Choosing Predictors for Imputation Models

Why Predictor Choice Matters

An imputation model uses observed information to generate plausible values for records with missing data. Its usefulness depends partly on which variables supply that information. If an imputation model leaves out a variable that is informative about the missing values, the generated values may be less accurate, and relationships needed for the planned regression may be weakened or distorted.

Choosing predictors is not simply a matter of including every available column. A candidate should have a defensible role: it may be needed by the planned analysis, related to whether a value is missing, or informative about the variable being imputed. The goal is a set of predictors that preserves relevant information without adding variables that are irrelevant, poorly measured, or unavailable where they are needed.

These roles can overlap. A variable in the planned regression may also predict missingness; an auxiliary variable may be associated with an analysis variable. Predictor choice concerns what information the imputation models should use, not a claim that every selected variable has a causal effect.

Three Reasons to Consider a Variable

Start with the variables required by the planned regression. Omitting an analysis variable from the imputation system can make the completed data fail to preserve the associations the regression is intended to estimate. The variables needed to fit or interpret the analysis should therefore be accounted for in the imputation plan, with the details depending on which variable is being imputed.

Next consider variables that predict missingness. If a measured characteristic helps distinguish records with and without a value, including it can help the imputation model reflect observed differences between those groups. This can support a plausible missing-at-random analysis when the relevant missingness differences are explained by observed information. It does not establish that missingness is at random: that mechanism cannot generally be verified from observed data alone.

Finally, consider auxiliary variables. A measure that is not part of the final regression can still carry useful information about the missing value. For example, a recent measurement might predict a later measurement, or a routinely collected measure might be associated with a value that is sometimes absent. Such variables can improve the information available to imputation and may reduce uncertainty, provided they are measured well and are available in enough records to be useful.

A variable does not become a good predictor merely because it is available. Check whether its meaning and timing make it relevant, whether it is observed for records needing imputation, and whether including it is practical in the imputation procedure. A variable that is itself largely missing may contribute little for records where it is unavailable. A measure collected after the event being studied may also be inappropriate if it would not be known or relevant at the time represented by the analysis.

Use Associations With the Values Being Imputed

Predicting missingness and predicting the missing value are related but distinct tasks. A variable can help explain which records have missing values without carrying much information about the values themselves. Conversely, a variable can predict the value being imputed even if it does not predict whether that value is missing. Examine both roles where the available data allow.

An auxiliary variable is especially promising when it is measured on many records that need imputation and is associated with the variable being imputed. The association should be plausible in context, not just selected because a small exploratory sample happens to show a large difference. Its usefulness also depends on whether the observed association is informative for the records whose values are missing.

No single screening statistic settles predictor choice. Correlations, group summaries, and missingness rates can help identify candidates, but they do not replace reasoning about the planned analysis and the data collection process. Avoid selecting variables solely because they pass a significance threshold in an exploratory test: a weak test result may reflect limited data, while a predictor with a modest association may still preserve an important analysis relationship.

Why Omitting Relevant Information Can Weaken Imputation

A simple conditional-mean argument shows why relevant information can matter. Suppose \(Z\) is observed and helps predict a variable \(X\). For squared prediction error, the best prediction based on \(Z\) is the conditional mean \(E(X\mid Z)\). If the model ignores \(Z\), it may use only the overall mean \(E(X)\), which cannot adapt to differences across values of \(Z\).

Here is the corresponding decomposition. Write \(\mu=E(X)\) and \(m(Z)=E(X\mid Z)\). Conditional on \(Z\), expand the squared error of predicting \(X\) by \(\mu\):

The cross-product has conditional expectation zero because \(E\{X-m(Z)\mid Z\}=0\). Thus

Averaging over \(Z\) gives

The final term is nonnegative. Therefore, using the conditional mean cannot have greater expected squared prediction error than using the overall mean; the improvement is strict when \(E(X\mid Z)\) varies across values of \(Z\). This argument concerns prediction under the stated population quantities. It does not say that every added variable improves a fitted imputation model in a finite sample, or that a conditional-mean prediction alone is a complete imputation. Imputation must also represent uncertainty and variation in plausible values.

Build and Review a Predictor Set

A practical review can be organized into a few explicit questions before fitting the imputation models.

This process is a reasoned construction, not an automatic variable-selection algorithm. A very large predictor set can create practical problems, especially when the sample is limited or many candidate predictors are themselves incomplete. In that situation, prioritize variables with clear analytical or predictive roles and document why they were retained. Do not remove a variable essential to the planned analysis merely because an automated screen gives it a weak score.

What Predictor Choice Cannot Do

Including variables related to missingness or the missing values can make an imputation model more informative, but it does not recover information that was never observed without assumptions. In particular, a predictor set does not prove that missingness is independent of the unseen values after conditioning on observed data. If missingness may depend on the unseen value itself, the observed data alone cannot settle that issue.

Nor does adding more predictors guarantee less bias or better performance. Variables that are measured poorly, unavailable for many incomplete records, or unrelated to the needed values may add little. A selected predictor may also need to be represented appropriately in the imputation model; merely naming it in a predictor list does not ensure the relevant relationship is captured.

A common mistake is to reason, “I included variables associated with missingness, so the data are now MAR and the imputation is unbiased.” The correction is that such variables can help make a MAR assumption more plausible by conditioning on observed information, but they cannot verify that assumption or rule out dependence on unobserved values. State the assumptions, use relevant information thoughtfully, and report the predictor choices clearly.