NAME

statreport - Summarize numeric samples from MCMC logs.

SYNOPSIS

statreport [OPTIONS] file1 [file2 ...]

DESCRIPTION

statreport computes summary statistics and sampling diagnostics for numeric fields in MCMC logs. It accepts tab-delimited files with a header row of column names, or BAli-Phy JSON logs. Use - as a filename to read standard input. Results are written to the terminal.

Supply multiple files to summarize independent runs of the same analysis. The selected columns must have matching names and order. Retained samples are pooled, so longer retained runs contribute more samples to the combined summaries.

By default, the report gives the median and a 95% highest-posterior-density credible interval for each parameter, together with applicable autocorrelation, effective sample size, and convergence diagnostics. The initial 10% of the shortest input is discarded as burn-in.

EXAMPLES

Summarize a single run:

statreport run-1/C1.log

Combine independent runs:

statreport run-1/C1.log run-2/C1.log

Discard the first 1000 data rows from each file:

statreport --skip=1000 run-1/C1.log

Show both means and medians:

statreport --mean --median run-1/C1.log

Analyze only columns 2 through 4:

statreport --select=2:4 run-1/C1.log

Row counts exclude the header. They correspond to MCMC iterations only when one row was logged per iteration.

OPTIONS

Sample selection

-s COUNT|PERCENT, --skip COUNT|PERCENT

Discard initial samples as burn-in. Accepts a data-row count or a percentage such as 10% (the default). A percentage is calculated from the shortest input after endpoint limiting and subsampling; the same number of retained rows is then removed from each file. With subsampling, an integer count is divided by the subsampling factor and rounded down before rows are removed.

-u NUM, --until NUM

Use only the first NUM data rows of each file, before subsampling and burn-in removal. This is a row limit, not a cutoff on the value of an iteration field. By default, read to the end.

-x NUM, --subsample NUM

Read every NUMth data row, starting with the first. The default is 1 (no thinning). Burn-in removal follows this selection.

--min NUM

Require at least NUM rows in each file after endpoint limiting and subsampling, but before burn-in removal. By default, no additional minimum is imposed.

Field selection

-S FIELD|RANGE, --select FIELD|RANGE

Analyze only the specified field name or inclusive column range, such as 2:4. Column numbers start at 1. Repeat the option to select multiple fields or ranges. Without this option, analyze all fields not excluded by --ignore.

-I FIELD|RANGE, --ignore FIELD|RANGE

Exclude a field name or column range. Repeat to exclude multiple fields or ranges. Exclusion takes precedence over selection. Fields chosen for analysis must be numeric.

Summaries and formatting

--mean

Show the mean and standard deviation. Suppresses the default median summary unless --median is also supplied.

--median

Show the median and credible interval. This is the default unless --mean is supplied.

--mode

Add an estimated mode and a measure of its resolution. See INTERPRETING THE OUTPUT.

--confidence PROBABILITIES

Credible-interval probabilities, separated by colons, such as 0.8:0.95. The default is 0.95. These control the intervals accompanying median summaries.

--BCI TYPE

Credible-interval type: HPD (the default) or central. HPD uses a shortest interval containing the requested posterior probability; central intervals leave equal probability in each tail.

-i, --individual

Include per-file summaries and diagnostics in addition to pooled results. The mode estimate and comparisons with --truth remain pooled in ordinary, unconditioned reports.

-p NUM, --precision NUM

Number of significant figures in the output. The default is 4.

Specialized analysis

--condition FIELD=VALUE

Summarize only retained samples whose numeric FIELD equals VALUE. VALUE must be finite. Only one condition is supported. See SPECIALIZED SUMMARIES for matching counts and diagnostic limitations.

--truth VALUE

Compare samples with a known numeric value, reporting bias, the fraction above VALUE minus one half (high), mean absolute error (absE), and root mean square error (rmsE). The same reference value is used for every analyzed field.

General options

-h, --help

Print usage information.

-V, --verbose

Print additional diagnostic messages to standard error.

INTERPRETING THE OUTPUT

A line such as parameter ~ 1.2 (0.8,1.7) @ 95% gives the median and credible interval. An interval is printed as (NA,NA) when the sample count is too small for the requested probability: the count multiplied by one minus that probability must be at least 10.

With --mean, E parameter = ... [+- ...] gives the mean and standard deviation, not the standard error of the mean. With --mode, parameter ^ ... [+- ...] gives the midpoint and full width of a narrow sample interval used to estimate the mode. This width is a resolution measure, not a credible interval or standard error. The mode is NA with fewer than 20 samples.

Constant fields are reported as single values. In ordinary reports, fields that only increase or only decrease are marked accordingly instead of receiving the usual summaries.

Sampling and convergence diagnostics

t @

Estimated autocorrelation time, in retained-sample units. Larger values indicate more dependence between successive samples.

Ne

Effective sample size: the retained sample count divided by the estimated autocorrelation time. This is an estimate of sampling information, not a count of independent observations.

burnin

A heuristic estimate based on repeated crossings of the range between the 5th and 95th percentiles of the final third of each input. The pooled report gives the largest estimate among runs. Not Converged! means a run did not complete those crossings before its final third. The estimate uses data before burn-in removal and is reported in original data-row units, even with subsampling. It does not change which samples --skip removes.

With multiple files, two additional measures compare chains using central 80% intervals, regardless of the chosen credible-interval options:

PSRF-80%CI

Width of the pooled interval divided by the average within-chain interval width. For integer-valued fields, widths count integer values, including both endpoints.

PSRF-RCF

For each chain's interval, the fraction of that chain's samples inside it divided by the fraction of pooled samples inside it, averaged over chains.

Values near 1 indicate agreement by these measures; larger values indicate differences between chains. These are interval-based diagnostics, not modern R-hat. Neither they nor the burn-in estimate establish convergence on their own.

SPECIALIZED SUMMARIES

Conditional summaries

--condition filters samples after endpoint limiting, subsampling, and burn-in removal. The condition field is read even when it is not selected for output. Matching sample counts are reported for each chain and pooled. No matching samples is reported explicitly and is not an error.

Conditional reports omit autocorrelation times, effective sample sizes, burn-in estimates, and PSRF measures: filtering changes the time sequence to which those diagnostics apply. Use the ordinary report to examine diagnostics for unfiltered parameter samples.

Log-odds summaries

Fields whose final name component begins with LogOdds, ignoring case, also receive a posterior-probability summary. Components are separated by / or :, so names such as P1/M8a_test:LogOddsPosSelection are recognized.

The summary converts each sampled natural-log odds to a probability, averages those probabilities, and reports both the average probability and its log odds. This differs from converting the average log odds to a probability. Ordinary summaries of the sampled log odds are retained.

REPORTING BUGS

BAli-Phy online help: https://www.bali-phy.org/docs.php.

Please send bug reports to bali-phy-users@googlegroups.com.