Pg categorical

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

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-categorical

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

Run or generate Pingouin code for categorical / contingency-table analyses — chi-square test of independence, McNemar's paired test, 2x2 crosstabs, and chi-square power — for psychology data with nominal variables.

SKILL.md

2.7 KB, 698 tokens by cl100k_base, as published. Nobody here has run it

PG Categorical

Use when both variables are categorical (nominal) and the question is association or change in proportions.

Load

Read:

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

Decision Rules

Association between two independent categorical variables (any R x C) -> pg.chi2_independence(data, x, y).
Change in a binary outcome for the same participants (paired 2x2) -> pg.chi2_mcnemar(data, x, y).
Just the 2x2 table from two binary columns -> pg.dichotomous_crosstab(data, x, y).
Sample size / power for a chi-square test -> pg.power_chi2(dof, w, n, power, alpha).

Required Inputs

Two categorical columns (row and column variables).
Whether observations are independent (chi-square) or paired within participants (McNemar).
For McNemar/crosstab: columns must be dichotomous (0/1 or two levels).
For power: effect size w (Cohen), degrees of freedom, and the unknown to solve (set to None).

Code Patterns

Chi-square test of independence (returns a 3-tuple):

expected, observed, stats = pg.chi2_independence(data=df, x="group", y="response")
pg.print_table(stats.round(3))          # read the "pearson" row

McNemar paired test (binary 0/1 columns; returns observed table + stats):

observed, stats = pg.chi2_mcnemar(data=df, x="before", y="after")
pg.print_table(stats.round(3))

2x2 crosstab:

ct = pg.dichotomous_crosstab(data=df, x="cond_a", y="cond_b")
print(ct)

Chi-square power / sample size:

power = pg.power_chi2(dof=1, w=0.3, n=100, alpha=0.05)   # solve n via n=None, power=0.8
print({"power": round(float(power), 3)})

Output Checks

chi2_independence stats rows include pearson; columns are test, lambda, chi2, dof, pval, cramer (effect size), power. Report Cramér's V, not just chi-square. chi2_mcnemar stats give chi2, dof, p_approx, p_exact; prefer p_exact for small discordant counts.

Guardrails

Chi-square needs adequate expected counts; if many cells < 5, note it and consider Fisher's exact (outside Pingouin).
McNemar requires paired dichotomous data, not two independent groups.
Report the contingency table and Cramér's V / effect size, not only the p-value.
Chi-square shows association, not causation or direction.
End result-bearing answers with one compact S0-S5 audit line.