Why the Outcome Belongs in Predictor Imputation

Why the Outcome Helps Impute a Predictor

Suppose the scientific analysis relates an outcome \(Y\) to a predictor \(X\), but some values of \(X\) are missing. For a record with an observed outcome, that outcome contains information about the plausible value of its missing predictor. An imputation model for \(X\) that ignores \(Y\) discards that information. More importantly, it may fail to preserve the relationship between \(X\) and \(Y\) that the analysis is intended to estimate.

In MICE, each incomplete variable is imputed in turn using a conditional model. When imputing \(X\), the predictors of that conditional model should generally include the observed outcome and the other variables needed to represent the analysis. This does not mean fitting the analysis model backwards and treating it as the imputation model: the imputation model predicts \(X\) given \(Y\), while the analysis model describes \(Y\) given \(X\).

The reason is conditional prediction. If people with larger outcomes tend to have larger values of \(X\), then knowing a person's observed \(Y\) shifts the distribution of that person's missing \(X\). A model that ignores \(Y\) instead draws from a distribution that does not reflect this person-specific information. Including \(Y\) is generally important for maintaining the association of interest, but it does not by itself ensure a valid imputation model. The conditional model must still be suitable for the data and analysis.

A Conditional-Normal Justification

A clear derivation is available when \(X\) and \(Y\) have a bivariate normal distribution. Write their means as \(\mu_X\) and \(\mu_Y\), their standard deviations as \(\sigma_X\) and \(\sigma_Y\), and their correlation as \(\rho\). Under this assumption, the conditional distribution of \(X\) given \(Y=y\) is normal. Its mean depends on \(y\), and its variance is smaller than the marginal variance of \(X\), unless \(X\) and \(Y\) are uncorrelated.

To justify this representation, define \(\varepsilon=X-\mu_X-\beta(Y-\mu_Y)\). Its covariance with \(Y\) is \(\operatorname{Cov}(X,Y)-\beta\operatorname{Var}(Y)=0\), by the definition of \(\beta\). Because \(X\) and \(Y\) are jointly normal, \(\varepsilon\) and \(Y\) are jointly normal; for jointly normal variables, zero covariance implies independence. Thus the conditional distribution of \(X\) given \(Y=y\) has mean \(\mu_X+\beta(y-\mu_Y)\) and variance \(\operatorname{Var}(\varepsilon)\). Expanding that variance gives:

Consequently, the conditional mean changes with the observed outcome whenever \(\rho\ne 0\), and conditioning reduces uncertainty from \(\sigma_X^2\) to \(\sigma_X^2(1-\rho^2)\). These formulas rely on joint normality. Outside that setting, the general principle remains that the imputation model should use outcome information, but the conditional form need not be normal or linear.

Using the Outcome in a Regression Imputation Model

In practice, an analysis may adjust for other covariates. If \(Z\) is such a covariate, the conditional model for the missing \(X\) should generally use both \(Y\) and \(Z\). The relevant question is whether \(Y\) adds information about \(X\) after accounting for the other predictors—not merely whether \(X\) and \(Y\) are associated before adjustment.

For a continuous \(X\), a common starting point is a regression model for \(X\) conditional on \(Y\), \(Z\), and any other selected predictors. A linear conditional mean with an appropriate residual distribution may be reasonable, but its adequacy depends on the data. The analysis model and imputation model have different roles: the first estimates the scientific relationship, while the second generates plausible values for missing data.

Applying the Principle in MICE

For an incomplete \(X\), the conditional imputation model should ordinarily include the observed outcome \(Y\), the other variables in the analysis model, and relevant auxiliary predictors that help explain missing values or the missingness process. The selection should preserve the analysis relationship while producing a credible conditional distribution for \(X\). If an outcome value is observed, it is available as a predictor when imputing \(X\); missing-outcome records require additional care and are not covered by the simple calculation above.

The usual justification assumes that the conditional imputation model is correctly specified, or sufficiently well specified for the target analysis; that the variables used as predictors are appropriately handled; and that the missingness assumptions needed for the analysis are defensible. Under a missing-at-random assumption, the probability that \(X\) is missing may depend on observed information such as \(Y\) and \(Z\), after conditioning on the variables in the model. Including \(Y\) can therefore help both represent the conditional distribution of \(X\) and account for its observed association with missingness. It does not make missing-at-random true, and it does not resolve missing-not-at-random mechanisms by itself.

Diagnostics, Alternatives, and Limitations

Including \(Y\) is a general principle, not a guarantee. Diagnose the conditional model as well as the MICE sequence. Useful checks include comparing observed and imputed \(X\) distributions, examining imputed values across ranges of \(Y\) and other predictors, checking whether fitted relationships are plausible, and reviewing convergence and chain behaviour. The earlier material on masked-value diagnostics is relevant to evaluating predictive performance where observed values can be temporarily hidden.

Warning signs include implausible imputations, poor prediction in parts of the outcome range, residual patterns inconsistent with the assumed model, or chains that do not stabilize. A linear conditional mean may be inadequate if the relationship is nonlinear; a normal residual distribution may be unsuitable for a skewed or bounded \(X\). These problems require revising the conditional model, not removing a useful observed outcome by default.

The main alternative is an imputation model that omits \(Y\), for example a model using only \(Z\) or a marginal draw for \(X\). Such an approach may be defensible in particular circumstances, but it should not be chosen merely because \(Y\) is the analysis response. Omitting a predictive outcome can lose information and inadequately represent the relationship targeted by the analysis. Conversely, including \(Y\) in a misspecified conditional model is not automatically safe. The choice must be assessed against the analysis structure and the data.

A common mistake is to reason, “The analysis predicts \(Y\) from \(X\), so the imputation should also predict \(X\) without using \(Y\).” That reverses the conditional distributions: the analysis model concerns \(Y\mid X\), while imputing a missing \(X\) requires a model for \(X\) conditional on observed information, usually including \(Y\). Another mistake is to treat the conditional mean as the completed value. Deterministic mean replacement ignores residual and model uncertainty and can make completed data too concentrated.