Missing data handling with MICE imputation

Description

Replacing missing survey or behavioural values with multiple plausible imputations, then pooling results, so estimates are not biased by who skipped questions.

Missing data are rarely random: busy, dissatisfied or less literate users skip more items. Listwise deletion discards cases and biases estimates unless data are missing completely at random (MCAR); single mean imputation shrinks variance. Multiple imputation by chained equations models each incomplete variable from the others in turn, generates m complete datasets with appropriate random variation, analyses each, and pools estimates with Rubin's rules so uncertainty from missingness is reflected in standard errors. It is valid under missing at random (MAR) assumptions given the variables included. Rigour comes from diagnosing the missingness pattern, including auxiliary variables that predict missingness, imputing enough datasets, checking imputed distributions and running sensitivity analyses for missing-not-at-random scenarios.

Attributes

Loss type
Unmigrated
Effort
Days

Steps

  1. Describe the missingness pattern: percentage missing per variable and per respondent, and which variables predict missingness.
  2. Decide the assumption (MCAR, MAR or MNAR) and include auxiliary variables related to missingness in the imputation model.
  3. Impute with MICE (e.g. scikit-learn IterativeImputer with sample_posterior=True, or statsmodels MICE), using m of at least 20 or roughly the percentage of incomplete cases.
  4. Check convergence across iterations and compare imputed with observed distributions.
  5. Fit the analysis model on each imputed dataset and pool estimates and standard errors with Rubin's rules.
  6. Compare against complete-case results and a delta-adjusted MNAR sensitivity scenario, and report all three.

Sources

  • factor_analyzer
  • semopy
  • pyreadstat (SPSS import)
  • scikit-learn (MICE)
  • pingouin (reliability)
  • pandas
  • Rubin, D. B. (1987) Multiple Imputation for Nonresponse in Surveys
  • van Buuren, S. (2018) Flexible Imputation of Missing Data, 2nd ed.
  • The Inquiry Series, C3 — Survey Pipeline in Python — From Raw Responses to Structural Equation Model

Also known as

  • Multiple imputation by chained equations
  • Fully conditional specification

Categories

  • Analytics & Behaviour
  • Statistics & Modelling