Pg Data Screening

PG Data Screening

Use before analysis when data has not been inspected in the current thread, or when a test fails because of missing values, shape, categories, or assumptions.

Load

Read:

• ../../references/supervision-gates.md
• ../../references/pingouin-api-quickref.md
• ../../references/pingouin-optimization.md
• ../../references/archive-contract.md when running inside an archive.

Workflow

1. Inspect data columns, dtypes, sample size, missingness, duplicates, and group sizes.
2. Confirm long versus wide format for the target analysis.
3. Confirm repeated-measures identifiers are unique enough for the design.
4. Run only relevant assumption checks.
5. Produce a short analysis-readiness summary and minimal corrective code.

Standard Code Skeleton

import pandas as pd
import pingouin as pg

df = pd.read_csv("data.csv")
print(df.info())
print(df.describe(include="all"))
print(df.isna().mean().sort_values(ascending=False).head(20))
print(df.duplicated().sum())

For group balance:

print(df.groupby("group", dropna=False).size())

For normality by group:

pg.normality(data=df, dv="score", group="group")

For homogeneity:

pg.homoscedasticity(data=df, dv="score", group="group")

For repeated-measures sphericity:

pg.sphericity(data=df, dv="score", within="condition", subject="id")

For robust outliers (MAD-median rule) and a normality Q-Q plot:

mask = pg.madmedianrule(df["score"].to_numpy())   # True = flagged outlier
print({"n_outliers": int(mask.sum())})

import matplotlib.pyplot as plt                    # qqplot needs matplotlib
pg.qqplot(df["score"], dist="norm")
plt.savefig("qqplot.png", dpi=150, bbox_inches="tight")

Reporting

Report:

• Number of rows and unique participants.
• Missingness by required variable.
• Any impossible values or suspicious categories.
• Whether the data shape matches the intended Pingouin function.
• Assumption diagnostics and practical implication.
• Exact rows removed if listwise deletion is used.

Guardrails

• Assumption tests are diagnostic; do not mechanically switch tests solely from a significant normality test in a large sample.
• If repeated measures exist, never analyze rows as independent.
• If IDs repeat unexpectedly within a condition, stop and flag the design/data issue.
• If transformations or exclusions are proposed, state whether they were planned or post hoc.
• Flag outliers with a rule chosen in advance (e.g. MAD-median); never delete points silently.
• End with a compact audit line using S0-S5 gate codes.