| name | pingouin-stat-router |
| description | 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. |
Pingouin Stat Router
Compatibility entry point. For end-to-end work, prefer pingouin-stat (adds intake, approval, and deliverable planning around routing).
Use this skill first when the requested statistical analysis is unclear or broad.
Load
Read only these references unless another skill requires more:
../../references/workflow-index.md../../references/supervision-gates.md../../references/pingouin-api-quickref.md only when generating concrete code.../../references/apa-output-template.md only if the user asks for report text or formatted output.
Goal
Select the smallest downstream skill and analysis plan that fits the study design. Do not run a statistical test until the design, variables, and data shape are clear enough to avoid an obvious wrong test.
Required Questions To Resolve
Resolve from the user's text or data before choosing the test:
- Outcome variable: continuous, ordinal, binary, count, or categorical.
- Predictors: between-subjects, within-subjects, covariates, mediators, or raters/items.
- Unit of analysis: participant, trial, item, dyad, classroom, or other cluster.
- Dependency: independent observations, repeated measures, paired samples, nested observations, or multiple raters.
- Number of levels per factor.
- Hypothesis direction: two-sided by default unless explicitly directional and preregistered.
- Planned comparisons versus exploratory post hoc tests.
- Missing-data handling.
If these cannot be inferred, ask the minimum necessary clarifying question. If a reasonable default is safe, state it explicitly.
Routing
- Data quality, missingness, assumptions, reshape, outliers ->
pg-data-screening.
- One/two/paired/Welch/pairwise mean comparisons ->
pg-mean-tests.
- ANOVA family: one-way, factorial, RM, mixed, Welch, ANCOVA ->
pg-anova.
- Correlation: Pearson/Spearman/Kendall, partial, pairwise, repeated ->
pg-correlations.
- Linear/logistic regression, mediation ->
pg-regression-mediation.
- Ordinal / non-normal outcomes, rank tests ->
pg-nonparametric.
- Categorical / contingency, chi-square / McNemar ->
pg-categorical.
- Several DVs together, Hotelling's T-squared ->
pg-multivariate.
- Bayes factors (t, r, proportion) ->
pg-bayesian.
- Cronbach alpha, ICC, inter-rater reliability ->
pg-reliability.
- Sample size, achieved power, detectable effect ->
pg-power.
- APA/Chinese result prose, polished tables ->
pg-reporting.
Output Format
Return this compact block:
Recommended: <skill> using <function(s)>.
Why: <design-to-test mapping>.
Need: <columns/data shape/missing item>.
Guardrail: <one highest-risk assumption or correction>.
Next: <code / run screening / one clarifying question>.
Audit: <S0-S5 compact gate status>.
Hard Rules
- Do not invent column names; use placeholders only in code templates and mark them clearly.
- Do not claim causality from correlational or cross-sectional data.
- Do not treat Likert single items as interval without noting the assumption.
- Do not ignore repeated measures or clustering.
- Prefer
pairwise_tests, not deprecated pairwise_ttests.
- Prefer
alternative="two-sided" unless the user gives a directional hypothesis.
