Congeniality Between Imputation and Analysis

Why Congeniality Matters

Multiple imputation separates two tasks. The imputer specifies how missing values are predicted from observed information and generates completed data sets. The analyst applies a chosen model to each completed data set and combines the resulting estimates. These tasks need not use identical formulas, but they must work together to support the inferential target.

Congeniality concerns that relationship. If the imputer’s probability model does not support the analyst’s model and inferential procedure, the completed data may still look plausible while the pooled regression estimate or its uncertainty is misleading. Congeniality is therefore a condition about the models behind inference, not a visual diagnostic or a guarantee that every imputed value resembles an observed value.

Definition: Let \(Y_{\mathrm{obs}}\) denote observed data and \(Y_{\mathrm{mis}}\) denote missing data. The imputer’s model specifies a predictive distribution for \(Y_{\mathrm{mis}}\) given \(Y_{\mathrm{obs}}\), often by integrating over uncertain imputation-model parameters. The analyst’s model specifies a likelihood for the complete data and an inferential procedure for a target parameter or estimand. In a formal Bayesian characterization, the procedures are congenial when there is a coherent joint model that yields the imputer’s predictive distribution and also yields the analyst’s inference for the target. In routine applications, this is an ideal that motivates checking whether the imputation procedure represents the structure and uncertainty needed by the analysis.

This definition is broader than asking whether the imputer and analyst write down the same equation. The imputer may include additional variables that help predict missing values, while the analyst may focus on a particular regression coefficient. Conversely, using the same named variables does not establish congeniality if the implied distributions or uncertainty calculations do not support the target inference.

The formal characterization assumes that both procedures can be understood within a coherent probability model. In practical MICE applications, conditional models are often specified separately; as established in the earlier material on fully conditional models, such conditionals need not correspond to any joint distribution. In that situation, exact Bayesian congeniality may not be available as a literal description. The applied question remains whether the imputation procedure adequately represents the analysis-relevant relationships and uncertainty.

A Linear Regression Model and Its Imputation Conditional

A simple normal regression illustrates how one model can imply a suitable conditional distribution for imputing a missing predictor. Let \(X\) be a continuous predictor and \(Y\) a continuous outcome. Suppose the complete-data model is

X \sim N(0,1), \qquad Y\mid X=x \sim N(2+3x,4).

Here the outcome has a linear mean in \(X\), with residual variance \(4\). The regression coefficient of interest is \(3\), measured in units of \(Y\) per unit of \(X\). If \(X\) is missing for some individuals but \(Y\) is observed, the joint model implies a predictive distribution for that missing \(X\) conditional on \(Y\).

Result: Under a bivariate normal model, the conditional distribution of \(X\) given \(Y=y\) is normal. Its mean is \(\mu_X+\frac{\operatorname{Cov}(X,Y)}{\operatorname{Var}(Y)}(y-\mu_Y)\), and its variance is \(\operatorname{Var}(X)-\frac{\operatorname{Cov}(X,Y)^2}{\operatorname{Var}(Y)}\).

The result follows by completing the square in the joint normal density. In this regression model, \(E(Y)=2\), \(\operatorname{Var}(X)=1\), \(\operatorname{Cov}(X,Y)=3\), and \(\operatorname{Var}(Y)=3^2(1)+4=13\). Substitution gives \(X\mid Y=y\sim N\left(\frac{3}{13}(y-2),\frac{4}{13}\right)\). Thus the regression model does not imply that missing \(X\) should be imputed independently of the observed outcome.

Worked Example: Deriving an Imputation Distribution

Suppose an individual has \(Y=5\), while \(X\) is missing. Use the stated joint model to calculate the conditional mean and variance for imputing \(X\). Treat the model parameters as known for this calculation.

The conditional mean is

E(X\mid Y=5)=\frac{3}{13}(5-2)=\frac{9}{13}=0.6923.

The conditional variance is

\operatorname{Var}(X\mid Y=5)=1-\frac{3^2}{13} =1-\frac{9}{13}=\frac{4}{13}=0.3077.

Therefore the missing predictor should be drawn from \(N(0.6923,0.3077)\), where the second argument is the variance. This calculation checks the model’s implied conditional distribution; it is not itself a complete multiple-imputation procedure, because it treats all parameters as known.

The example also clarifies the meaning of alignment. If the analyst’s complete-data model is the stated normal regression, an imputer using its implied \(X\mid Y\) distribution is working from the same joint structure. Under a fully specified Bayesian procedure, parameter uncertainty would also be represented through posterior draws. The imputer’s conditional distribution and the analyst’s regression inference could then arise from one coherent model, meeting the formal idea of congeniality.

Parameter Uncertainty Is Part of the Inferential Agreement

A frequent mistake is to regard the conditional mean as the whole imputation model. Replacing each missing value with a predicted mean suppresses individual-level uncertainty. Even drawing values from a conditional distribution with fixed estimated parameters can understate uncertainty, because the parameters used to form that distribution are themselves estimated.

In proper multiple imputation, the completed data sets represent both uncertainty about missing values and uncertainty in the imputation model, typically by drawing model parameters and then drawing missing values conditional on those draws. This distinction matters because the analyst’s standard errors and confidence intervals are intended to reflect uncertainty from incomplete data, not just variation among observed cases.

Worked Example: Why Fixed-Parameter Draws Can Miss Uncertainty

Continue with the normal regression example, but suppose the slope \(3\) is estimated rather than known. For a simple illustration, let its estimated standard error be \(0.50\). An imputer that always uses the fixed estimate \(3\) incorporates conditional variation in \(X\) but does not represent uncertainty in the estimated slope.

A parameter-drawing approach might, in a simplified approximation, draw a slope \(b\) from \(N(3,0.50^2)\), then use that draw when generating missing values. The variance of \(b\) is \(0.50^2=0.25\), so the imputer now allows the relationship used to predict missing \(X\) to vary across imputations. This is not a calculation of the exact additional variance of \(X\mid Y\); that would depend on the full parameterization and posterior distribution. It does show what is absent when the slope is held fixed.

The implication is inferential: imputations generated with fixed fitted parameters can be too similar, producing too little between-imputation variation. The analyst may then obtain intervals that are too narrow. Drawing parameters is not a universal cure for model misspecification, but it is part of representing the uncertainty required by a proper imputation procedure.

Implications for Regression Inference

For each of \(m\) completed data sets, the analyst obtains an estimate \(Q_i\) of the target regression coefficient and an estimated within-imputation variance \(U_i\). Define the pooled estimate \(\bar{Q}\), average within-imputation variance \(\bar{U}\), between-imputation variance \(B\), and total variance \(T\) by

\bar{Q}=\frac{1}{m}\sum_{i=1}^{m}Q_i,\qquad \bar{U}=\frac{1}{m}\sum_{i=1}^{m}U_i,\qquad B=\frac{1}{m-1}\sum_{i=1}^{m}(Q_i-\bar{Q})^2,

T=\bar{U}+\left(1+\frac{1}{m}\right)B,\qquad SE(\bar{Q})=\sqrt{T}.

The first part of \(T\) reflects average sampling variance within completed data sets. The second reflects variation in the estimates across imputations, with a finite-\(m\) adjustment. These formulas are meaningful only when the imputation and analysis procedures jointly support the target inference. Congeniality provides a principled setting for that interpretation; it does not make an incorrect regression model correct.

Worked Example: Pooling a Regression Slope

In a hypothetical study, the target is a regression slope measured in outcome units per predictor unit. Five imputed data sets yield slope estimates \(1.8, 2.0, 1.9, 2.1,\) and \(2.2\). Suppose each fitted regression reports within-imputation variance \(U_i=0.16\).

The pooled estimate is

\bar{Q}=\frac{1.8+2.0+1.9+2.1+2.2}{5}=2.00.

The mean within-imputation variance is \(\bar{U}=0.16\). The deviations from \(2.00\) are \(-0.20,0, -0.10,0.10,0.20\), so

B=\frac{0.20^2+0^2+0.10^2+0.10^2+0.20^2}{5-1} =\frac{0.10}{4}=0.025.

Thus

T=0.16+\left(1+\frac{1}{5}\right)(0.025)=0.190, \qquad SE(\bar{Q})=\sqrt{0.190}=0.4359.

For a large-sample illustration, the usual Rubin degrees-of-freedom approximation is \(\nu=(m-1)\left[1+\bar{U}/\{(1+1/m)B\}\right]^2\). Here \(\nu=4[1+0.16/(1.2\times0.025)]^2=160.4\). A 95% interval using the corresponding \(t\) critical value, approximately \(1.975\), is \(2.00\pm1.975(0.4359)\), or approximately \((1.14,2.86)\). It represents uncertainty about the regression slope, in outcome units per predictor unit, under the stated imputation and analysis assumptions. It does not mean there is a 95% probability that the fixed slope lies in this particular interval.

Assumptions, Diagnostics, and Alternatives

Congeniality is a model-based condition, not a property that can be established from a single diagnostic plot. The example’s formal interpretation relies on a correctly specified joint normal model, a suitable parameter-uncertainty procedure, and an ignorable missingness mechanism for the inference being made. In the usual likelihood setting, ignorability requires missing at random conditional on observed data and distinct parameters for the data and missingness models. The observed data must also provide support for the relationships needed to predict the missing values.

Practical checks can reveal problems without proving congeniality. Review whether the imputation conditionals preserve the outcome-predictor relationships required by the analysis, whether the imputed values have plausible distributions, and whether regression estimates and standard errors are stable across imputations and reasonable model variations. Earlier material on observed-imputed comparisons and masked-value diagnostics can help identify distributional discrepancies and predictive failures. Passing those checks does not establish that the analyst’s inferential procedure is supported by the imputer’s model.

Warning sign: implausibly similar imputations. If model parameters are treated as fixed or missing values are replaced by predictions, between-imputation variation may be too small.
Warning sign: unstable estimates. Large changes in the pooled coefficient under modest, defensible changes to the imputation model suggest that the target inference depends heavily on modeling choices.
Warning sign: a conditional model that ignores analysis structure. The imputation procedure may fail to preserve a relationship needed to estimate the stated regression target, even if marginal distributions look reasonable.
Warning sign: unsupported extrapolation. Sparse observed data in regions where values are missing make conditional predictions depend strongly on assumptions that the data cannot check.

Alternatives include complete-case analysis, a single imputation, or a different joint or conditional imputation model. Complete-case analysis avoids specifying an imputation distribution but can lose precision and may be biased when its assumptions fail. Single imputation does not represent missing-value uncertainty in the same way as multiple imputation. A different imputation model may better support the target, but changing models does not by itself guarantee congeniality. Each approach has assumptions; none is a substitute for defining the estimand and checking whether the modeling procedure supports its inference.

Common Mistake

Common mistake: “The imputer and analyst both used \(X\) and \(Y\), so their models are congenial.” Sharing variables is not sufficient. The distributions, parameter uncertainty, and inferential target must be mutually supportable. A better question is whether a coherent model could generate the imputer’s predictive distribution and the analyst’s inference for the target. If that cannot be justified, treat the resulting inference cautiously and examine defensible alternative specifications.

Key takeaway: Congeniality links the imputer’s predictive model to the analyst’s inferential model. In regression, that link concerns the relationships used to predict missing data and the uncertainty carried into pooled estimates—not merely whether the same variables appear in both procedures.

Check Your Understanding

Answer each question, then compare your reasoning with the brief solutions below.

In the Gaussian example, what are \(\operatorname{Cov}(X,Y)\) and \(\operatorname{Var}(Y)\), and what do they imply about \(X\mid Y=y\)?
Why does using the same variables in the imputation and analysis procedures not by itself establish congeniality?
For five imputations, if \(B=0\) and \(\bar{U}=0.16\), what does the total-variance formula give? What concern might remain?
Name one assumption used for the formal interpretation of the Gaussian example and one diagnostic that could reveal a practical problem.

Solutions. (1) Since \(Y=2+3X+\epsilon\), with \(\operatorname{Var}(X)=1\) and \(\operatorname{Var}(\epsilon)=4\), \(\operatorname{Cov}(X,Y)=3\) and \(\operatorname{Var}(Y)=9+4=13\). Thus \(X\mid Y=y\) is normal with mean \(3(y-2)/13\) and variance \(4/13\). (2) Variable overlap does not ensure that the predictive distribution, parameter uncertainty, and analyst’s target inference arise from a coherent model. (3) \(T=\bar{U}+(1+1/5)B=0.16\). Zero observed between-imputation variance may signal overly similar imputations or too few imputations; it does not prove that missing-data uncertainty is absent. (4) One assumption is a correctly specified joint normal model; one practical diagnostic is checking imputed distributions and their relationships with observed variables, while recognizing that diagnostics cannot prove congeniality.