Multiple Imputation as Repeated Plausible Completion

Why Complete the Data More Than Once?

When values are missing, a regression cannot use those values as if they were observed. Replacing each missing value with one fixed guess produces a single completed data set, but it hides uncertainty about what the missing values might have been. Regression imputation with residual variation takes a step toward representing that uncertainty. Multiple imputation extends the idea by creating several completed data sets, each with a different plausible set of missing values.

The goal is not to choose the completion that looks most convenient or to treat the several data sets as if they were new independent samples. They are alternative representations of the same incomplete sample. The analyst fits the same regression to each one, then combines the estimates and their uncertainty. Variation among the estimates carries information about how much the missing values matter to the result.

The Basic Sequence

Suppose the scientific question concerns a regression of an outcome \(Y\) on a predictor \(X\). For example, \(Y\) might be a blood-pressure measurement and \(X\) might be hours of exercise per week. The intended regression specification is fixed before examining the results—for a simple linear regression, it could be \(Y_i=\beta_0+\beta_1X_i+\varepsilon_i\), where \(\beta_1\) is the population slope and \(\varepsilon_i\) represents deviations from the regression mean.

If some \(X\) or \(Y\) values are missing, multiple imputation follows this general sequence:

The imputations should be stochastic: repeated completions need not assign exactly the same values to missing entries. The draws should also reflect uncertainty in estimating the imputation model, rather than merely inserting predictions from one fixed fitted model. Otherwise the completed data sets can look more certain than the information warrants.

What the Two Sources of Uncertainty Mean

Each completed-data regression has ordinary sampling uncertainty: even if that particular completed data set were known, a coefficient estimate would vary from sample to sample. This is represented by the estimate's standard error within that regression. There is also missing-data uncertainty: different plausible values for the unobserved entries can lead to different fitted coefficients.

The distinction can be expressed using the law of total variance. Let \(Z_{\text{mis}}\) denote the missing values and let \(Q\) be a regression quantity of interest, such as a slope. Conditional on the observed data, uncertainty about \(Q\) can be separated into uncertainty remaining after the missing values are specified and uncertainty arising because those values are not known:

The first term corresponds to average within-completion uncertainty; the second corresponds to between-completion variation. Multiple imputation aims to represent both. This explains why averaging estimates alone does not finish the uncertainty calculation: the estimates' variation across imputations and their within-regression standard errors also matter. The exact pooling calculations and inferential procedures are separate details; the essential point here is that neither source of uncertainty should be discarded.

The same regression specification is important. If one completion uses a different predictor set or functional form, the resulting estimates no longer answer exactly the same question. Averages across unlike analyses can be difficult to interpret, and their variation cannot be attributed simply to uncertainty about missing values.

Conditions and Limits of Interpretation

Multiple imputation is a way to propagate uncertainty under a model; it does not make missingness harmless by itself. A common working condition is that missingness is missing at random (MAR) conditional on information included in the analysis of missingness. In this setting, the probability that an entry is missing may depend on observed information, but not on the unobserved value after conditioning on that information. If missingness still depends on the unobserved value after conditioning, a MAR-based analysis may not capture the relevant uncertainty.

The imputation model also needs to be credible. It should preserve important relationships needed for the regression analysis, and its distributional assumptions should be suitable for the variables being imputed. In addition, the draws need to represent plausible variability rather than repeated copies of a single prediction. A large number of completed data sets cannot repair a systematically inappropriate model or a missingness assumption that does not fit the setting.

A practical consequence is that multiple imputation is not a method for recovering the unknowable true values of every missing observation. The individual imputed values are plausible draws, not observed facts. The target is a regression analysis whose estimate and uncertainty appropriately reflect the available data and the assumptions used to complete it.

Why Repeated Plausible Completion Matters

A single completed data set can make the analysis appear more precise than it is because it acts as though the inserted values were known. Multiple imputation instead asks how the regression would look across several plausible completions of the same incomplete sample. When the resulting estimates are similar, the missing values have relatively little influence on that coefficient under the imputation model. When they differ substantially, the uncertainty about the missing entries is consequential and should be visible in the final inference.

A common mistake is to fit the regression in every imputed data set, average the coefficient estimates, and then report the standard error from just one of those fits. That standard error reflects uncertainty within only that completion. Correction: use a pooling procedure that incorporates within-completion sampling uncertainty and between-completion variation. The average coefficient is a useful combined point estimate, but it does not by itself provide a complete account of uncertainty.