Statistics Glossary → Multiple imputation by chained equations (FCS/MICE) and PMM
Multiple imputation by chained equations (FCS/MICE) and PMMconcept
A multiple imputation approach that fills in missing values sequentially and iteratively with a separate model predicting each variable from the others, producing several completed data sets.
In fully conditional specification (FCS), or MICE, a suitable model is chosen for each incomplete variable: linear regression or PMM for continuous variables, logistic regression for binary ones and polytomous regression for multi-category ones. The algorithm cycles through the variables, predicts each from the current values of the others and fills it with randomly drawn values that also reflect parameter uncertainty. The cycle is repeated several times, and the process is run independently m times to obtain m completed data sets; the analysis is carried out in each and the results are pooled with Rubin's rules.
In predictive mean matching (PMM) the imputed value is not the regression prediction itself: one of several donors whose predicted value is closest to that of the incomplete case is chosen, and its observed value is used. Imputed values therefore always stay in a realistic range (no negative ages or half children) and skewed distributions are preserved. The method assumes that data are missing at random (MAR); convergence of the chain should be checked with trace plots.
An everyday example
In a health screening, weight is missing for 15% of participants and income for 10%. MICE cycles by predicting weight from age, height and income, then income from age, education and weight. Thanks to PMM, imputed weights are drawn from weights actually observed; a regression is fitted in each of 20 completed data sets, the coefficients are pooled and the extra uncertainty caused by missingness is reflected in the standard errors.
In MerQur
The 'Imputation method' option in Multiple Imputation (by default FCS/MICE with PMM for numeric and logreg/polyreg for categorical variables; alternatively FCS norm or IterativeImputer + BayesianRidge); the 'Convergence trace' under the 'Chart' option shows the mean and SD of the FCS chain across iterations.
Analyses where this concept appears
Related concepts
Rubin's rules (pooling imputation results)Auxiliary variables (imputation model)