Alterlab Missing Data

Missing Data — Name the Mechanism, Impute Multiply, Pool by Rubin's Rules

Skill type: ANALYSIS MODULE. Missing data is not a nuisance to delete or fill with a mean. The discipline: state the missingness mechanism, then use a method whose uncertainty is honest — multiple imputation with Rubin's-rules pooling, or FIML. The dangerous shortcut is single imputation, which treats guessed values as observed and understates standard errors.

Core Mission

STATE THE MECHANISM (MCAR / MAR / MNAR). MULTIPLY IMPUTE AND POOL BY RUBIN'S RULES —
SINGLE IMPUTATION FAKES CERTAINTY IT DOESN'T HAVE.

When to Use This Skill

• "My dataset has missing values — how should I handle them?"
• "Should I use multiple imputation? How many imputations?"
• "How do I pool results across imputed datasets?"
• "Is mean imputation / listwise deletion okay here?"

Does NOT Trigger

Step 1 — state the mechanism (Rubin)

The mechanism is an assumption, largely untestable (MAR vs MNAR especially) — state it and justify it; inspect the missingness pattern (mice::md.pattern) and test MCAR (Little's test) as supporting evidence, not proof.

Step 2 — multiple imputation with Rubin's-rules pooling

Create m completed datasets (each imputes with added noise reflecting uncertainty), analyze each, then pool: the point estimate is the mean across imputations; the SE combines within-imputation and between-imputation variance (Rubin's rules), so it honestly reflects imputation uncertainty.

Python — statsmodels MICE:

from statsmodels.imputation import mice
imp = mice.MICEData(df)                                  # chained-equations imputer over the frame
fit = mice.MICE("y ~ x1 + x2", sm.OLS, imp).fit(n_burnin=10, n_imputations=20)
fit.summary()                                            # pooled estimates + SEs (Rubin's rules)

R — mice (field standard):

library(mice)
md.pattern(data)                                         # inspect the missingness pattern
imp  <- mice(data, m = 20, method = "pmm")               # m completed datasets
fit  <- with(imp, lm(y ~ x1 + x2))
summary(pool(fit))                                       # Rubin's-rules pooling (fmi, lambda, df)

FIML — for likelihood/SEM models under MAR, full-information ML uses all available data without explicit imputation; route to alterlab-sem-psychometrics when the model is latent-variable.

The single-imputation trap (verified caveat)

• Mean / regression / single imputation treats imputed values as if observed → biased, understated SEs. Not acceptable as a final analysis.
• scikit-learn IterativeImputer returns a single completed dataset and does not pool by Rubin's rules — the sklearn docs say so explicitly. It is fine for a prediction pipeline, but using it as "multiple imputation" for inference understates uncertainty. If you must, run it repeatedly with sample_posterior=True and different seeds and pool manually — but mice/statsmodels MICE do this correctly out of the box.

How many imputations (m)

Rule of thumb: m at least the percentage of incomplete cases (higher fraction of missing information → larger m); modern guidance often uses m ≈ 20–50. Report m and the fraction of missing information (FMI).

Reporting checklist

MECHANISM:   MCAR / MAR / MNAR — stated and justified; pattern inspected (md.pattern)
METHOD:      MICE (m, imputation model) with Rubin's-rules pooling | FIML | (deletion only if MCAR)
POOLING:     estimates + SEs pooled by Rubin's rules; report FMI
SENSITIVITY: for MNAR, a sensitivity analysis (selection / pattern-mixture)

References

• references/mechanisms_and_pooling.md — MCAR/MAR/MNAR detail, Rubin's-rules math, FIML vs MI, m guidance, the sklearn caveat.

Part of the AlterLab Academic Skills suite.