Pg bayesian

Skill Exekiel179/pingouin-psych-stats/skills/pg-bayesian

Skill plugin for Claude Code, Codex & other AI agents: small, safety-first Pingouin workflows for psychology statistics — assumption checks, S0–S5 supervision gates, reproducible Python, and APA-style reporting.

Install

npx -y skills add Exekiel179/pingouin-psych-stats --skill pg-bayesian

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Compute or generate Pingouin Bayes factors for t tests, correlations, and proportions (bayesfactor_ttest, bayesfactor_pearson, bayesfactor_binom), and read the BF10 already returned by pg.ttest and pg.corr.

SKILL.md

2.2 KB, 601 tokens by cl100k_base, as published. Nobody here has run it

PG Bayesian

Use when the user wants Bayesian evidence (Bayes factors) alongside or instead of p-values, e.g. to quantify support for the null.

Load

Read:

../../references/supervision-gates.md
../../references/pingouin-api-quickref.md
../../references/apa-output-template.md if writing results.

Decision Rules

t test already run -> read the BF10 column from pg.ttest(...); no extra call needed.
Correlation already run -> read BF10 from pg.corr(...).
From a t statistic -> pg.bayesfactor_ttest(t, nx, ny=None, paired=False).
From a correlation r and n -> pg.bayesfactor_pearson(r, n).
Proportion vs a chance value -> pg.bayesfactor_binom(k, n, p).

Required Inputs

For t: the t value, group sizes (nx, ny), paired flag, prior scale r (default 0.707).
For pearson: r and n.
For binom: successes k, trials n, null probability p (default 0.5).

Code Patterns

Bayes factor from a t test (BF10 is also already in the ttest table):

tt = pg.ttest(x, y, paired=False)
print({"BF10_from_table": float(tt["BF10"].iloc[0])})
bf = pg.bayesfactor_ttest(float(tt["T"].iloc[0]), nx=len(x), ny=len(y))
print({"BF10": round(float(bf), 3)})

Bayes factor for a correlation:

bf = pg.bayesfactor_pearson(r=0.30, n=60)
print({"BF10": round(float(bf), 3)})

Bayes factor for a proportion:

bf = pg.bayesfactor_binom(k=55, n=100, p=0.5)
print({"BF10": round(float(bf), 3)})

Interpretation

BF10 > 1 favors the alternative; BF10 < 1 favors the null. Rough labels: 1-3 anecdotal, 3-10 moderate, 10-30 strong, >30 very strong. BF01 = 1 / BF10.

Guardrails

State the prior scale (r) used; the Bayes factor depends on it.
A Bayes factor is relative evidence between two models, not the probability a hypothesis is true.
Do not equate a Bayes factor with a p-value threshold; report it as graded evidence.
Pair BF10 (or BF01) with the effect size and CI.
End result-bearing answers with one compact S0-S5 audit line.