exekiel179/pingouin-psych-stats/pg-regression-mediation

PG Regression Mediation

Use for regression-style questions that Pingouin supports directly.

Load

Read:

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

Function Choice

• Continuous outcome, additive linear predictors -> pg.linear_regression.
• Binary outcome -> pg.logistic_regression with X, y.
• Single or multiple mediator path model -> pg.mediation_analysis.

Required Inputs

• Outcome variable and scale.
• Predictor list.
• Covariates and whether they are theoretical controls.
• Binary coding for logistic regression.
• Mediation paths: x, m, y.
• Bootstrap count and seed for mediation.

Code Patterns

Linear regression:

vars_needed = ["outcome", "x1", "x2"]
tmp = df.dropna(subset=vars_needed)
res = pg.linear_regression(tmp[["x1", "x2"]], tmp["outcome"],
                           add_intercept=True).round(3)
pg.print_table(res)

Logistic regression:

vars_needed = ["binary_outcome", "x1", "x2"]
tmp = df.dropna(subset=vars_needed)
res = pg.logistic_regression(tmp[["x1", "x2"]], tmp["binary_outcome"],
                             remove_na=False).round(3)
pg.print_table(res)

Mediation:

res = pg.mediation_analysis(data=df, x="x", m="mediator", y="outcome",
                            covar=["age"], n_boot=5000,
                            seed=42).round(3)
pg.print_table(res)

Reporting

• Regression: report coefficient, SE, statistic, p, CI, and model R-squared/adjusted R-squared if returned.
• Logistic: report coefficient and convert to odds ratio when useful with np.exp(coef).
• Mediation: report total, direct, indirect paths, bootstrap CI, seed, and bootstrap count.

Guardrails

• Pingouin linear regression is OLS; it is not a mixed model or SEM package.
• For Pingouin 0.6.1, logistic_regression takes X, y; verify the signature before using examples from older wrappers.
• Check residual assumptions and influential cases when substantive conclusions depend on them.
• Do not present mediation as causal unless temporal order, design, and assumptions support it.
• Do not hide automatic missing-data removal; report effective n.
• Mediation bootstrap p-values vary with seed and bootstrap count.
• End result-bearing answers with one compact S0-S5 audit line.