MICE — Multivariate Imputation by Chained Equations
- Introduced by
- Stef van Buuren & Karin Groothuis-Oudshoorn
- Year
- 2011
- Type
- Iterative multiple imputation algorithm
- Family
- Missing data
What does it do?
Multivariate Imputation by Chained Equations (MICE) is an iterative procedure for handling missing data in multivariate datasets.
Introduced by Stef van Buuren and Karin Groothuis-Oudshoorn through the R package mice (2011), the algorithm fills each missing variable using a separate regression model conditioned on all other variables, cycling through variables repeatedly until the imputed values converge. The result is m completed datasets that are analysed separately and combined using Rubin's rules.
The intuition behind the method
Imagine a partially completed crossword puzzle: each blank word is filled in by looking at the crossing letters already present. If a newly filled word reveals more letters, you revisit earlier blanks with the new information. MICE works the same way — it fills in one missing column at a time using all other columns as clues, then revisits every column again, iterating until the answers stabilise across many candidate puzzles.
History
The conceptual foundation of chained equations imputation emerged in the early 1990s from work by van Buuren, Boshuizen, and Knook on incomplete health survey data. The Fully Conditional Specification framework was formalised through the 2000s, with van Buuren and Groothuis-Oudshoorn releasing the mice R package in 2011 along with a comprehensive methodological paper in the Journal of Statistical Software, which became the canonical reference with tens of thousands of citations.
When to use it
MICE is appropriate when data are missing at random (MAR) and missingness occurs across multiple variables simultaneously. It suits continuous, binary, ordinal, and count outcomes through its flexible per-variable model specification. It is less suitable for data missing not at random (MNAR) without auxiliary modelling, for very high-dimensional datasets where the imputation model becomes unwieldy, or when missingness rates exceed 50% in key variables. Single imputation or complete-case analysis may suffice under MCAR with low missingness rates.
Limitations and considerations
- Assumes data are missing at random (MAR); violations require sensitivity analyses or pattern-mixture extensions
- Computational cost grows with the number of variables, iterations, and imputations (m)
- Convergence of the Gibbs-like sampler is not mathematically guaranteed for incompatible conditional models
- Model specification (predictor selection, imputation method per variable) requires careful analyst judgment
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Handles mixed variable types (continuous, binary, ordinal, count) via tailored per-column models
- Properly propagates uncertainty through multiple imputed datasets and Rubin's pooling rules
- Flexible predictor specification allows domain knowledge and auxiliary variables to improve imputation quality
- Well-validated R implementation (mice package) with extensive diagnostics and active maintenance
Applications
- Clinical trials and epidemiological cohort studies with intermittent follow-up measurements
- Survey data harmonisation where different questionnaire modules yield different missing patterns
- Machine learning preprocessing pipelines requiring complete feature matrices before model training
- Health economics and policy research combining administrative records with self-reported questionnaire data
- Longitudinal educational assessments with item non-response across multiple test waves
How it works
How it works
Before any cycling begins, MICE replaces all missing values with a quick placeholder — typically the observed column mean for continuous variables or the mode for categorical ones. This initial fill ensures that every variable has at least some values available as predictors when the first imputation cycle starts.
Formula
X^{(0)}_{\text{mis}} \leftarrow \text{simple imputation}(X_{\text{obs}})For each variable j with missing entries, MICE fits a predictor-appropriate model (linear regression, logistic regression, predictive mean matching, etc.) using all other variables X_{-j} as covariates. Missing values of X_j are then drawn from the posterior predictive distribution of that model, so imputed values reflect both the fitted relationship and parameter uncertainty.
Formula
\hat{X}_j \sim P\!\left(X_j \mid X_{-j},\, \theta_j\right), \quad j = 1, \ldots, pAfter imputing all p variables once, one full cycle is complete. MICE repeats the cycle — updating each variable's imputation model with the newly imputed values of its covariates — for a user-defined number of iterations (typically 5–20). Convergence is monitored by inspecting trace plots of imputed means and standard deviations across iterations.
Formula
\theta_j^{(t)} = \arg\max_{\theta_j} \prod_{i: x_{ij} \text{ obs}} P\!\left(x_{ij} \mid x_{i,-j}^{(t-1)},\, \theta_j\right)The full procedure (initialisation + cycling) is repeated m times independently to produce m completed datasets. Each dataset is analysed with the intended statistical method, yielding m point estimates and standard errors. Rubin's combining rules then merge these into a single estimate and a total variance T that correctly propagates both within-imputation and between-imputation uncertainty.
Formula
\bar{Q} = \frac{1}{m}\sum_{l=1}^{m} \hat{Q}_l, \quad T = \bar{U} + \left(1 + \frac{1}{m}\right) BCommon pitfalls
- Using too few imputed datasets (m < 5) underestimates standard errors and inflates false-positive rates
- Including the outcome variable in the imputation model is essential; omitting it biases regression coefficients toward null
- Ignoring interaction terms or non-linear relationships in the imputation model can distort downstream analyses
- Failing to check convergence via trace plots may lead to pooling results from non-converged chains
Frequently asked questions
How many imputed datasets (m) should I create?
A common default is m = 5, sufficient when missingness rates are low. White, Royston, and Wood (2011) recommend setting m approximately equal to the percentage of incomplete cases. For missingness rates above 20–30%, m = 20 or more is advisable to stabilise pooled estimates and standard errors.
Can MICE handle missing not at random (MNAR) data?
Standard MICE assumes MAR. Under MNAR the imputation model is misspecified, potentially biasing results. Analysts should conduct sensitivity analyses — for example, using delta-adjustment or pattern-mixture models — to assess how results change under plausible MNAR departures from the MAR assumption.
What imputation method should I choose for each variable?
The mice package selects defaults based on variable class: predictive mean matching (pmm) for continuous variables, logistic regression for binary, polytomous regression for unordered factors, and proportional odds for ordered factors. These defaults work well in most applied settings, but analysts may substitute random forests or other flexible learners for complex distributions.
Related methods
Sources
van Buuren, S., & Groothuis-Oudshoorn, K. (2011). mice: Multivariate imputation by chained equations in R. Journal of Statistical Software, 45(3), 1–67.
Open DOI referenceCopied.