exekiel179/pingouin-psych-stats/pg-reliability

PG Reliability

Use for internal consistency and inter-rater/test-retest reliability.

Load

Read:

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

Function Choice

• Internal consistency of multiple scale items -> pg.cronbach_alpha.
• Inter-rater, test-retest, or target-by-rater reliability -> pg.intraclass_corr.

Required Inputs

• Item columns and reverse-coded item list for alpha.
• Whether items are all intended to measure one construct.
• Long-format ICC columns: target, rater, rating.
• ICC model decision: single vs average, consistency vs agreement, one-way vs two-way.

Code Patterns

Cronbach alpha:

items = df[["item1", "item2", "item3", "item4"]]
alpha, ci = pg.cronbach_alpha(data=items)
print({"alpha": round(alpha, 3), "ci95": ci})

Item-total screening:

for col in items.columns:
    alpha_drop, ci_drop = pg.cronbach_alpha(data=items.drop(columns=col))
    print(col, round(alpha_drop, 3), ci_drop)

ICC:

icc = pg.intraclass_corr(data=df, targets="target",
                         raters="rater", ratings="rating").round(3)
pg.print_table(icc)

Reporting

• Alpha: number of items, alpha, CI, and whether any reverse coding was applied.
• ICC: exact row/type from Pingouin output, ICC value, CI, F, df, p if available, and model interpretation.

Guardrails

• Alpha does not prove unidimensionality; mention factor structure if relevant.
• Low alpha can reflect few items, multidimensionality, poor items, or restricted range.
• Do not choose ICC row after seeing the most favorable value; pick model from design first.
• Make sure each target-rater pair is valid; duplicated ratings can corrupt ICC.
• End result-bearing answers with one compact S0-S5 audit line.