Missing-Indicator Methods and Their Pitfalls

Why a Missingness Flag Is Not a Replacement Covariate

In a genetic association model, an incomplete covariate can be consequential when it helps control confounding or otherwise defines the adjusted association of interest. A tempting shortcut is to fill in its missing values with a fixed number and add a binary indicator for whether the value was missing. The indicator records that a value is absent; it does not generally recover what the value would have been. Those are different pieces of information.

Let \(G\) denote genotype dosage, \(C\) an incomplete covariate, and \(Y\) an outcome. Define \(R=1\) when \(C\) is observed and \(R=0\) otherwise, and set \(M=1-R\). For simplicity, suppose the placeholder is zero (for example, after centering \(C\)), so the coded covariate is \(C^*=RC\). A common indicator-coded regression includes \(G\), \(C^*\), and \(M\). The missingness flag can allow a different intercept for records with missing \(C\), but in that group \(C^*=0\) for everyone. The fitted model therefore has no information about individual differences in the unobserved values of \(C\).

To see the problem, consider the correctly specified full-data linear mean model \(\mathbb{E}(Y\mid G,C)=\alpha+\beta G+\gamma C\), where \(\beta\) is the genotype association conditional on \(C\), and \(\gamma\) is the covariate effect. If the distribution of \(C\) among records with missing values varies with genotype, then \(C\) can still be associated with \(G\) in precisely the records where its value is unavailable. A flag that is identical for all those records cannot adjust for that variation.

Deriving the Remaining Bias

The issue can be stated precisely for ordinary least squares. Assume the full-data conditional mean is linear as above and the error has conditional mean zero given \(G\), \(C\), and \(M\). Fit a population regression of \(Y\) on an intercept, \(G\), \(C^*\), and \(M\). Let \(\widetilde{G}\) be the residual from projecting \(G\) on the other included predictors: the intercept, \(C^*\), and \(M\). The coefficient of \(G\) in this regression is

This is the omitted-variable bias formula applied after accounting for the predictors that the indicator-coded model actually includes. The full outcome equation contains \(C\), but the fitted model uses \(C^*\), which equals zero rather than \(C\) when \(M=1\). Unless the residualized genotype is uncorrelated with the true covariate, the second term does not vanish. Adding \(M\) removes the part of the missingness pattern represented by \(M\); it does not guarantee that \(\widetilde{G}\) and the unobserved \(C\) are uncorrelated.

What Happens Within the Missing Group

A second view is to compare expected outcomes across genotype values among records with \(M=1\). Under the linear model,

The missingness flag is fixed at one within this group. If the mean covariate value changes with genotype, the outcome difference includes both the genetic effect and a covariate-composition difference. An indicator can shift the fitted intercept for this group, but a constant shift cannot generally remove a difference that varies with genotype.

When Indicator Coding Can Be Useful

Indicator coding is not universally useless. It can serve a limited purpose when the scientific target or the assumptions make the missing values themselves unnecessary for estimating the coefficient of interest. The purpose must be stated clearly: a model that accounts for missingness status is not automatically a model adjusted for the unobserved covariate.

• Missingness is itself a meaningful category. If the variable of interest is whether a measurement was recorded, then “not recorded” is an actual status category. Modeling that status can be appropriate for describing or predicting outcomes by recorded-data status. It does not estimate the association adjusted for the missing underlying value.
• The incomplete covariate does not confound the genotype coefficient under the specified model. In the linear setting above, if \(\operatorname{Cov}(\widetilde{G},C)=0\), the bias term is zero. This is a substantive condition about the relation between genotype and the true covariate after adjustment, not a consequence of adding \(M\). For example, a covariate may be prognostic for \(Y\) but not associated with \(G\) after the other model terms are accounted for.
• A randomized exposure provides a distinct setting. In a randomized study, random assignment can make the exposure independent of baseline covariates by design. That can protect a treatment contrast from baseline-covariate confounding under suitable analysis assumptions, even if an incomplete baseline covariate is not fully adjusted. Genotype in an observational association study is not automatically randomized independently of ancestry or other structure, so this rationale cannot simply be transferred to GWAS.

These are restricted cases, not general guarantees. The zero-covariance condition is tied to the stated linear coefficient and model; it should not be assumed to protect a different estimand or regression scale. Likewise, if missingness status is included as a category for a descriptive purpose, the resulting contrast is conditional on recorded status, not on the covariate value that was never observed.

Choices for an Association Analysis

Before using a missingness indicator, identify the estimand: for example, a genotype association adjusted for the true covariate \(C\), or an association conditional on whether \(C\) was recorded. Then ask whether the indicator-coded regression identifies that target under defensible assumptions. If adjustment for \(C\) is needed and those assumptions are not credible, coding a placeholder and adding a flag is not a substitute for an appropriate missing-data analysis. Multiple imputation is one approach already used for incomplete covariates, although its validity depends on its own model and assumptions.

The indicator procedure also does not automatically fix other features of the analysis. It does not guarantee a correct outcome model, eliminate selection effects, or ensure that the reported standard error is appropriate. The coefficient's units and interpretation still depend on the outcome scale and the model. For example, a linear-regression coefficient is a difference in expected outcome per allele copy, conditional on included terms; it is not, by itself, a causal effect.

Common Mistake

A reliable interpretation therefore names exactly what was modeled. “Adjusted for the observed covariate values and an indicator of missingness” is precise. “Adjusted for the covariate” is justified only when the analysis actually supports adjustment for the underlying covariate under its assumptions, not merely because a flag was added.