Skip to content
Back to search results

Multiple Imputation — MICE

Introduced by
Donald B. Rubin
Year
1987
Type
Missing-data handling procedure
Font size

What does it do?

Multiple Imputation (MI), formally introduced by Donald B. Rubin in 1987, is a principled statistical procedure for handling missing data.

Rather than replacing each missing value once, MI fills the gaps m times — each time drawing plausible values from the posterior predictive distribution of the missing data — producing m complete datasets. Each dataset is analysed independently, and the results are combined into a single set of estimates using Rubin's pooling rules. The MICE variant (Multivariate Imputation by Chained Equations), popularised by van Buuren and Groothuis-Oudshoorn (2011), extends the approach to mixed variable types by imputing each variable in turn through a sequence of conditional regression models.

The intuition behind the method

Think of missing values as blank entries in a survey. Instead of guessing once and hoping for the best, multiple imputation makes m educated guesses — each slightly different, reflecting genuine uncertainty about what the true value might have been. The spread across those guesses carries the uncertainty forward into the analysis, so the final standard errors are honest rather than artificially narrow. If you had guessed only once, you would have pretended to know more than you really did; by guessing m times and combining the results, the uncertainty due to missingness is preserved in the reported estimates.

History

The formal statistical framework for multiple imputation was developed by Donald B. Rubin and presented in his 1987 monograph Multiple Imputation for Nonresponse in Surveys. Rubin derived the pooling rules — now universally called Rubin's rules — that combine estimates across imputed datasets in a way that properly accounts for the uncertainty introduced by missingness. The approach gained practical traction with the development of the MICE algorithm, which extends Rubin's framework to datasets with variables of mixed types by imputing each variable sequentially through a chain of conditional models. The landmark software implementation mice for R, published by van Buuren and Groothuis-Oudshoorn in the Journal of Statistical Software in 2011, made the method accessible to applied researchers across fields and established MICE as the dominant implementation strategy.

When to use it

Multiple imputation is appropriate when a dataset has missing values and the missingness mechanism is plausibly MAR or MCAR. It applies to continuous, categorical, binary, ordinal, and mixed-type variables and fits cross-sectional, longitudinal, and panel data structures. A minimum sample size of approximately 30 is required. The method suits descriptive, predictive, and relational research purposes. When missingness is MNAR, standard multiple imputation is insufficient and must be supplemented by a sensitivity analysis or a selection model. Listwise deletion (complete-case analysis) is acceptable only when data are MCAR and the proportion of missing data is very small; otherwise MI is the preferred approach.

Limitations and considerations

  • Requires the MAR or MCAR assumption; cannot correct for MNAR missingness without additional modelling.
  • Computationally more demanding than single imputation or listwise deletion, especially with large m or large datasets.
  • The quality of imputed values depends on the imputation model — omitting strong predictors of missingness from the imputation model biases results.
  • With high fractions of missing data (above ~50 %), even correctly specified multiple imputation may yield wide confidence intervals.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Produces unbiased estimates under MAR, unlike listwise deletion which requires the stronger MCAR assumption.
  • Propagates uncertainty from missing data into standard errors and confidence intervals through Rubin's pooling rules, avoiding the false precision of single imputation.
  • Handles mixed variable types — continuous, binary, ordinal, nominal — within the same MICE framework.
  • Compatible with virtually any downstream analysis model, requiring no modification to the intended analysis.
  • Well-supported in standard statistical software (R mice, SPSS MULTIPLE IMPUTATION, Python IterativeImputer / miceforest).

Applications

  • Handling item non-response in large-scale survey datasets in social and health sciences.
  • Preparing clinical trial data for analysis when follow-up measurements are missing.
  • Longitudinal and panel studies where participants drop out over time.
  • Econometric modelling with administrative records that have incomplete fields.
  • Preprocessing step in any analysis pipeline where downstream models require complete data.

How it works

How it works

1

Before imputing, the missingness pattern is examined to assess whether the mechanism is plausibly Missing Completely At Random (MCAR) or Missing At Random (MAR). MCAR means the probability of missingness does not depend on any observed or unobserved data. MAR means it may depend on observed values but not on the unobserved missing values themselves. If the mechanism is Missing Not At Random (MNAR) — where the probability of a value being missing depends on that value itself — simple multiple imputation is not sufficient and a sensitivity analysis is required. The fraction of missing data informs how many imputed datasets m are needed; a common guideline is m ≥ 100 × (fraction missing).

2

The MICE algorithm cycles through each incomplete variable and fits a regression model predicting that variable from all others (including previously imputed values). A draw from the posterior of this model is used to fill in the missing entries. The cycle repeats for all incomplete variables and is iterated until the imputed distributions stabilise — commonly 50 or more iterations. This produces one completed dataset. The whole process is repeated from scratch m times with different random seeds, yielding m independently imputed complete datasets. The choice of imputation model (predictive mean matching, logistic regression, classification trees, etc.) depends on the variable type: continuous, binary, ordinal, or nominal.

3

Each of the m completed datasets is analysed with whatever statistical model the researcher originally intended — regression, ANOVA, structural equation modelling, and so on. This step produces m sets of parameter estimates and standard errors, all slightly different because the m datasets differ in how the missing values were filled.

4

The m sets of estimates are combined into a single set using Rubin's rules. The pooled point estimate is the arithmetic mean of the m individual estimates. The pooled variance has two components: the average within-imputation variance (what the analysis would have given with no missingness) and the between-imputation variance (the extra uncertainty from not knowing the missing values). The ratio of between-imputation to total variance is the fraction of missing information, a useful diagnostic. Degrees of freedom are adjusted accordingly. The result is a single table of estimates and confidence intervals that honestly reflects both sampling variability and the uncertainty due to missing data.

Common pitfalls

  • Setting m too low — a common rule of thumb is m ≥ 5, but with non-trivial missing fractions a much higher m (e.g., m ≥ 100 × fraction missing) is needed to achieve stable, reproducible estimates.
  • Omitting the outcome variable or other important predictors from the imputation model produces biased imputations; the imputation model must be at least as rich as the analysis model.
  • Applying multiple imputation when the mechanism is MNAR without a supplementary sensitivity analysis gives results that appear valid but may be substantially biased.
  • Pooling results with Rubin's rules assumes that the analysis model is the same for all m datasets and that convergence of the MICE algorithm has been checked — inspect trace plots of imputed means and variances to verify convergence.

Frequently asked questions

How many imputed datasets (m) do I need?

The traditional minimum of m = 5 is often too low. A widely cited practical guideline is m ≥ 100 × (fraction of missing data expressed as a percentage), so that 10 % missingness calls for at least m = 10 and 30 % missingness calls for at least m = 30. Higher m reduces Monte Carlo error in the pooled estimates and is especially important for reproducibility.

What is the difference between MCAR, MAR, and MNAR?

MCAR (Missing Completely At Random) means the probability of a value being missing is unrelated to any data, observed or unobserved — data are missing by pure chance. MAR (Missing At Random) means the probability of missingness can depend on observed values but not on the unobserved missing value itself — once you condition on what you have observed, the missingness is random. MNAR (Missing Not At Random) means the probability of missingness depends on the missing value itself. Multiple imputation is valid under MCAR and MAR; under MNAR it can still be run but must be supplemented by a sensitivity analysis.

Does multiple imputation create real data?

No. Multiple imputation creates plausible values drawn from a model of the missing data distribution; these are not the true missing values. The method's purpose is not to recover the original data but to produce estimates and standard errors for the analysis model that correctly account for the uncertainty introduced by missingness.

Can I use multiple imputation with categorical or ordinal variables?

Yes. The MICE framework handles mixed variable types by fitting a type-appropriate conditional model for each variable: predictive mean matching or linear regression for continuous variables, logistic regression for binary variables, polytomous logistic regression or ordinal regression for nominal or ordinal variables, and classification and regression trees as a flexible non-parametric option. Each variable is imputed in turn using the models appropriate to its type.

Related methods

Sources

  1. Rubin, D.B. (1987). Multiple Imputation for Nonresponse in Surveys. Wiley.

    Open DOI reference
    Copied.
  2. van Buuren, S. & Groothuis-Oudshoorn, K. (2011). mice: Multivariate Imputation by Chained Equations in R. Journal of Statistical Software, 45(3), 1–67.

    Open publication reference
    Copied.