菜单
Run or generate Pingouin code for rank-based non-parametric tests — Mann-Whitney U, Wilcoxon signed-rank, Kruskal-Wallis, Friedman, and Cochran Q — when outcomes are ordinal or parametric assumptions fail.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Screen psychology datasets before Pingouin analysis: variable types, missingness, duplicates, long/wide shape, assumption checks, and safer preprocessing.
Run or generate Pingouin code for multivariate comparisons — Hotelling's T-squared test (multivariate_ttest) — plus the multivariate assumption checks box_m (equal covariance) and multivariate_normality, for designs with several dependent variables.
Route psychology statistics requests to the smallest Pingouin workflow. Use when the user asks what analysis to run, gives a psychology design, asks to analyze data with Pingouin, or needs lower-token guidance before code generation.
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.
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.
Review and approve Pingouin psychology analyses before final reporting. Use after data screening or statistical output and before interpretation to catch common problems: wrong test, ignored repeated measures, missing assumption checks, multiple-comparison errors, unsupported causal claims, incomplete reporting, or output-format mismatches.
| name | pg-nonparametric |
| description | Run or generate Pingouin code for rank-based non-parametric tests — Mann-Whitney U, Wilcoxon signed-rank, Kruskal-Wallis, Friedman, and Cochran Q — when outcomes are ordinal or parametric assumptions fail. |
Use when the outcome is ordinal, or continuous but non-normal with small n, so a rank-based test is safer than a t test or ANOVA.
Read:
../../references/supervision-gates.md../../references/pingouin-api-quickref.md../../references/apa-output-template.md if writing results.pg.mwu(x, y) (Mann-Whitney U).pg.wilcoxon(x, y) (signed-rank).pg.kruskal(data, dv, between).pg.friedman(data, dv, within, subject).pg.cochran(data, dv, within, subject).pg.pairwise_tests(..., parametric=False, padjust="holm").holm.two-sided (mwu/wilcoxon only).Mann-Whitney U (independent):
x = df.loc[df["group"].eq("A"), "score"]
y = df.loc[df["group"].eq("B"), "score"]
res = pg.mwu(x, y, alternative="two-sided").round(3)
pg.print_table(res)
Wilcoxon signed-rank (paired):
res = pg.wilcoxon(df["pre"], df["post"], alternative="two-sided").round(3)
pg.print_table(res)
Kruskal-Wallis (k independent groups):
res = pg.kruskal(data=df, dv="score", between="group").round(3)
pg.print_table(res)
Friedman (k repeated conditions):
res = pg.friedman(data=df, dv="score", within="condition", subject="id").round(3)
pg.print_table(res)
Cochran Q (binary repeated):
res = pg.cochran(data=df, dv="passed", within="condition", subject="id").round(3)
pg.print_table(res)
mwu/wilcoxon return U_val/W_val, p_val, RBC (rank-biserial), CLES. kruskal returns H, ddof1, p_unc; friedman returns Kendall W, Q, p_unc; cochran returns Q, dof, p_unc. Report the effect size, not just p.
mwu is unpaired only; wilcoxon is paired only — do not swap them.