Why These Mistakes Matter
Regression imputation is not just a way to fill empty cells. In MICE, each incomplete variable is drawn from a conditional model, and the completed data are then used for an analysis. If that conditional model discards a relationship the analysis relies on, the resulting imputations can weaken or reshape the target relationship. Plausible-looking values and good marginal distributions do not guarantee that the analysis relationship has been preserved.
This tutorial focuses on three recurring errors: leaving the outcome out when imputing an incomplete predictor, omitting interaction terms, and using a linear imputation relationship when the analysis relies on a nonlinear one. The central question is not whether every imputation model must copy the analysis formula mechanically. It is whether the conditional distribution used to impute missing values preserves the relationships needed by the intended analysis.
Compatibility is not a command to put every analysis term into every conditional model in the same algebraic form. The conditional distribution of \(X\) given \(Y\) and \(Z\) implied by an analysis model can have a different functional form from the analysis model for \(Y\) given \(X\) and \(Z\). The examples below make that distinction explicit.
Mistake 1: Omitting the Outcome
When \(X\) is incomplete and the analysis regresses \(Y\) on \(X\), \(Y\) will often carry information about the missing values of \(X\). An imputation model for \(X\) that ignores \(Y\) cannot use that information. Under restrictive special conditions, omitting \(Y\) may do little harm for a particular estimand, but it is not a safe default. The usual practice is to include the outcome and relevant analysis covariates when imputing an incomplete predictor.
Worked Example: Leaving the Outcome Out Under MCAR
Suppose \(X\) is a centered predictor with \(X\sim N(0,1)\), and the outcome follows \(Y=2X+\varepsilon\), where \(\varepsilon\sim N(0,1)\) is independent of \(X\). Assume \(X\) is missing completely at random: its observation indicator \(R\) is independent of \((X,Y,\varepsilon)\), with \(P(R=1)=q=0.60\). Compare imputing missing \(X\) from its marginal distribution with using a model that conditions on \(Y\).
Under the stated model, \(\operatorname{Cov}(X,Y)=2\) and \(\operatorname{Var}(Y)=5\). The conditional distribution is \(X\mid Y\sim N(0.4Y,0.20)\): the conditional mean coefficient is \(\operatorname{Cov}(X,Y)/\operatorname{Var}(Y)=2/5=0.4\), and the conditional variance is \(1-2^2/5=0.20\). Thus an imputation draw conditional on \(Y\) uses the outcome information.
Now instead draw each missing \(X\) independently from its marginal \(N(0,1)\), ignoring \(Y\). Call the completed predictor \(X_c\). Because \(R\) is independent of the data, an observed value contributes the original covariance, while a marginal draw for a missing value has zero covariance with \(Y\):
Both the observed \(X\) and the independent marginal draws have mean zero and variance one, so \(\operatorname{Var}(X_c)=1\). The population least-squares slope of \(Y\) on this completed predictor is therefore \(1.20/1=1.20\), rather than the complete-data slope \(2\). This calculation illustrates how ignoring the outcome can weaken an association. It relies on the explicit MCAR assumption and the specified independent marginal draws; without those assumptions, the missing cases need not have the marginal distribution of \(X\).
This example concerns stochastic draws from a marginal distribution, not deterministic replacement by a mean. The behavior of mean replacement can differ, and neither shortcut should be treated as a general substitute for a suitable conditional model. Including \(Y\) is not, by itself, sufficient: the conditional model may still omit important structure, as the next examples show.
Mistake 2: Dropping an Interaction
Suppose the analysis includes an interaction between \(X\) and \(Z\). Then the relationship between \(Y\) and \(X\) changes with \(Z\). An additive imputation model for \(X\) that gives \(Y\) the same effect at every value of \(Z\) can erase that variation. The right conditional form for imputing \(X\) should be judged against the distribution implied by the analysis model, not chosen by copying terms without checking.
Worked Example: An Interaction Changes the Conditional Imputation Slope
Let \(Z\) be binary, \(X\mid Z\sim N(0,1)\), and let \(Y=(1+2Z)X+\varepsilon\), where \(\varepsilon\sim N(0,1)\) is independent of \((X,Z)\). This is an analysis model with an \(X\times Z\) interaction: the coefficient of \(X\) is 1 when \(Z=0\), and 3 when \(Z=1\). Derive the conditional mean of \(X\) for imputing missing \(X\).
For a fixed value \(Z=z\), write \(b(z)=1+2z\). The model is \(Y=b(z)X+\varepsilon\), with prior \(X\mid Z=z\sim N(0,1)\) and error variance 1. The normal conditional distribution has mean \(b(z)Y/[1+b(z)^2]\) and variance \(1/[1+b(z)^2]\). Thus its mean coefficient for \(Y\) is:
An additive imputation regression \(X\) on \(Y\) and \(Z\) imposes one common coefficient for \(Y\). It cannot reproduce the two conditional mean slopes, 0.50 and 0.30. Because \(Z\) is binary here, including \(Y\times Z\) permits a separate \(Y\) slope in each stratum and can represent the derived conditional mean. For continuous \(Z\), the coefficient \(b(z)/[1+b(z)^2]\) is generally a nonlinear function of \(z\); merely adding \(Y\times Z\) need not be exact. The calculations depend on the stated normality, independence, and variance assumptions.
The example also explains why “include the interaction” is a useful warning but not a complete recipe. The analysis relationship and the conditional imputation relationship point in related directions, yet their functional forms need not match term for term. A suitable conditional model may need stratification, flexible terms, or a method that directly respects the substantive model.
Mistake 3: Using a Linear Model for a Nonlinear Relationship
A linear regression of \(X\) on \(Y\) can have a zero slope even when \(Y\) is strongly informative about \(X\). This occurs when the relevant information concerns a nonlinear feature, such as the magnitude of \(X\), rather than its sign. In such cases, a linear conditional mean is not a sufficient description of the information needed for the analysis.
Worked Example: A Quadratic Analysis Relationship
Let \(X\sim N(0,1)\), and let \(Y=X^2+\varepsilon\), where \(\varepsilon\sim N(0,0.25)\) is independent of \(X\). Suppose the intended analysis uses the quadratic relationship between \(X\) and \(Y\), and some values of \(X\) are missing. Check what a linear regression of \(X\) on \(Y\) captures.
The joint distribution is symmetric under replacing \(X\) with \(-X\), while \(Y\) is unchanged by that replacement. Therefore \(E(X\mid Y)=0\), and the population linear regression slope of \(X\) on \(Y\) is zero because \(\operatorname{Cov}(X,Y)=0\). A linear conditional-mean model can therefore suggest that \(Y\) carries no information about \(X\).
But the relevant magnitude, \(X^2\), is informative about \(Y\). For a standard normal variable, \(E(X^2)=1\) and \(\operatorname{Var}(X^2)=2\). Independence and the zero mean of \(\varepsilon\) give:
Also, \(\operatorname{Var}(Y)=2+0.25=2.25\), so the linear regression slope of \(Y\) on \(X^2\) is \(2/2.25\approx0.8889\). The outcome is strongly associated with the squared predictor even though its linear association with the signed predictor is zero. Here \(X^2\) is nonconstant, and the moments used in the covariance calculation exist. A marginal draw of \(X\) independent of \(Y\) would lose the relationship between \(Y\) and the magnitude of \(X\).
Symmetry alone does not guarantee this magnitude information: if \(X\) were restricted to \(+1\) and \(-1\), then \(X^2\) would be constant. The example works because \(X^2\) varies. More generally, for \(Y=X^2+\varepsilon\) with independent mean-zero error and finite second moments, \(\operatorname{Cov}(X^2,Y)=\operatorname{Var}(X^2)\); this is positive when \(X^2\) is nonconstant.
How to Review a Regression-Imputation Plan
A compatibility review should start with the intended analysis and then examine each conditional imputation model. The purpose is to find missing information or forced functional forms, not to add terms indiscriminately. A practical review follows this sequence:
Name the outcome, incomplete predictors, covariates, interactions, and nonlinear terms used by the analysis.
Check whether it includes the outcome and relevant covariates, and whether it allows the conditional relationship to vary as required.
Do not assume the imputation mean must copy the analysis formula. Derive or reason about the form of \(X\mid Y,Z\), including its dependence on \(Z\).
Compare observed and imputed values across outcome and covariate ranges, including strata or nonlinear features that matter to the analysis.
Useful diagnostics include plotting observed and imputed \(X\) against \(Y\) within levels or ranges of \(Z\), checking whether interactions appear weakened, and comparing distributions of analysis predictors and derived terms across imputations. A fitted-model summary alone is not enough: similar marginal distributions can conceal a changed conditional relationship. Sparse data, separation, extrapolation beyond observed combinations, and overly flexible models can also make the imputation unstable or implausible.
| Approach | What it can address | Main limitation |
|---|---|---|
| Add outcome and analysis covariates to the conditional model | Uses outcome information for missing predictors | Does not automatically handle interactions or nonlinear forms |
| Add interaction or nonlinear terms | Allows selected relationships to vary or curve | A convenient term set may still not match the implied conditional distribution |
| Stratify or use flexible conditional models | Can represent varied relationships with fewer functional restrictions | May be unstable with sparse strata or overfit limited data |
| Use a substantive-model-compatible approach | Can build the substantive analysis structure into imputation more directly | Requires additional modeling choices and careful implementation |
Common Pitfalls and Limits
- “The outcome is not a predictor, so it should be excluded.” The analysis may predict \(Y\) from \(X\), but the imputation task predicts missing \(X\). Conditioning on \(Y\) can be important for preserving their association.
- “I included \(Y\), so the model is compatible.” Including the outcome does not fix an omitted interaction, nonlinear relationship, or other misspecified conditional form.
- “The analysis has \(X\times Z\), so adding \(Y\times Z\) always solves imputation.” This may work in a particular setting, such as the binary-\(Z\) example, but the implied conditional relationship can be nonlinear in \(Z\).
- “A zero linear slope means no useful information.” The relationship may concern \(X^2\), a threshold, or another feature; inspect the analysis-relevant structure.
- “A good-looking observed-versus-imputed plot proves compatibility.” Diagnostics can reveal problems but cannot establish that all assumptions or the missing-data mechanism are correct.
No finite diagnostic can guarantee compatibility, and no adjustment can compensate for information that the data and assumptions do not support. Model complexity must be balanced against data sparsity and stability. The goal is a defensible conditional model that retains the structure needed by the analysis, with its assumptions made explicit.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- An incomplete predictor \(X\) is used in a regression of \(Y\) on \(X\) and \(Z\). Why can omitting \(Y\) from the imputation model for \(X\) be problematic?
- In the interaction example, what are the conditional-mean slopes of \(X\) on \(Y\) at \(Z=0\) and \(Z=1\)?
- For \(Y=X^2+\varepsilon\), what additional condition beyond symmetry supports the claim that \(Y\) carries information about \(|X|\)?
- A linear regression of \(X\) on \(Y\) has slope zero. Does that rule out an association relevant to an analysis using \(X^2\)? Explain.
Solutions. (1) The outcome can contain information about missing \(X\); omitting it prevents the imputation from preserving their association, as the MCAR numerical example demonstrates for marginal stochastic draws. (2) The slopes are \(1/(1+1^2)=0.50\) and \(3/(1+3^2)=0.30\), so a common additive \(Y\) slope cannot match both. (3) \(X^2\) must be nonconstant, with the relevant moments finite; then \(\operatorname{Cov}(X^2,Y)=\operatorname{Var}(X^2)>0\). (4) No. Symmetry can make the signed linear association zero while \(X^2\) remains associated with \(Y\), so the analysis-relevant nonlinear relationship must be checked directly.