DataTerminology.comNEW
Also known as: Missing Value Imputation
Replacing missing data with plausible values to maintain completeness and reduce bias in analyses.
DESCRIPTION
Imputation addresses gaps in datasets by substituting missing values with estimates. Simple methods include mean, median, and mode; model-based methods include regression, k-NN, and EM. Multiple imputation propagates uncertainty by creating several completed datasets and combining estimates. Method choice depends on missingness mechanisms (MCAR, MAR, MNAR) and analysis goals. Poorly chosen imputations can bias results and overstate confidence. Rigorous diagnostics and sensitivity analyses help validate assumptions and impacts.
ALTERNATIVE DEFINITIONS
Wikipedia
Imputation is the process of replacing missing data with substituted values.
Direct, concise definition focused on missing data.
OECD Glossary of Statistical Terms
Imputation is the process used to determine and assign replacement values for missing, invalid or inconsistent data.
Broadens scope to invalid or inconsistent data beyond missingness.
RELATED TERMS
Missing Data MCAR MAR MNAR Multiple Imputation KNN Imputation Expectation-Maximization Systematic Error
COMPETING TERMS
Listwise Deletion Pairwise Deletion
RELATIONS
KNOWLEDGE AREAS
Data Science Mathematics Data Quality Testing & Validation
TARGET AUDIENCE
Data Scientist Analytics Engineer Data Analyst Data Steward
EXTERNAL SOURCES
- Wikipedia - Imputation (statistics)
- scikit-learn - API and guides for imputers
- statsmodels - Missing data handling in statsmodels
- National Library of Medicine - Tutorial on multiple imputation
TOP QUESTIONS
When should I use Multiple Imputation instead of single imputation?
Use Multiple Imputation when uncertainty needs to be reflected in downstream inference. It reduces bias and underestimation of variance compared to single fills.
Category: Practice
How do missingness mechanisms affect Imputation choice?
MCAR supports simpler methods; MAR often requires model-based imputations; MNAR may need explicit modeling or sensitivity analysis. Assumptions should be tested where possible.
Category: Requirements
What are the risks of naive mean imputation?
It distorts distributions, attenuates variance, and weakens correlations, potentially biasing models and hypothesis tests.
Category: Risks
How do I evaluate Imputation quality?
Use holdout missingness simulations, compare distributions, check model performance, and run sensitivity analyses across imputation strategies.
Category: Testing
Is Imputation always better than deletion?
Not always. With MCAR and small missing rates, listwise deletion can be unbiased and simpler. Evaluate trade-offs per analysis.
Category: Opinion
Can Imputation harm causal inference?
Yes, if it violates ignorability or introduces collider bias. Carefully align imputations with causal assumptions and models.
Category: Risks
Which algorithms commonly support Imputation in practice?
MICE, k-NN imputer, Bayesian regression, and iterative EM-based methods are widely supported in libraries like scikit-learn and statsmodels.
Category: Tools
How does Imputation interact with feature scaling and encoding?
Impute before scaling to avoid leaking information. For categorical data, use special categories or model-based imputations that respect levels.
Category: Practice
What documentation should accompany Imputation in production?
Record missingness patterns, chosen methods, assumptions, parameter settings, and monitoring plans for drift or changing data mechanisms.
Category: Governance