strain in brown. These vectors indicate directional stretching.

A description of thin-plate spline transformation grids is given by Dryden & Mardia (1998).

Partial warps and scores

From the thin-plate spline window, you can choose to see the partial warps for a particular spline

deformation. The first partial warp will represent some long-range (large scale) deformation of the

grid, while higher-order warps are normally connected with more local deformations. The affine

component of the warp (also known as zeroth warp) represents linear translation, scaling, rotation,

and shearing.

When you increase the amplitude factor from zero, the original landmark configuration and a grid

will be progressively deformed according to the selected partial warp.

The partial warp scores of all the specimens are given in a table. Each partial warp score has two

components (x and y).

Reference

Dryden, I.L. & K.V. Mardia 1998. Statistical Shape Analysis. Wiley.

Linear regression of 2D landmarks

Expects specimens in rows, with a single column of independent data (e.g. size) followed by pairs of

columns with Procrustes-fitted landmark positions. Output includes deformation grids and

displacement vectors, drawn from the mean to the deformed (dots) landmark positions.

270

Common Allometric Component analysis for 2D landmarks

Common Allometric Component (CAC) analysis for landmarks was first suggested by Mitteroecker et

al. (2004). The principle is simple and logical: Do a linear regression of shape as a function of size (the

allometric component) and then a PCA on the residuals (the residual shape components). The

required data are one column of sizes, followed by pairs of columns containing Procrustes-fitted x-y

coordinates of the landmarks. The data can be obtained from the original landmarks using the

“Transform->Landmarks->Procrustes” function, and selecting “Add size column”. A final feature of

CAC analysis is that if groups are specified, then the data are centered on group means prior to

analysis, effectively removing inter-group variation.

X is the nxm matrix of (Procrustes-fitted) shape coordinates, centered on group means. s is the n-

vector with the logarithms of centroid sizes (in Past you can also choose to skip the log-transform of

size). The vector of the “common allometric component” is

퐗^ꢡ퐬

퐚 =

퐬^ꢡ퐬

normalized as a’=a/√a^ta. This common allometric component is projected out to produce a reduced

data matrix

(

)^ꢡ)

퐖 = 퐗 퐈 − 풂′ 풂′

The principal components of W constitute the residual shape components. In Past, you can produce

scatter plots of scores on the common allometric component and the residual shape components,

and landmark deformations along all components can be visualized using vector displacements and

thin-plate spline grids, relative to mean shape.

Reference

Mitteroecker, P., Gunz, P., Bernhard, M., Schaefer, K., Bookstein, F.L. 2004. Comparison of cranial

ontogenetic trajectories among great apes and humans. Journal of Human Evolution 46:679-698.

271

PCA of 3D landmarks

Specimens in rows, 3D landmarks in triplets of columns (should be Procrustes-fitted first). The

module is similar to the standard PCA module, but allows visualization of the principal components

as 3D vectors (arrows) away from the mean configuration.

Linear regression of 3D landmarks

Expects specimens in rows, with a single column of independent data (e.g. size) followed by triplets

of columns with Procrustes-fitted landmark positions. Output includes 3D plot of the displacement

vectors, drawn from the mean to the deformed landmark positions.

Common Allometric Component analysis for 3D landmarks

See above for a description of CAC for 2D landmarks. Expects specimens in rows, with a single

column of sizes followed by triplets of columns with Procrustes-fitted landmark positions. In this

module, the displacement vectors can be visualized in 3D.

272

Size from landmarks (2D or 3D) NOT YET IN PAST 4

Digitized x/y or x/y/z landmark coordinates. Specimens in rows, coordinates with alternating x and y

(and z for 3D) values in columns. Must not be Procrustes fitted or normalized for size!

Calculates the centroid size for each specimen (Euclidean norm of the distances from all landmarks to

the centroid).

The values in the 'Normalized' column are centroid sizes divided by the square root of the number of

landmarks - this might be useful for comparing specimens with different numbers of landmarks.

Normalize size

The 'Normalize size' option in the Transform menu allows you to remove size by dividing all

coordinate values by the centroid size for each specimen. For 2D data you may instead use

Procrustes coordinates, which are also normalized with respect to size.

See Dryden & Mardia (1998), p. 23-26.

Reference

Dryden, I.L. & K.V. Mardia 1998. Statistical Shape Analysis. Wiley.

273

Distance from landmarks (2D or 3D) NOT YET IN PAST 4

Digitized x/y or x/y/z landmark coordinates. Specimens in rows, coordinates with alternating x and y

(and z for 3D) values in columns. May or may not be Procrustes fitted or normalized for size.

Calculates the Euclidean distances between two fixed landmarks for one or many specimens. You

must choose two landmarks - these are named according to the name of the first column for the

landmark (x value).

All distances from landmarks (EDMA) NOT YET IN PAST 3

Digitized x/y or x/y/z landmark coordinates. Specimens in rows, coordinates with alternating x and y

(and z for 3D) values in columns. May or may not be Procrustes fitted or normalized for size.

This function will replace the landmark data in the data matrix with a data set consisting of distances

between all pairs of landmarks, with one specimen per row. The number of pairs is N(N-1)/2 for N

landmarks. This transformation will allow multivariate analysis of distance data, which are not

sensitive to rotation or translation of the original specimens, so a Procrustes fitting is not mandatory

before such analysis. Using distance data also allows log-transformation, and analysis of fit to the

allometric equation for pairs of distances.

Missing data is supported by column average substitution.

274

Edit landmark lines/polygons

This function in the Geometry menu under Landmarks (2D) and Landmarks (3D) allows the selection

of landmarks to be linked with lines in the morphometric plots (PCA, thin-plate splines, etc.), to

improve readability. The landmarks must be present in the main spreadsheet before links can be

defined.

The window contains a list of lines or closed polygons. In the example below, the user has specified

one polygon (consisting of four landmarks) and one line (two landmarks). Click on a landmark and

press the Delete key to delete it from the list. The data will be written to the Past file when saved.

275

Elliptic Fourier shape analysis

Requires digitized x/y coordinates around outlines. Specimens in rows, coordinates of alternating x

and y values in columns. Elliptic Fourier shape analysis is in several respects superior to simple

Fourier shape analysis. One advantage is that the algorithm can handle complicated shapes which

may not be expressible as a unique function in polar co-ordinates. Elliptic Fourier shapes is now a

standard method of outline analysis. The algorithm used in PAST is described by Ferson et al. (1985).

EFA coefficients

Cosine and sine components of x and y increments along the outline for the first 30 harmonics are

given, but only the first N/2 harmonics should be used, where N is the number of digitized points.

Size and positional translation are normalized away, and do not enter in the coefficients. The size

(before normalization) is given in the first column. The optional standardization for rotation or

starting point, following Ferson et al., sometimes flips shapes around . This should be checked with

the ‘Shape view’ (see below) – it may be necessary to remove such specimens.

The coefficients can be copied to the main spreadsheet for further analysis such as PCA and

discriminant analysis.

The ‘Shape view’ window allows graphical viewing of the elliptic Fourier shape approximation(s).

EFA PCA

Principal Components Analysis of the EFA coefficients of the given outlines, with visualization of the

principal components as EFA deformations. For more details on PCA in Past, see the description of

PCA.

Reference

Ferson, S.F., F.J. Rohlf & R.K. Koehn. 1985. Measuring shape variation of two-dimensional outlines.

Systematic Zoology 34:59-68.

276

Hangle Fourier shape analysis

Requires digitized x/y coordinates around outlines. Specimens in rows, coordinates of alternating x

and y values in columns.

The “Hangle” method for analysing closed outlines, proposed by Haines & Crampton (2000) is a

competitor to Elliptic Fourier Analysis. Hangle has certain advantages over EFA, the most important

being that fewer coefficients are needed to capture the outline to a given precision. This is of

importance for statistical testing (e.g. MANOVA) and discriminant analysis. The implementation in

Past is based on the Hangle/Hmatch/Htree/Hshape package of Haines & Crampton (thanks to the

authors for providing the source code).

The output consists of 46 Fourier coefficients, which are the cos and sin coefficients of the first 24

harmonics (modes), starting on harmonic number 2. Copy these numbers back to a Past spreadshhet

for further multivariate shape analysis.

Starting point normalization

Usually leave at ‘Match all’, either with the ‘Hmatch’ or (perhaps preferably) the ‘Htree’ method to

align all the outlines. Alternatively, select 2.-4. harmonic, which will phase shift each outline

according to the selected mode (see Haines & Crampton 2000).

Smoothing

Increasing the smoothing parameter can reduce high-frequency noise, at the cost of dampening

potentially informative high-frequency shape information.

Shape view

Use this function to inspect the shapes reconstructed from the Fourier coefficients. Check that the

matching routine has not rotated any shape incorrectly. Also, use this function to select the minimum

number of modes necessary for capturing the shape. In the example above, the number of modes

has been set to 14, which captures 99.88% of the total integrated power (amplitude squared) of the

selected shape. The number of modes is shown by the red line in the power spectrum – make sure

that the main features of the spectrum are to the left of this line for all the shapes.

Note: PCA visualization and regression (as for EFA) has not yet been implemented for Hangle.

Reference

Haines, A.J. & J.S. Crampton. 2000. Improvements to the method of Fourier shape analysis as applied

in morphometric studies. Palaeontology 43:765-783

277

Eigenshapes

Eigenshape analysis (Lohmann, 1983; Rohlf, 1986; MacLeod, 1999) can, somewhat imprecisely, be

thought of as PCA of the raw outlines, without going through a transformation stage such as Fourier

analysis. Past carries out the following steps:

1. Produce an equally spaced set of points by interpolation between the original points. The number

of interpolated points along the outline is optimized automatically.

2. Going along the outline from a fixed point (homologous on all outlines), calculate the tangent

angle from one point to the next. For m interpolated points on a closed outline there are m tangent

angles, constituting a vector φ describing the shape.

Given a set of n equally spaced (x, y)-coordinates around a contour, the tangent angles are calculated

in the following way (modified after Zahn and Roskies, 1972). First, compute the angular orientations

of the tangent to the curve:

푦_ꢄ− 푦_ꢄ−1

휑_ꢄ= tan⁻¹

푥_ꢄ− 푥_ꢄ−1

For a closed contour of n points, we use

푦₁− 푦_푛

휑₁= tan⁻¹

푥₁− 푥_푛

This vector is then normalized by subtracting the angular orientations that would be observed for a

circle:

(

)

2휋 푖 − 1

휑^∗= 휑_ꢄ− 휑₁−

ꢄ

ꢀ

3. The shape vectors for the n shapes are subjected to PCA, giving principal components that are

referred to as eigenshapes. The eigenshapes are themselves tangent angle vectors, given in

decreasing order of amount of shape variation they explain. The first (most important) eigenshapes

define a low-dimensional space into which the original specimens can be projected.

References

Lohmann, G.P. 1983. Eigenshape analysis of microfossils: a general morphometric method for

describing changes in shape. Mathematical Geology 15: 659–672.

MacLeod, N. 1999. Generalizing and extending the eigenshape method of shape space visualization

and analysis. Paleobiology 25: 107–138.

Rohlf, F.J. 1986. Relationships among eigenshape analysis, Fourier analysis and analysis of

coordinates. Mathematical Geology 18: 845–854.

Zahn, C.T., Roskies, R.Z. 1972. Fourier descriptors for plane closed curves. IEEE Transactions,

Computers C-21: 269–281.

278

Coordinate transformation

Conversion between geographical coordinates in different grids and datums. The number of input

columns depends on the data type, as described below.

Decimal degrees (WGS84)

Two columns: Latitude and longitude, in decimal degrees (60.5 is 60 degrees, 30 minutes). Negative

values for south of equator and west of Greenwich. Referenced to the WGS84 datum.

Deg/decimal mins (WGS84)

Four columns: Latitude degrees, deci mal minutes (40.5 is 40 minutes, 30 seconds), longitude

degrees, decimal minutes. Referenced to the WGS84 datum.

Deg/min/sec (WGS84)

Six columns: Latitude degrees, minutes, seconds, longitude degrees, minutes, seconds. Referenced to

the WGS84 datum.

UTM-ED50 (Intl 1924)

Three columns: Easting (meters), northing (meters), and zone. Use negative zone numbers for the

southern hemisphere. The handling of UTM zones takes into account the special cases of Svalbard

and western Norway. Referenced to the ED50 European datum at Potsdam.

UTM-WGS84 (WGS84)

Three columns: Easting (meters), northing (meters), and zone. Referenced to the WGS84 datum.

UTM-NAD27 (Clarke 1866)

Three columns: Easting (meters), northing (meters), and zone. Referenced to the NAD27 datum.

Conversion to/from this format is slightly inaccurate (5-6 meters).

UTM-NAD83 (GRS80)

Three columns: Easting (meters), northing (meters), and zone. Referenced to the NAD83 datum

(practically identical to WGS84).

Sweden (RT90)

Two columns: Easting (meters) and northing (meters).

The transformations are based on code generously provided by I. Scollar.

279

Open Street Map

Takes two columns of latitudes and longitudes in decimal degrees (WGS84) and shows an Open

Street Map window with graphic elements at the given coordinates in one of these ways:

•

Points with symbols and colours taken from the Past spreadsheet

Filled polygons. The colour is taken from the row colour of the first point. Use an additional

group column to specify multiple polygons.

•

Multi-segment lines, with colour and groups as for filled polygons

Bubbles, with the radius in km of each bubble given in the third data column.

Pie charts, with multiple data columns starting in the third column.

An optional third column can be used to specify the sizes of bubbles (radius in km) or thicknesses of

lines (pixels).

This module requires Internet connection.

Please note: Use of these maps requires that you include the appropriate copyright notice. For

OpenStreetMap, this is:

For the open topomap:

280

Measure on image

A simple tool to digitize point co-ordinates, distances and directions on images. Click “Open” to open

an image. A list of measurements is given on the right. If you start this tool with selected data in the

main spreadsheet, these data will be pre-loaded into the measurement list. The measurement list

can be copy-pasted back to the Past spreadsheet with the “Copy all” button.

New measurements will enter in the selected row in the list. You can move a point by clicking on the

row in the list, then clicking on the new position in the image. You can also delete or insert points in

the list.

281

Stratigraphy menu

Unitary Associations

Unitary Associations analysis (Guex 1991; Guex et al. 2016) is a method for biostratigraphical

correlation (see Angiolini & Bucher 1999 for a typical application). The data input consists of a

presence/absence matrix with samples in rows and taxa in columns. Samples belonging to the same

section (locality) must be assigned to the same group, and ordered stratigraphically within each

section such that the lowermost sample enters in the lowest row.

Overview of the method

The method of Unitary Associations is logical, but rather complicated, consisting of a number of

steps. For details, see Guex (1991). The implementation in PAST includes most of the features found

in the original program, called BioGraph (Savary & Guex 1999), and thanks to a fruitful co-operation

with Jean Guex it also includes a number of additional options and improvements.

The basic idea is to generate a number of assemblage zones (similar to 'Oppel zones') which are

optimal in the sense that they give maximal stratigraphic resolution with a minimum of

superpositional contradictions. One example of such a contradiction would be a section containing a

species A above a species B, while assemblage 1 (containing species A) is placed below assemblage 2

(containing species B). PAST carries out the following steps:

282

1. Residual maximal horizons

The method makes the range-through assumption, meaning that taxa are considered to have been

present at all levels between the first and last appearance in any section. Then, any samples with a

set of taxa that is contained in another sample are discarded. The remaining samples are called

residual maximal horizons. The idea behind this throwing away of data is that the absent taxa in the

discarded samples may simply not have been found even though they originally existed. Absences

are therefore not as informative as presences.

2. Superposition and co-occurrence of taxa

Next, all pairs (A,B) of taxa are inspected for their superpositional relationships: A below B, B below

A, A together with B, or unknown. If A occurs below B in one locality and B below A in another, they

are considered to be co-occurring although they have never actually been found together.

The superpositions and co-occurrences of taxa can be viewed in the biostratigraphic graph. In this

graph, taxa are coded as numbers. Co-occurrences between pairs of taxa are shown as solid blue

lines. Superpositions are shown as arrow lines, pointing from the below-occurring taxon to the

above-occurring taxon.

Some taxa may occur in so-called forbidden sub-graphs, which indicate inconsistencies in their

superpositional relationships. Two of the several types of such sub-graphs can be plotted in PAST: C_n

cycles, which are superpositional cycles (A->B->C->A), and S₃circuits, which are inconsistencies of the

type 'A co-occurring with B, C above A, and C below B'. Interpretations of such forbidden sub-graphs

are suggested by Guex (1991).

283

3. Maximal cliques

Maximal cliques are groups of co-occurring taxa not contained in any larger group of co-occurring

taxa. The maximal cliques are candidates for the status of unitary associations, but will be further

processed below. In PAST, maximal cliques receive a number and are also named after a maximal

horizon in the original data set which is identical to, or contained in (marked with asterisk), the

maximal clique.

4. Superposition of maximal cliques

The superpositional relationships between maximal cliques are decided by inspecting the

superpositional relationships between their constituent taxa, as computed in step 2. Contradictions

(some taxa in clique A occur below some taxa in clique B, and vice versa) are resolved by a 'majority

vote'. The contradictions between cliques can be viewed in PAST.

The superpositions and co-occurrences of cliques can be viewed in the maximal clique graph. In this

graph, cliques are coded as numbers. Co-occurrences between pairs of cliques are shown as solid

blue lines. Superpositions are shown as red arrow lines, pointing from the below-occurring clique to

the above-occurring clique. Also, cycles between maximal cliques (see below) can be viewed as green

lines.

5. Resolving cycles

It will sometimes be the case that maximal cliques are now ordered in cycles: A is below B, which is

below C, which is below A again. This is clearly contradictory. The 'weakest link' (superpositional

relationship supported by fewest taxa) in such cycles is destroyed.

6. Reduction to unique path

At this stage, we should ideally have a single path (chain) of superpositional relationships between

maximal cliques, from bottom to top. This is however often not the case, for example if A and B are

below C, which is below D, or if we have isolated paths without any relationships (A below B and C

below D). To produce a single path, it is necessary to merge cliques according to special rules.

7. Post-processing of maximal cliques

Finally, a number of minor manipulations are carried out to 'polish' the result: Generation of the

'consecutive ones' property, reinsertion of residual virtual co-occurrences and superpositions, and

compaction to remove any generated non-maximal cliques. For details on these procedures, see

Guex (1991). At last, we now have the Unitary Associations, which can be viewed in PAST.

284

The unitary associations have associated with them an index of similarity from one UA to the next,

called D:

D_i=|UA_i-UA_i-1| / |UA_i| +|UA_i-1-UA_i| / |UA_i-1|

8. Correlation using the Unitary Associations

The original samples are now correlated using the unitary associations. A sample may contain taxa

which uniquely places it in a unitary association, or it may lack key taxa which could differentiate

between two or more unitary associations, in which case only a range can be given. These

correlations can be viewed in PAST.

9. Reproducibility matrix

Some unitary associations may be identified in only one or a few sections, in which case one may

consider to merge unitary associations to improve the geographical reproducibility (see below). The

reproducibility matrix should be inspected to identify such unitary associations. A UA which is

uniquely identified in a section is shown as a black square, while ranges of UAs (as given in the

correlation list) are shown in gray.

10. Reproducibility graph and suggested UA merges (biozonation)

The reproducibility graph (Gk' in Guex 1991) shows the superpositions of unitary associations that

are actually observed in the sections. PAST will internally reduce this graph to a unique maximal path

(Guex 1991, section 5.6.3), and in the process of doing so it may merge some UAs. These mergers are

285

shown as red lines in the reproducibility graph. The sequence of single and merged UAs can be

viewed as a suggested biozonation.

Special functionality

The implementation of the Unitary Associations method in PAST includes a number of options and

functions which have not yet been described in the literature. For questions about these, please

contact us.

References

Angiolini, L., Bucher, H. 1999. Taxonomy and quantitative biochronology of Guadalupian brachiopods

from the Khuff Formation, Southeastern Oman. Geobios 32:665-699.

Guex, J. 1991. Biochronological Correlations. Springer Verlag.

Guex, J., Galster, F., Hammer, Ø. 2016. Discrete Biochronological Time Scales. Springer.

Savary, J., Guex, J. 1999. Discrete Biochronological Scales and Unitary Associations: Description of the

BioGraph Computer Program. Meomoires de Geologie (Lausanne) 34.

286

Ranking-Scaling

Ranking-Scaling (Agterberg & Gradstein 1999) is a method for quantitative biostratigraphy based on

events in a number of wells or sections. The data input consists of wells in rows with one well per

row, and events (e.g. FADs and/or LADs) in columns. The values in the matrix are depths of each

event in each well, increasing upwards (you may want to use negative values to achieve this).

Absences are coded as zero. If only the order of events is known, this can be coded as increasing

whole numbers (ranks, with possible ties for co-occurring events) within each well.

The implementation of ranking-scaling in PAST is not comprehensive, and advanced users are

referred to the RASC and CASC programs of Agterberg and Gradstein.

Overview of the method

The method of Ranking-Scaling proceeds in two steps:

1. Ranking

The first step of Ranking-Scaling is to produce a single, comprehensive stratigraphic ordering of

events, even if the data contains contradictions (event A over B in one well, but B over A in another),

or longer cycles (A over B over C over A). This is done by 'majority vote', counting the number of

times each event occurs above, below or together with all others. Technically, this is achieved by

Presorting followed by the Modified Hay Method (Agterberg & Gradstein 1999).

2. Scaling

The biostratigraphic analysis may end with ranking, but additional insight may be gained by

estimating stratigraphic distances between the consecutive events. This is done by counting the

number of observed superpositional relationships (A above or below B) between each pair (A,B) of

consecutive events. A low number of contradictions implies long distance.

Some computed distances may turn out to be negative, indicating that the ordering given by the

ranking step was not optimal. If this happens, the events are re-ordered and the distances re-

computed in order to ensure only positive inter-event distances.

RASC in PAST

Parameters

•

Well threshold: The minimum number of wells in which an event must occur in order to be

included in the analysis

Pair threshold: The minimum number of times a relationship between events A and B must

be observed in order for the pair (A,B) to be included in the ranking step

Scaling threshold: Pair threshold for the scaling step

•

Tolerance: Used in the ranking step (see Agterberg & Gradstein)

Ranking

The ordering of events after the ranking step is given, with the first event at the bottom of the list.

The "Range" column indicates uncertainty in the position.

287

Scaling

The ordering of the events after the scaling step is given, with the first event at the bottom of the list.

For an explanation of all the columns, see Agterberg & Gradstein (1999).

Event distribution

A plot showing the number of events in each well, with the wells ordered according to number of

events.

Scattergrams

For each well, the depth of each event in the well is plotted against the optimum sequence (after

scaling). Ideally, the events should plot in an ascending sequence.

Dendrogram

Plot of the distances between events in the scaled sequence, including a dendrogram which may aid

in zonation.

Variance analysis

For each event, this function plots the deviations from the line of correlation (see above) across all

the wells. This gives a graphical representation of the biostratigraphic quality of each event.

Reference

Agterberg, F.P. & F.M. Gradstein. 1999. The RASC method for Ranking and Scaling of Biostratigraphic

Events. In: Proceedings Conference 75th Birthday C.W. Drooger, Utrecht, November 1997. Earth

Science Review 46(1-4):1-25.

288

Constrained optimization (CONOP)

Requires a table of depths/levels, with wells/sections in rows and event pairs in columns: FADs in odd

columns and LADs in even columns. Missing events coded with zeros.

PAST includes a simple version of Constrained Optimization (Kemple et al. 1989). Both FAD and LAD

of each taxon must be specified in alternate columns. Using so-called Simulated Annealing, the

program searches for a global (composite) sequence of events that implies a minimal total amount of

range extension (penalty) in the individual wells/sections. The parameters for the optimization

procedure include an initial annealing temperature, the number of cooling steps, the cooling ratio

(percentage lower than 100), and the number of trials per step. For explanation and

recommendations, see Kemple et al. (1989).

Output windows include the optimization history with the temperature and penalty as function of

cooling step, the global composite solution and the implied ranges in each individual section.

The implementation of CONOP in PAST is based on a FORTRAN optimization core provided by Sadler

and Kemple.

Reference

Kemple, W.G., P.M. Sadler & D.J. Strauss. 1989. A prototype constrained optimization solution to the

time correlation problem. In Agterberg, F.P. & G.F. Bonham-Carter (eds), Statistical Applications in

the Earth Sciences. Geological Survey of Canada Paper 89-9:417-425.

289

Range confidence intervals

Estimation of confidence intervals for first or last appearances or total stratigraphic range, for one

taxon.

Assuming a random (Poisson) distribution of fossiliferous horizons, and given the first occurrence

datum (level), last occurrence datum, and total number of horizons where the taxon is found, we can

calculate confidence intervals for the stratigraphic range of one taxon (Strauss & Sadler 1989,

Marshall 1990).

No data are needed in the spreadsheet. The program will ask for the number of horizons where the

taxon is found, and levels or dates for the first and last appearances. If necessary, use negative values

to ensure that the last appearance datum has a higher numerical value than the first appearance

datum. 80%, 95% and 99% confidence intervals are calculated for the FAD considered in isolation, the

LAD considered in isolation, and the total range.

The value α is the length of the confidence interval divided by the length of the observed range.

For the single endpoint case:

⁄(

−1 ꢔ−1

⁾− 1

(

)

ꢒ = 1 − ꢉ₁

where C₁is the confidence level and H the number of fossiliferous horizons.

For the joint endpoint (total range) case, α is found by iterative solution of the equation

(

− ꢔ−1

)

(

− ꢔ−1

)

(

)

(

)

ꢉ₂= 1 − 2 1 + ꢒ

+ 1 + 2ꢒ

The assumption of random distribution will of course not hold in many real situations.

Solow’s method

Past also includes a method due to Solow (2003) that does not assume a uniform (stationary)

distribution of finds. It only uses the occurrence datums of the lowest, second lowest, highest and

second highest finds.

References

Marshall, C.R. 1990. Confidence intervals on stratigraphic ranges. Paleobiology 16:1-10.

Solow, A.R. 2003. Estimation of stratigraphic ranges when fossil finds are not randomly distributed.

Paleobiology 29:181-185.

Strauss, D. & P.M. Sadler. 1989. Classical confidence intervals and Bayesian probability estimates for

ends of local taxon ranges. Mathematical Geology 21:411-427.

290

Distribution-free range confidence intervals

Estimation of confidence intervals for first or last appearances. Assumes there is no correlation

between stratigraphic position and gap size. Section should be continuously sampled. Expects one

column per taxon, with levels or dates of all horizons where the taxon is found. This method

(Marshall 1994) does not assume random distribution of fossiliferous horizons. It does require that

the levels or dates of all horizons containing the taxon are given. The program outputs upper and

lower bounds on the lengths of the confidence intervals, using a 95 percent confidence probability,

for confidence levels of 50, 80 and 95 percent. Values which cannot be calculated are marked with an

asterisk (see Marshall 1994).

Reference

Marshall, C.R. 1994. Confidence intervals on stratigraphic ranges: partial relaxation of the assumption

of randomly distributed fossil horizons. Paleobiology 20:459-469.

291

Sequence slotting

Requires two groups (marked with a group column) of univariate or multivariate data, the rows in

each group arranged in sequence. Sequence slotting will merge the two sequences into one, in a way

that minimizes the total path length from point to point, according to the selected distance measure.

A typical application is the correlation of geophysical or geochemical data from two wells or sections.

The algorithm is based on Clark (1985).

An optional third group with up to three columns specifies constraints on the sequence. Each row in

this group specifies one constraint using three numbers: a constraint type code followed by up to

two integer parameters A and B, referring to row numbers within each original sequence.

Code

1

2

3

4

5

6

7

8

A

s

B

t

Function

A_sand B_tare near-neighbors

A_sand B_tare adjacent

A_sprecedes B_t

A_sfollows B_t

Subsequence A_s, B_t, A_s+1

Subsequence B_t, A_s, B_t+1

Sequence starts with A₁

Sequence starts with B₁

Sequence ends with A_m

Sequence ends with B_n

No B between A_sand A_s+1

No A between B_tand B_t+1

9

10

11

12

s

t

Reference

Clark, R.M. 1985. A FORTRAN program for constrained sequence-slotting based on minimum

combined path length. Computers & Geosciences 11: 605-617.

292

Diversity and rates

This module computes curves for diversity (species richness), origination and extinction from a

presence-absence (binary) matrix with the oldest sample or interval in the last (lowest) row, and the

youngest in the first row.

The many options reflect the considerable literature on such curves – for a short overview see

Hammer and Harper (2024, pp. 192-196).

Richness

Species richness is estimated simply by counting the number of presences in each row. The Range-

through option will fill in any absences between the first and last appearance of each species. The

Endpoint adjust option will count any species beginning or ending in the row as 0.5 unit, while

species confined to the row (singletons) will count 1/3 unit. This option should be used if rows

represent time intervals instead of point samples (Foote 2000; Hammer & Harper 2024).

Origination and extinction – method “Simple”

Counts metric: The simple number of first appearances in the row (originations) and the number of

last appearances (extinctions). These numbers are calculated with or without range-through, as

selected by the Range-through option.

Ratios metric: Simple origination/extinction counts divided by the number of presences in the row.

Rates metric: Calculated according to Foote (2000) as follows.

+푁

ꢛꢶ

Origination rate: 푅_ꢧ= ln ^푁

^ꢶwhere N_tis the number of taxa present both in the focus row and

ꢛꢶ

푁

the row above, and N_btis the number of taxa present both in the focus row and in the rows below

and above.

+푁

Extinction rate: 푅_ꢣ= ln ^푁

^ꢛwhere N_bis the number of taxa present both in the focus row and the

ꢛꢶ

푁

ꢛꢶ

row below.

Origination and extinction, method “Three-timer”

The data are not subjected to range-through.

Counts metric: Not available

Ratios metric: Calculated according to Alroy (2008) as follows.

ꢡ

3

Per-interval sampling probability 푠 =

where t₃is number of taxa present both in the focus row

ꢡ +ꢑ

and in the rows below and above (“thr³ee-timers”), and p is the number of taxa present below and

above but not in the focus row (“part-timers”).

293

Origination ratio: 푅푎_ꢧ= 푠 ^ꢡ^푈where t₂^Uis the number of species present in the focus row and the row

2

ꢡ

above (“upper two-timers”)³.

ꢷ

Extinction ratio: 푅푎_ꢣ= 푠 ^ꢡwhere t₂is the number of species present in the focus row and the row

2

L

ꢡ

3

below (“lower two-timers”).

Rates metric: Logarithms of the ratios.

Origination and extinction, method “Gap-filler”

Counts metric: Not available

Ratios metric: Calculated according to Alroy (2014) as

푈

2

ꢡ

+ꢑ

Origination ratio: 푅푎_ꢧ=

where g is the number of species present in the row above and two

ꢡ +ꢑ+ꢅ

rows below (“gap fillers”).³It does not matter if it was found in the focus row.

ꢷ

ꢡ +ꢑ

2

Extinction ratio: 푅푎_ꢣ=

where g is the number of species present in the row below and two

ꢡ +ꢑ+ꢅ

3

rows above.

Rates metric: Logarithms of the ratios.

References

Alroy, J. 2008. Dynamics of origination and extinction in the marine fossil record. Proceedings of the

National Academy of Sciences USA 105:11,536–11,542.

Alroy, J. 2014. Accurate and precise estimates of origination and extinction rates. Paleobiology 40:

374–397.

Foote, M. 2000. Origination and extinction components of taxonomic diversity: general problems.

Paleobiology 26:74–102.

Hammer, Ø. & Harper, D.A.T. 2024. Paleontological Data Analysis, 2^nded. Elsevier.

294

Stratigraphic chart

This flexible plotting module can produce well logs and logs of stratigraphic sections. It provides for

multiple plots, called panels, of different types, making it possible to combine e.g. biozones,

lithology, geochemical and geophysical logs and pollen diagram in one figure. The plotting setup is

automatically saved in your Past file together with the data in the spreadsheet.

Note: This module is under construction; expect bugs and lack of features. Especially, you should not

delete, add or rearrange columns in the spreadsheet while working with a stratigraphic chart, this

will cause unpredictable behaviour.

The spreadsheet should contain one or more columns of stratigraphic levels (typically meter levels),

and several columns of data collected at the given levels. When opening the module you will see a

blank page. Use the “Add panel” button to add plots. The Zoom buttons are useful for navigating in

large charts.

Global settings

Above the plot are found the settings affecting all the panels. The Chart height is the height of the

chart in pixels. The Bottom and Top levels are the vertical limits of the chart. The Levels are depths

option controls the vertical orientation – select this if your levels increase downwards in the section

or core.

Panel settings

These are the settings affecting each panel. The Width is the width of the panel in pixels. Level sets

the column containing the vertical levels for each data point. Data start and Data end set the range

of columns for the data in this panel. Often, each panel will show only one data series, and then Data

start and Data end should be identical.

Type sets the plot type (Line, Dots, Line+dots, Silhouette, Bars, Stacked, Spindle, Zones, Lithology).

The spindle diagram is used for fossil occurrences (abundances or presence-absence).

295

The Zones plot displays a stratigraphic succession of intervals such as biozones, periods or stages (see

the leftmost panel in the example above). The levels represent the basal levels of each interval. An

extra data point is required for the top of the last interval. The data should be a single column of type

‘String’, with the names of the intervals. Chronostratigraphic units such as ‘Precambrian’,

‘Ordovician’, ‘Upper Ordovician’, ‘Pliensbachian’ are recognized by the program and will be plotted

using the colors specified by the International Commission on Stratigraphy.

The Lithology plot displays a column with lithology patterns. Levels are given as for the Zones plot

(above). The data should consist of a column of type ‘String’, containing lithology codes (not case

sensitive). A second data column may be given with a width in percent (0-100).

Sst

Cst

Cgl

Slt

Sandstone

Coarse sandstone

Conglomerate

Siltstone

Sha

Lst

Sls

Ign

Mis

Shale

Limestone

Silty limestone

Igneous rock

Missing/covered (cross)

296

Radiocarbon calibration

This module expects one or more rows of radiocarbon dates, with two columns for uncalibrated age

(BP 1950) and its laboratory-reported standard error (one-sigma). Three calibration curves are

included: IntCal20 (Reimer et al. 2020) for atmospheric (terrestrial) samples, SHcal20 (Hogg et al.

2020) for atmospheric southern hemisphere samples, and MarineCal20 (Heaton et al. 2020) for

marine samples. Please include a reference to Reimer et al. (2020), Hogg et al. (2020) or Heaton et al.

(2020) when reporting calibration results (as well as Past of course!).

For the marine calibration, a globally averaged reservoir age (around 400 years, but varying with

time) is included in the calibration. However, a local reservoir age correction (delta R) and its one-

sigma error should be entered. These values can be looked up e.g. in the Marine Reservoir Correction

Database at http://calib.org/marine/

The output ages can be reported in cal BP 1950 (years before 1950), BP 2000, or CE (Common Era).

The CE values are reported with negative or positive sign and with year zero included (ISO 8601).

Very young and old samples can be somewhat inaccurately calibrated because of the limited range of

the calibration curve. Check the calibration plot to see if the probability curve is truncated.

Plotted probability curves are normalized to a peak value of 1.

Calibrated ages should always be reported with confidence intervals (and ideally the complete

probability curve), but the median and mode (age of peak probability) are also given by Past as

convenient short-hands. The median is more commonly used, but Michczyński (2007) found the

mode to perform better in simulation studies.

297

Confidence intervals

For consistency with other modules in Past, the confidence intervals on calibrated ages are given

with respect to the 95% range, not 95.4% as in some other calibration software. This can give slightly

shorter confidence intervals in Past.

The “Percentile” option calculates a single confidence interval from the 2.5 to the 97.5 percentile

points. The “Level set” option calculates a probability threshold such that the sum of all probabilities

larger than the threshold is 0.95. All the corresponding t values are included in the CI. Because the

probability curve can be multimodal, this can give rise to several disjoint segments in the CI. Past

reports the three of these segments containing the largest areas (reported in %), which is usually

sufficient to reach 95%.

Both these options produce valid 95% confidence intervals. The “Level set” option gives a more

detailed view, and is used in most other calibration software, but the “Percentile” option is more

convenient and gives a sufficiently informative CI for most purposes (personal opinion!). If only a

single segment is sufficient for the level set CI, this will normally be very similar to the percentile CI.

The confidence interval is shown as red lines in the plot.

Computational details

The pointwise calculation of the probability curve p(t) as a function of calibrated time follows

Equation (2) in Bronk Ramsey (2001):

2

( )

²( )

exp �−0.5�ꢂ_ꢊ− ꢂ 푡 � ��휎_ꢊ+ 휎 푡 ��

( )

ꢕ 푡 =

2

�

²( )

휎_ꢊ+ 휎 푡

where the calibration curve is r(t) with standard error σ(t), and the given radiocarbon date is r_m± σ_m.

This equation is evaluated with a time step of 1 year.

298

Values for r(t) and σ(t) are linearly interpolated from the tables, which is reasonably accurate given

the fine spacing of the IntCal20 and MarineCal20 curves.

The given error in delta R is included in σ_m(added in quadrature) before calculation of p(t).

References

Bronk Ramsey, C. 2001. Development of the radiocarbon calibration program. Radiocarbon 43:355-

363

Heaton TJ, Köhler P, Butzin M, Bard E, Reimer RW, Austin WEN, Bronk Ramsey C, Grootes PM,

Hughen KA, Kromer B, Reimer PJ, Adkins J, Burke A, Cook MS, Olsen J, Skinner LC. 2020. Marine20-

the marine radiocarbon age calibration curve (0-55,000 cal BP). Radiocarbon 62. doi:

10.1017/RDC.2020.68.

Hogg AG, Heaton TJ, Hua Q, Palmer JG, Turney CSM, Southon J, Bayliss A, Blackwell PG, Boswijk G,

Bronk Ramsey C, Pearson C, Petchey F, Reimer P, Reimer R, Wacker L. 2020. SHCal20 Southern

Hemisphere calibration, 0-55,000 years cal BP. Radiocarbon 62. doi: 10.1017/RDC.2020.59.

Michczyński, A. 2007. Is it possible to find a good point estimate of a calibrated radiocarbon date?

Radiocarbon 49:393-401.

Reimer P, Austin WEN, Bard E, Bayliss A, Blackwell PG, Bronk Ramsey C, Butzin M, Cheng H, Edwards

RL, Friedrich M, Grootes PM, Guilderson TP, Hajdas I, Heaton TJ, Hogg AG, Hughen KA, Kromer B,

Manning SW, Muscheler R, Palmer JG, Pearson C, van der Plicht J, Reimer RW, Richards DA, Scott EM,

Southon JR, Turney CSM, Wacker L, Adolphi F, Büntgen U, Capano M, Fahrni S, Fogtmann-Schulz A,

Friedrich R, Köhler P, Kudsk S, Miyake F, Olsen J, Reinig F, Sakamoto M, Sookdeo A, Talamo S. 2020.

The IntCal20 Northern Hemisphere radiocarbon age calibration curve (0-55 cal kBP). Radiocarbon 62.

doi: 10.1017/RDC.2020.41.

299

Grain size analysis

Classical plotting and characterization of grain size distributions.

The input data should consist of, first, one column of sieve sizes, in mm, from smallest to largest. The

smallest size can be 0, which is interpreted as phi = 10, or 0.001 mm (1 µm). This can be used for

unanalyzed sediment in the bottom pan. The largest sieve size should be larger than the largest

particle in the sample and will contain no sediment. This is used for defining an upper size limit.

The second column contains sample weights for each size fraction, in any unit, or percent. The last

value is assumed to be 0 and does not need to be entered. Additional columns can contain sample

weights for additional samples. The example below defines two samples, with the sieve sizes in the

first column:

0

4.0

5.2

5.3

7.1

8.0

9.1

4.3

2.0

1.5

0.020 7.1

0.063 9.3

0.125 9.4

0.25

0.5

1

11.3

8.1

4.2

2.1

2

4

Past follows recommendations and equations of Blott and Pye (2001). Grain sizes are reported in the

logarithmic phi scale (Krumbein scale), where φ = 0 corresponds to grain diameter 1 mm, more

positive values are finer, and more negative coarser:

ꢥ

휑 = −log

²1 푚푚

where D is in mm. The denominator makes the φ scale dimensionless.

Gravel is defined as φ < -1 (D > 2 mm); sand is -1 < φ < 4 (0.0625 mm > D > 2 mm); mud is φ > 4 (D <

0.0625 mm).

The sediment name (textural group) is defined on the proportions of gravel, sand and mud, following

Folk (1954) and with linear interpolation when sieve sizes do not correspond to the definition

boundaries. The refined classification that also includes proportions of silt and clay within the mud

class is not included in Past.

Mean grain size is computed both with the method-of-moments (MOM) and the Folk & Ward (1957)

“graphical” method. MOM is potentially more accurate when the sediment analysis is complete and

finely binned, but the F&W method is more robust and therefore recommended for less complete

analyses. Results are given both in φ units and in mm.

Mean grain size in φ units, method of moments:

∑

푓푚_ꢸ

푥̅_ꢸ

=

∑

푓

where f are the given sieve sample weights and m_φare the midpoints of the sieve size classes.

300

Mean grain size in φ units, Folk & Ward:

푀_ꢹ=

휑₁₆+ 휑₅₀+ 휑₈₄

3

where the φ are the 16, 50 and 84 percentile points of the cumulative grain size distribution.

Sorting is defined using standard deviations, as follows.

Sorting using φ units, method of moments:

2

∑

푓�푚_ꢸ− 푥_ꢸ

̅

�

휎_ꢸ=

∑

푓

Sorting using φ units, Folk & Ward:

휑

₈₄− 휑₁₆

휑

₉₅− 휑₅

휎_ꢺ=

+

4

6.6

Standardized names of degree of sorting follow Folk & Ward (1957), from “very well sorted” (σ <

0.35) to “extremely poorly sorted” (σ > 4.00)

Skewness is a somewhat less commonly used parameter. Because skewness is very sensitive to

incomplete sieve analysis, only the more robust Folk & Ward method is calculated:

휑₁₆+ 휑₈₄− 2휑₅₀

휑₅+ 휑₉₅− 2휑₅₀

ꢋꢳ_ꢺ=

+

(

)

(

)

2 휑₈₄− 휑₁₆

2 휑₉₅− 휑₅

Standardized names of skewness follow Folk & Ward (1957), from “very coarse skewed” (Sk_I< -0.3)

to “very fine skewed” (Sk_I> 0.3).

References

Blott, S.J., Pye, K. 2001. GRADISTAT: A grain size distribution and statistics package for the analysis of

unconsolidated sediments. Earth Surface Processes and Landforms 26:1237–1248.

Folk, R.L. 1954. The distinction between grain size and mineral composition in sedimentary-rock

nomenclature. Journal of Geology, 62:344–359.

Folk, R.L., Ward, W.C. 1957. Brazos River bar: a study in the significance of grain size parameters.

Journal of Sedimentary Petrology 27:3–26.

301

Scripting

Past includes a fairly rich scripting language, which allows you to program your own modules taking

advantage of the Past user interface, graphics package and mathematical and statistical routines. The

scripter is already fully functional, but it will be substantially extended in future versions.

Language structure

Scripts are written in Pascal-style syntax. The fundamental language elements are:

begin .. end constructor

procedure and function declarations

if .. then .. else constructor

for .. to .. do .. step constructor

while .. do constructor

repeat .. until constructor

try .. except and try .. finally blocks

case statements

array constructors (x:=[ 1, 2, 3 ];)

^ , * , / , and , + , - , or , <> , >=, <= , = , > , < , div , mod , xor , shl , shr operators

access to object properties and methods (ObjectName.SubObject.Property)

Script structure

A script is made of two major blocks: a) procedure and function declarations and b) main block. Both

are optional, but at least one should be present. There is no need for the main block to be inside

begin..end. It could be a single statement. Some examples:

SCRIPT 1:

procedure DoSomething;

begin

CallSomethingElse;

end;

begin

DoSomething;

end;

SCRIPT 2:

begin

CallSomethingElse;

end;

302

SCRIPT 3:

function MyFunction;

begin

result:=’Ok!’;

end;

SCRIPT 4:

CallSomethingElse;

Statements should be terminated by the “;” character. Begin..end blocks are used to group

statements.

Identifiers

Identifier names in script (variable names, function and procedure names, etc.) should begin with a

character (a..z or A..Z), or ‘_’, and can be followed by alphanumeric chars or the ‘_’ char. They cannot

contain any other characters or spaces.

Valid identifiers:

VarName

_Some

V1A2

_____Some____

Invalid identifiers:

2Var

My Name

Some-more

This,is,not,valid

Assign statements

Assign statements (assigning a value or expression result to a variable or object property) are built

using “:=”. Examples:

MyVar := 2;

Button.Caption := ‘This ‘ + ‘is ok.’;

303

Character strings

Strings (sequence of characters) are declared using single quote (‘) characters. Double quotes (“) are

not used. You can also use #nn to declare a character inside a string. There is no need to use the ‘+’

operator to add a character to a string. Some examples:

A := ‘This is a text’;

Str := ‘Text ‘+’concat’;

B := ‘String with CR and LF char at the end’#13#10;

C := ‘String with ‘#33#34’ characters in the middle’;

Comments

Comments are defined by // chars or (* *) or { } blocks. With the // char, the comment will finish at

the end of line.

//This is a comment before ShowMessage

ShowMessage(‘Ok’);

(* This is another comment *)

ShowMessage(‘More ok!’);

{ And this is a comment

with two lines }

ShowMessage(‘End of okays’);

Variables

There is no need to declare variable types. Thus, you declare a variable just using the var directive

and its name. Also, it is optional to declare variables at all. Variables and their types are implicitly

declared at first usage. Examples:

SCRIPT 1:

procedure Msg;

var S;

begin

S:=’Hello world!’;

ShowMessage(S);

end;

SCRIPT 2:

var A;

begin

304

A:=0;

A:=A+1;

end;

SCRIPT 3:

var S: string;

begin

S:=’Hello World!’;

ShowMessage(S);

end;

Var declarations are not strictly necessary in any of scripts above.

Indexing

Strings, arrays and array properties can be indexed using “[” and “]” chars. For example, if Str is a

string variable, the expression Str[3] returns the third character in the string denoted by Str, while

Str[I + 1] returns the character immediately after the one indexed by I. More examples:

MyChar:=MyStr[2];

MyStr[1]:=’A’;

MyArray[1,2]:=1530;

Lines.Strings[2]:=’Some text’;

Arrays

To construct an array, use “[” and “]” chars. You can construct multi-index arrays, nesting array

constructors. You can then access arrays using indexes. A variable is an array if it was assigned using

an array constructor or if it was created using the array or vector procedures. Some examples:

NewArray := [ 2,4,6,8 ];

Num:=NewArray[1]; //Num receives “4”

MultiArray := [[’green’,’red’,’blue’] , [’apple’,’orange’,’lemon’]];

Str:=MultiArray[0,2]; //Str receives ‘blue’

MultiArray[1,1]:=’new orange’;

V:=vector(100);

A:=array(100,100);

Arrays defined using the array constructors can contain elements of any type, but arrays defined by

the vector and array procedures are of type Double.

305

Arrays constructed using the array constructors are indexed from 0. Also, arrays defined using the

vector and array procedures are indexed from 0, but the 0 element is often not used and these

arrays contain n+1 elements, indexed from 0 to n.

If statements

There are two forms of if statement: if...then and if...then...else. If the if expression is true, the

statement (or block) is executed. If there is an else part and the expression is false, the statement (or

block) after else is executed. Examples:

if J <> 0 then Result := I/J;

if J = 0 then Exit else Result := I/J;

if J <> 0 then

begin

Result := I/J;

Count := Count + 1;

end else

Done := True;

while statements

A while statement is used to repeat a statement or a block, while a control condition (expression) is

evaluated as true. The control condition is evaluated before the statement. Hence, if the control

condition is false at first iteration, the statement sequence is never executed. The while statement

executes its constituent statement (or block) repeatedly, testing the expression before each

iteration. As long as expression returns True, execution continues. Examples:

while Data[I] <> X do I := I + 1;

while I > 0 do

begin

if Odd(I) then Z := Z * X;

I := I div 2;

X := Sqr(X);

end;

while not Eof(InputFile) do

begin

Readln(InputFile, Line);

Process(Line);

end;

306

repeat statements

The syntax of a repeat statement is

repeat statement_1; ...; statement_n; until expression

where expression returns a Boolean value. The repeat statement executes its sequence of

constituent statements continually, testing the expression after each iteration. When expression

returns True, the repeat statement terminates. The sequence is always executed at least once

because expression is not evaluated until after the first iteration. Examples:

repeat

K := I mod J;

I := J;

J := K;

until J = 0;

repeat

Write(‘Enter a value (0..9): ‘);

Readln(I);

until (I >= 0) and (I <= 9);

for statements

For statements have the following syntax:

for counter := initialValue to finalValue do statement

The For statement sets counter to initialValue, repeats execution of the statement (or block) and

increments the value of counter until counter reaches finalValue. Examples:

SCRIPT 1:

for c:=1 to 10 do

a:=a+c;

SCRIPT 2:

for i:=a to b do

begin

j:=i^2;

sum:=sum+j;

end;

case statements

Case statements have the following syntax:

307

case selectorExpression of

caseexpr1: statement1;

...

caseexprn: statementn;

else

elsestatement;

end

If selectorExpression matches the result of one of caseexprn expressions, the respective statement

(or block) will be executed. Otherwise, elsestatement will be executed.The Else part of the case

statement is optional. A Case statement doesn’t need to use only ordinal values. You can use

expressions of any type in both the selector expression and the case expression. Example:

case uppercase(Fruit) of

‘lime’: ShowMessage(‘green’);

‘orange’: ShowMessage(‘orange’);

‘apple’: ShowMessage(‘red’);

else

ShowMessage(‘black’);

end;

function and procedure declaration

Declaration of functions and procedures are similar to Pascal, with the difference that you don’t

specify variable types. To return function values, use the implicitly declared result variable.

Parameters by reference can also be used, with the restriction mentioned: no need to specify

variable types. Some examples:

procedure HelloWord;

begin

ShowMessage(‘Hello world!’);

end;

procedure UpcaseMessage(Msg);

begin

ShowMessage(Uppercase(Msg));

end;

function TodayAsString;

begin

result:=DateToStr(Date);

end;

function Max(A,B);

begin

if A>B then

result:=A

else

result:=B;

end;

308

procedure SwapValues(var A, B);

Var Temp;

begin

Temp:=A;

A:=B;

B:=Temp;

end;

The output window

When you run a script, an output window will be opened automatically. It contains three tabs: Text,

graphic and table.

The text tab

The text tab contains a window to which the script can write output. The text can be copy-pasted to

other programs by the user. The following procedures are available:

cleartext

Clears the text window

textout(s)

Writes a line to the text window. Handles numerical, string, vector and array types.

The graphic tab

A resizable graphic canvas with the usual Past functionality such as a graph preferences window with

export to vector (SVG or PDF) or bitmap formats. The window will automatically scale to its contents,

so you do not need to consider the scale of coordinates. For efficiency, no graphics will appear until

the redraw procedure is called.

Colors must be given as one of the following constants: black, red, blue, green, purple, yellow, gray,

brown.

redraw

Redraw the graphic window with automatic axis ranges

Redraw the graphic with the given axis ranges

Clears the graphic window

setaxes(x1, x2, y1, y2)

cleargraphic

savegraphic(filename)

Depending on the file extension, will save the graphic in one

of the following formats: svg, pdf, jpg, tif, gif, png, bmp

drawpoints(x, y, color)

Draws one point (if x and y are single numbers) or several (if x

and y are vectors). Color is a single integer (see above).

drawsymbols(x, y, color, symbol)

Draws one symbol (if x and y are single numbers) or several

(if x and y are vectors). Color and symbol are single integers,

see ‘spreadsheet_symbols’ for symbol coding.

309

drawline(x1, y1, x2, y2, color)

drawpolyline(x, y, color)

A line from (x1, y1) to (x2, y2)

A polyline with lists of x and y coordinates in vectors x and y

A rectangle with the given corners

drawrectangle(x1, y1, x2, y2, color)

drawellipse(x, y, major, minor, angle, color)

An ellipse with center (x, y), given major and minor axes and

with the major axis at the given angle (radians) to the x axis.

drawtext(x, y, string)

Draws text at position (x, y) – may reposition to reduce

overlap

drawmatrix(A, interpolate)

drawhistogram(V, nbins, color, kde)

Draws the matrix A. Set interpolate to false or true to select

drawing mode. Does an automatic redraw.

A histogram of vector V, with the given number of bins. If

kde=true, a kernel density estimate is also drawn.

drawbars(V, color)

A bar chart of vector V.

drawboxplot(V, x, outliers)

A box plot of vector V at the given x position. Outliers is true

or false.

drawconvexhull(Vx, Vy, color)

drawrose(V, n, equalarea, kde)

The convex hull of the points in vectors Vx and Vy.

A rose plot of angles in V (degrees), with n bins. Equalarea

and kde (kernel density estimate) are true or false.

The table tab

A table (spreadsheet) window with copy-paste function.

tablesize(rows, columns: integer)

tableout(row, col, value)

Set the number of rows and columns in the table

Write value to a particular cell in the table, indexing starting

at 0.

310

Accessing the main Past spreadsheet and menus

clickmenu(name: string)

Executes an item in the Past menu. Specify the name as given in the

menu, e.g. clickmenu(‘save as’). Parts of the name in brackets are not

included: Use ‘species packing’, not ‘species packing (Gaussian)’.

spreadsheet_array

Returns an array containing the selected area in the Past

spreadsheet. Group columns are not included

spreadsheet_column(n: integer)

spreadsheet_groups(n: integer)

Returns a vector with the numbers in column n in the Past

spreadsheet.

Returns a vector with group numbers corresponding to the

rows in spreadsheet. For n=1, the first group column is

returned, for n=2 the second group column (if any), etc.

spreadsheet_rowlabels

spreadsheet_columnlabels

spreadsheet_symbols

Returns a string vector with the row labels in the selected area.

Returns a string vector with the column labels in the selected area.

Returns a vector with numbers (0-15) identifying the symbols

corresponding to the rows in spreadsheetarray. 0=dot, 1=+,

2=square, 3=X, 4=triangle, 5=O, 6=diamond, 7=-, 8=l, 9=fillsquare,

10=*, 11=oval, 12=filltriangle 13=invtriangle, 14=fillinvtriangle,

15=filldiamond

spreadsheet_set(row, col, s)

Sets the contents of the cell at (row, col) in the Past spreadsheet to s

(number or string). Indexing starts from 0 (label cells).

311

Array and vector operations

array(m, n: integer)

Allocates and returns a Float (Double) array with m rows and n columns,

indexing starting at 1.

Allocates and returns a vector (one-dimensional array) with n elements,

indexing starting at 1

vector(n: integer)

column(A, n)

row(A, m)

ncols(A)

nrows(A)

inv(A)

Returns column n in array A, as a vector

Return row n in array A, as a vector

Returns the number of columns in array A

Returns the number of rows in array A

Inverse of square matrix A

arrmult(A, B)

mean(V)

variance(V)

median(V)

skew(V)

Returns array multiplication A*B

Mean of vector V

Variance of vector V

Median of vector V

Skew of vector V

kurtosis(V)

svd(A)

Kurtosis of vector V

Singular Value Decomposition of A, returning V augmented by an extra column

containing D.

cov(A)

eig(A)

Returns variance-covariance matrix of A.

Returns the eigenvectors of A, augmented by an extra column with the eigenvalues.

pearsonr(X, Y) Returns the Pearson correlation (r) between vectors X and Y

spearmanrs(X, Y) Returns the Spearman rank-order correlation (rs) between vectors X and Y

linfit(X, Y) Ordinary least-squares regression of vectors X, Y. Returns a 4-vector with slope,

intercept, standard error of slope, standard error of intercept.

Some common mathematical functions are available for arrays and vectors, for efficiency. The

function is applied to each element of the array or vector, returning an array or vector.

arrAbs(A)

arrCos(A)

arrExp(A)

arrLn(A)

arrSin(A)

arrSqrt(A)

arrTan(A)

Absolute value

Cosine (radians)

e^x

Natural logarithm (base e)

Sine (radians)

Square root

Tangent (radians)

Scalar math functions

abs(x)

Absolute value

arctan(x)

arctan2(y, x)

cos(x)

Inverse tangent (radians)

Inverse tangent of y/x extended to correct quadrant.

Cosine (radians)

cumulnorm(x, s)

exp(x)

Cumulative normal distribution function, mean=0, stdev=s

e^x

frac(x)

Fractional part of x

fresnel(x)

ln(x)

Returns a 2-vector with the S and C Fresnel integrals to x.

Natural logarithm (base e)

normal(m, s)

gamma(k)

invnorm(x)

Normally distributed random number, mean=m, stdev=s.

Gamma distributed random number, shape=k, scale=1.

Inverse of the cumulative normal distribution, mean=0, stdev=1, 0<x<1

312

invchi2(x, df)

odd(x)

Inverse of the cumulative chi-squared distribution with df degrees of freedom

True if x is odd

random

round(x)

sin(x)

Random number, uniform distribution, 0 <= x < 1

Rounds to the nearest integer

Sine (radians)

sqr(x)

sqrt(x)

Square (x*x)

Square root

studentp(t, df) Two-tailed p value from Student’s t distribution with df degrees of freedom

tan(x)

trunc(x)

Tangent (radians)

Rounds down

File I/O

function Append(var F: File): Integer;

Prepares an existing file for adding text to its end. F is a text file variable and must be

associated with an external file, using AssignFile. If the external file does not exist, an error

occurs. If F is already open, it is closed, then reopened. The current file position is set at the

end of the file.

function AssignFile(var F: File; FileName: String): Integer;

Associates the name of an external file with a file variable. After calling AssignFile, F is

associated with the external file until F is closed. All further operations on the file variable F

operate on the external file named by FileName.

procedure ChDir(S: string);

Changes the current directory to the path specified by S.

procedure CloseFile(var F: File);

Terminates the association between a file variable and an external disk file. F is a file variable

opened using Reset, Rewrite, or Append. The external file associated with F is completely

updated and then closed, freeing the file handle for reuse.

function Eof(var F: File): Boolean;

Tests whether the file position is at the end of a file.

function FilePos(var F: File): Integer;

Use on an open file to determine the current position. If the current position is at the

beginning, FilePos returns 0. Otherwise, FilePos returns the byte offset from the beginning of

the file.

function FileSize(var F: File): Integer;

313

Returns the number of records in a file.

function ReadLn(var F: File): String;

Reads a line of text and then skips to the next line of the file.

procedure Reset(var F: File);

Opens the existing external file with the name assigned to F. An error results if no existing

external file of the given name exists or if the file cannot be opened. If F is already open, it is

first closed and then reopened. The current file position is set to the beginning of the file.

procedure Rewrite(var F: File);

Creates a new external file with the name assigned to F. F is associated with an external file

using AssignFile. If a file with the same name already exists, it is deleted and a new empty file

is created in its place. If F is already open, it is first closed and then re-created. The current

file position is set to the beginning of the empty file.

procedure WriteLn(var F: File; S: string);

Writes to a text file and adds an end-of-line marker.

String operations

function Chr(X: Byte): Char;

Returns the character for a specified ASCII value.

function CompareStr(S1, S2: string): Integer;

Compares S1 to S2, with case sensitivity. The return value is less than 0 if S1 is less than S2, 0

if S1 equals S2, or greater than 0 if S1 is greater than S2.

function CompareText(S1, S2: string): Integer;

Compares S1 to S2, without case sensitivity. The return value is less than 0 if S1 is less than

S2, 0 if S1 equals S2, or greater than 0 if S1 is greater than S2.

function Copy(S: string; Index: Integer; Count: Integer): string;

Returns a substring of a string S. Index and Count are integer-type expressions. Copy returns

a substring or subarray containing Count characters or elements starting at S[Index].

function FloatToStr(Value: Double): string;

Converts the floating-point value given by Value to its string representation. The conversion

uses general number format with 15 significant digits.

314

procedure Insert(Source: string; var Dest: string; Index: Integer);

Inserts a substring into a string, from a specified position. If Index is less than 1, it is set to 1.

If it is past the end of Dest, it is set to the length of Dest, turning the operation into an

append.

function IntToStr(Value: Integer): string;

Converts an integer to a string that contains its decimal representation.

function Length(S: string): Integer;

Returns the number of characters in a string.

function LowerCase(S: string): string;

Returns a string with the same text as in S, but with all letters converted to lowercase.

function Pos(SubStr, Str: string): Integer;

Returns an index of the first occurrence of Substr in Str. Returns zero if Substr is not found.

function StrToFloat(S: string): Double;

Converts a string to a floating-point value (leading and trailing blanks are ignored).

function StrToInt(S: string): Integer;

Converts a string that represents an integer into a number.

function StrToIntDef(S: string; Default: Integer): Integer;

Converts the string S, which represents an integer, into a number. If S does not represent a

valid number, StrToIntDef returns Default.

function Trim(S: string): string;

Trims leading and trailing spaces and control characters from a string.

function TrimRight(S: string): string;

Trims trailing spaces and control characters from a string.

function UpperCase(S: string): string;

Returns a copy of a string in uppercase.

315

Other functions

procedure ShowMessage(S: string);

Shows a message box and waits for user to click Ok.

function InputQuery(Caption, Prompt: string; var Value: string):

Boolean;

Displays an input dialog box that lets the user enter a value. Caption is the caption of the

dialog box. Prompt is the text that prompts the user to enter input. Value is the value that

appears in the edit box when the dialog first appears and returns the value that the user

enters. InputQuery returns true if the user chooses OK, false if the user chooses Cancel or

presses Esc.

procedure sleep(ms: integer);

Suspends the execution of the script for the given number of milliseconds.

316

Calling dll functions (Windows only)

Past allows importing and calling external DLL functions, by declaration of script routines, indicating

library name and, optionally, the calling convention, in addition to the function signature. External

libraries are loaded by Past on demand, before function calls, if not already loaded (dynamically or

statically). To load and unload libraries explicitly, functions LoadLibrary and FreeLibrary from unit

Windows can be used.

Syntax

function functionName(arguments): resultType; [callingConvention];

external ‘libName.dll’ [name ExternalFunctionName];

For example, the following declaration:

function MyFunction(arg: integer): integer;

external ‘CustomLib.dll’;

imports a function called MyFunctionfrom CustomLib.dll. Default calling convention, if not

specified, is register. Past also allows declaring a different calling convention (stdcall, register, pascal,

cdecl or safecall) and to use a different name for the DLL function, like the following declaration:

function MessageBox(hwnd: pointer; text, caption: string; msgtype:

integer): integer; stdcall;

external ‘User32.dll’ name ‘MessageBoxA’;

that imports ‘MessageBoxA’ function from User32.dll (Windows API library), named ‘MessageBox’ to

be used in the script.

The declaration above can be used for functions and procedures (routines without result value).

Supported types

Past supports the following basic data types on arguments and result of external functions:

Integer

Boolean

Char

Extended

String

Pointer

PChar

Object

Class

WideChar

PWideChar

AnsiString

Currency

Variant

317

Interface

WideString

Longint

Cardinal

Longword

Single

Byte

Shortint

Word

Smallint

Double

Real

DateTime

TObject descendants (class must be registered in scripter with DefineClass)

Other types (records, arrays, etc.) are not supported yet. Arguments of above types can be passed by

reference, by adding var in the param declaration of the function.

Libraries and classes

318

Forms and components

For user input more complex than provided by the InputQuery function (see above), you can build

your own forms (windows) with the following components: Labels, edits (where the user can enter

text and numbers), buttons and check boxes.

Form

A form is defined with a variable of type TForm, and created using “TForm.CreateNew(nil, 0)”. Some

useful properties of the TForm class are:

Caption

Width

Height

The text at the top of the form

Width in pixels

Height in pixels

The method showModal displays the form, returning mrOk or mrCancel depending on buttons

clicked in the form (see Button below).

Label

A label, of type TLabel, created with Tlabel.create(parentform) shows a simple text. Properties of the

TLabel class include

Parent

Text

The parent form, must be specified here in addition to in TLabel.create

The text of the caption

Position.x

Position.y

Width

x position, in pixels, relative to the parent form

y position, in pixels

Width in pixels

Edit

A box, of type TEdit, created with TEdit.create(parentform), where the user can enter text or

numbers. Properties of the TEdit class include

Parent

Text

The parent form, must be specified here in addition to in TEdit.create

The text of the caption

Position.x

Position.y

Width

x position, in pixels, relative to the parent form

y position, in pixels

Width in pixels

Button

A button, of type TButton, created with TButton.create(parentform). Properties of the TButton class

include

Parent

Text

The parent form, must be specified here in addition to in TButton.create

The text of the caption

Position.x

Position.y

Width

x position, in pixels, relative to the parent form

y position, in pixels

Width in pixels

319

Height

modalResult

Height in pixels

Can be set to mrOk or mrCancel. When the button is clicked, this value is returned by

the form’s showModal method.

Check box

A check box, of type TCheckBox, created with TCheckBox.create(parentform). Properties of the

TCheckBox class include

Parent

Text

The parent form, must be specified here in addition to in TCheckBox.create

The text of the caption

Position.x

Position.y

Width

x position, in pixels, relative to the parent form

y position, in pixels

Width in pixels

isChecked

Boolean (True or False). Read only, don’t set to True from code, it won’t work

Example

The following script shows a form with a label, an edit control and an OK button. When the user

clicks the button, the text of the edit control is written to the text window.

var

fm: TForm;

lb: TLabel;

ed: TEdit;

bt: TButton;

begin

fm : = TForm.CreateNew(nil, 0);

fm.Caption := ‘A brand new form!’;

fm.Width := 300;

fm.Height := 150;

lb := TLabel.Create(fm);

lb.Parent := fm;

lb.Position.X := 10;

lb.Position.Y := 10;

lb.Text := ‘Your name:’;

ed := TEdit.Create(fm);

ed.Parent := fm;

ed.Position.X := lb.Position.X;

ed.Position.Y := lb.Position.Y + lb.Height + 10;

bt := TButton.Create(fm);

bt.Parent := fm;

bt.Position.X := ed.Position.X;

bt.Position.Y := ed.Position.Y + ed.Height + 10;

bt.Text := ‘Ok’;

bt.Default := True;

bt.ModalResult := mrOk;

fm.ActiveControl := ed; // Sets the focus to the edit control

320

if fm.ShowModal = mrOk then

textout(‘Hello ‘+ed.Text+’!’);

fm.Free;

end;

321