Multiple imputation by chained equations (MICE) is a method of missing-data imputation that represents uncertainty about missing observations by generating several plausible completed datasets. It specifies a separate conditional model for each variable with missing values and repeatedly updates the missing entries using the current values of the other variables. Each completed dataset is analyzed separately, and the resulting estimates and uncertainty measures are combined. MICE is particularly useful for datasets containing a mixture of continuous, binary, categorical, and other variable types.
Statistical basis
MICE implements multiple imputation, a framework in which missing values are treated as uncertain rather than replaced by a single definitive guess. Differences between completed datasets reflect uncertainty about the unobserved values; uncertainty remaining within each completed dataset is represented by the statistical analysis performed on that dataset.
For example, a dataset might contain income, age, employment status, and educational attainment, with missing entries in several columns. A chained-equations procedure can model income using the other variables, then model employment status using income and the remaining variables, and continue through the incomplete columns. Later updates use values imputed during earlier updates.
Unlike methods that begin with one explicit joint probability distribution for all variables, MICE specifies their conditional distributions individually. This approach is called fully conditional specification. It provides modeling flexibility, but separately chosen conditional models do not necessarily correspond to a coherent joint distribution.
Missingness assumptions
The validity of an imputation analysis depends on both its models and the process that generated the missing values. Three standard missing-data mechanisms distinguish the main cases:
- Missing completely at random (MCAR): missingness does not depend on either observed or unobserved data values.
- Missing at random (MAR): conditional on the observed data, missingness does not additionally depend on the unobserved values.
- Missing not at random (MNAR): missingness still depends on unobserved values after conditioning on the observed data.
Standard MICE analyses commonly rely on MAR, together with suitable model specification and conditions that permit the missingness process to be ignored for the intended inference. MAR does not mean that missingness is arbitrary. For example, income may be more often missing among younger respondents without violating MAR if age is observed and adequately represented in the imputation models.
MAR generally cannot be distinguished from MNAR using the observed data alone. Standard chained equations do not automatically address MNAR. Such analyses require additional assumptions about missingness or the unobserved values, often examined through sensitivity analysis.
Algorithm
Let the dataset contain variables , and let denote all variables except . For each incomplete variable, a conditional model is specified:
where denotes that model’s parameters.
A typical implementation proceeds as follows:
- Initialize missing entries. Supply provisional values, for example by sampling from observed entries.
- Select an incomplete variable. Temporarily regard its originally missing entries as unknown.
- Fit its conditional model. Use records in which that variable was observed, with the current observed or imputed values of its predictors.
- Draw model parameters. In a Bayesian implementation, draw from its posterior distribution. Other implementations use suitable resampling or approximate procedures to represent parameter uncertainty.
- Draw replacement values. Generate values for the missing entries from the conditional predictive distribution.
- Cycle through the remaining incomplete variables.
- Repeat the cycles. After an initial stabilization period, retain a completed dataset.
The process is repeated to obtain completed datasets. Implementations may use separately initialized chains or retain appropriately spaced states from a longer run. Observed entries remain unchanged in ordinary imputation.
At cycle , a sequential update has the schematic form
Here each combines its original observed values with its current imputations. Variables visited earlier in a cycle contribute their newly updated values; variables visited later contribute values from the preceding cycle.
The number of cycles and the number of completed datasets serve different purposes. Additional cycles may improve stabilization within a chain, whereas additional imputations reduce simulation error in pooled results. Increasing either number does not correct a misspecified model.
Conditional models
Different variables can use different model families:
| Variable type | Common conditional model |
|---|---|
| Approximately continuous | [[linear-regression |
| Binary | [[logistic-regression |
| Unordered categorical | Multinomial logistic regression |
| Ordered categorical | Ordinal regression |
| Count | Poisson or negative binomial regression |
| Continuous with a suitable observed donor pool | Predictive mean matching |
Choosing a model involves more than matching the broad variable type. Important features include skewness, bounds, excess zeros, nonlinear relationships, and interactions. For example, ordinary Gaussian regression can generate impossible negative values for a strictly positive variable.
Predictive mean matching
Predictive mean matching uses a fitted prediction model to identify observed records whose predicted values are close to those of a record with a missing value. One of these donors is selected, and its observed value supplies the imputation.
Because it uses observed donor values, the method can preserve the observed support and reduce sensitivity to some distributional misspecifications. It does not eliminate modeling assumptions. Sparse donor pools, poor predictors, or inadequate representation of a subgroup can produce weak imputations, and the method cannot naturally extrapolate beyond the available donor values.
Parameter and residual uncertainty
For proper multiple imputation, imputations generally need to represent both uncertainty in fitted model parameters and unexplained variation conditional on those parameters. Replacing missing entries only with fitted conditional means usually understates uncertainty and compresses variability.
Drawing regression coefficients alone is also insufficient for an ordinary continuous-response model: predictive draws ordinarily include residual variation. The precise uncertainty-generating procedure depends on the imputation method.
Combining completed-data analyses
Suppose a scalar quantity , such as a regression coefficient, is estimated in each of completed datasets. Let be its estimate and its estimated complete-data variance in dataset .
The pooled estimate is
The average within-imputation variance is
and the between-imputation variance is
Rubin’s rules give the total variance estimate
The pooled standard error is . The between-imputation component accounts for uncertainty associated with the missing data, while the factor includes a finite-imputation adjustment.
Confidence intervals and significance tests commonly use a Student’s -distribution with an appropriate degrees-of-freedom approximation. Small-sample adjustments may be important. Parameters with strongly asymmetric distributions may require transformation before pooling; for example, odds ratios are commonly pooled on the logarithmic scale.
Pooling is performed on analysis results, not by averaging the completed datasets into a single dataset. Likewise, simply averaging separate -values is not a general multiple-imputation testing procedure. Multivariate tests and some performance measures require specialized pooling methods.
Compatibility and convergence
Compatibility of conditional models
A collection of conditional models is compatible if there exists a joint distribution having those models as its full conditional distributions. Arbitrarily chosen regression equations need not satisfy this requirement.
When the updates are genuine full conditional draws from a coherent joint model, the procedure is closely related to Gibbs sampling, a form of Markov chain Monte Carlo. General MICE procedures do not automatically have this interpretation.
Under incompatible specifications, theoretical guarantees can be weaker, and limiting behavior may depend on the update order. Nevertheless, incompatibility alone does not establish that a particular inferential result is invalid: performance also depends on the estimand, the missingness pattern, and the relevant model properties.
Compatibility with the analysis
A second issue is the relationship between the imputation models and the eventual analysis model. An imputation model that omits an important interaction or nonlinear term can weaken that relationship in the completed data.
For example, imputing a covariate using only additive linear relationships may distort an analysis that depends on its squared value or its interaction with another variable. Substantive-model-compatible approaches construct imputations to respect specified features of the analysis model. More generally, congeniality concerns whether the imputation and analysis procedures support mutually consistent inferential assumptions.
Diagnostics
Common diagnostics include:
- Trace plots of imputed means, standard deviations, or selected model parameters across cycles.
- Comparisons between independently initialized chains.
- Comparisons of observed and imputed distributions, including comparisons within relevant subgroups.
- Checks for impossible combinations, boundary violations, and implausible relationships.
- Assessment of the stability of pooled estimates as the number of imputations increases.
Observed and imputed distributions need not match: under MAR, the records with missing values may systematically differ from fully observed records. Distributional differences therefore require interpretation rather than automatic rejection.
Stable chains do not establish that MAR is correct or that the imputation models are adequate. Posterior predictive checking and related model checks can expose some shortcomings but cannot verify assumptions about unobserved values.
Model construction and data structure
Imputation models often include variables used in the final analysis, predictors of incomplete variables, and observed predictors of missingness. Auxiliary variables can make MAR more plausible and improve precision, although excessive or poorly supported predictors can cause instability.
In explanatory analyses, an observed outcome is often included when imputing missing predictors so that their associations are preserved. In predictive machine learning, however, evaluation requires a separate consideration: preprocessing and imputation must not use information from held-out observations that would be unavailable at prediction time. Imputation belongs within the training portion of each relevant resampling split.
Derived variables require consistent handling. A total, ratio, or squared term may be recalculated from its component variables after updates, a procedure often called passive imputation. Maintaining an arithmetic identity in this way does not by itself guarantee compatibility with the analysis model.
Clustered, longitudinal, and survey data introduce additional structure. Ordinary single-level MICE can misrepresent within-group dependence or uncertainty if it ignores clustering. Multilevel conditional models, appropriately specified cluster effects, or survey-aware procedures may be needed. Structural missingness—such as a response that is not applicable—must also be distinguished from an unknown but meaningful value.
Applications, limitations, and software
MICE is used in survey research, observational studies, economics, and other settings with incomplete multivariable data. Its main practical advantage is the ability to combine variable-specific models without requiring a simple joint distribution for the entire dataset.
Important limitations include sparse categories, separation in logistic models, strongly correlated predictors, weak information about heavily incomplete variables, and unsupported extrapolation. Conditional models may need regularization or other stabilization. No imputation procedure can reliably recover information absent from both the observed data and the assumptions supplied by the model.
There is no universally sufficient number of imputations. The required number depends on the amount of missing information and the desired precision of estimates, standard errors, intervals, and tests. A number adequate for a stable point estimate may be inadequate for stable uncertainty estimates.
Widely used implementations include the R package mice and Stata’s mi impute chained. Software differs in its conditional models, predictor specifications, handling of parameter uncertainty, diagnostics, and support for complex data structures. The method name alone therefore does not fully specify an imputation analysis.