SMC-FCS for Product Terms in Regression

Why a Product Term Changes the Imputation

Suppose a regression relates an outcome to an incomplete covariate \(X\), an observed covariate \(Z\), and their product \(XZ\). The interaction means that the association between \(X\) and the outcome depends on \(Z\). Consequently, the information about a missing \(X\) supplied by the observed outcome also depends on \(Z\). An imputation model that does not respect this structure can draw plausible values of \(X\) while failing to preserve the substantive regression relationship.

A passive product term is still needed: after imputing \(X\), calculate \(XZ\) from the completed \(X\) and observed \(Z\). But recalculating the product does not by itself determine a suitable distribution for missing \(X\). SMC-FCS addresses that second problem by constructing the covariate draw so it is compatible with the specified substantive model. The focus here is the case of one incomplete continuous covariate and one observed interacting covariate in a linear regression.

The proportionality is with respect to \(x\); its normalizing constant makes the density integrate to one. This expression is the conditional distribution used to generate a missing covariate value under the two stated models. It is not enough to draw from \(f(x\mid z)\) alone, because that omits the outcome information, nor is it enough to choose a separate imputation model without considering the substantive likelihood.

Deriving the Interaction-Aware Conditional Distribution

For the normal models in the definition, the target density can be calculated exactly. Write \(c(z)=\beta_0+\beta_Zz\), \(a(z)=\beta_X+\beta_{XZ}z\), and \(\mu_X(z)\) for the mean of \(X\mid Z=z\). Ignoring terms that do not depend on \(x\), the log target density is

Collecting the terms involving \(x^2\) and \(x\) gives a quadratic kernel. Completing the square shows that the conditional distribution is normal. Its variance is the reciprocal of the coefficient multiplying \(x^2\) in the corresponding precision form, and its mean is that variance times the coefficient of \(x\):

The calculation makes the interaction’s role explicit: \(a(z)=\beta_X+\beta_{XZ}z\) is the outcome slope for \(X\) at the observed value \(z\). It affects both the conditional mean and the conditional variance of the imputed \(X\). If \(a(z)\) is near zero, the outcome contributes relatively little information about \(X\); if its magnitude is large, the outcome likelihood is more informative, all else equal.

What the Interaction Does to the Draw

The interaction affects more than the derived column that will eventually be analyzed. It changes how much a particular observed outcome should shift the imputed \(X\). The next calculation isolates that effect. Suppose \(X\mid Z\sim N(0,1)\), and the hypothetical substantive model is \(Y=(1+Z)X+\epsilon\), where \(\epsilon\sim N(0,1)\). This model has an \(X\) main effect of 1, a product coefficient of 1, and no intercept or \(Z\) main effect.

The calculation assumes the outcome model has the stated linear mean and constant residual variance, and that the normal covariate model is appropriate at both values of \(Z\). In applications, the conditional distributions may need to be checked over the observed range of \(Z\); a formula can be evaluated even where the data provide little support, but such extrapolation is not made reliable by the formula.

Generating the Product in Each Completed Data Set

SMC-FCS imputes the incomplete covariate, not an independently varying copy of the interaction term. Once a missing \(X\) has been drawn, the product is determined by the completed \(X\) and the observed \(Z\). Keeping this distinction prevents a completed data set from containing a product value inconsistent with its own covariates.

Assumptions, Diagnostics, and Alternatives

The derivation depends on a model for the outcome conditional on the covariates and a model for the incomplete covariate conditional on the observed covariates. In the worked normal case, these models specify a linear mean for \(Y\), normally distributed errors with variance \(\sigma^2\), and a normal conditional distribution for \(X\) with variance \(\tau^2\). The simplified setup also assumes independent records. In practice, the substantive model, covariate model, and any additional predictors needed to make the missingness assumption plausible must be chosen for the application.

The missingness mechanism is commonly assumed to be missing at random conditional on variables included in the analysis or imputation process, together with distinct model parameters. SMC-FCS does not make this assumption true; it works under the specified missingness and model assumptions. If important predictors of missingness or of \(X\) are omitted, or if the outcome mean, error distribution, or covariate distribution is badly misspecified, the resulting imputations can still be biased or poorly calibrated. There must also be adequate data support across relevant combinations of \(X\) and \(Z\); an interaction estimated in a sparse region can be unstable.

In an actual iterative procedure, model parameters are not generally known. They are estimated or drawn as part of the imputation process so that uncertainty in those parameters is represented across completed data sets. The formulas above show the covariate update conditional on specified parameter values; repeatedly inserting a single set of fitted values and using only conditional means would not provide proper stochastic imputations.

Useful diagnostics include checking that the product is exactly the product of the completed variables, inspecting the distribution of imputed \(X\) across values of \(Z\), and comparing observed and imputed covariate distributions where comparisons are informative. Also examine whether conditional draws are implausible, concentrated at limits, or unusually variable in regions with little data. These checks can reveal coding errors, poor covariate-model fit, or inadequate overlap. They do not prove that the models are correct, especially because missing values cannot be directly checked against their unobserved truth.

A common mistake is to reason, “I have passively recalculated \(XZ\), so the imputation preserves the interaction.” Recalculation preserves the algebraic identity, but it does not make the distribution of \(X\) conditional on \(Y\) and \(Z\) agree with the substantive model. The correction is to use the substantive likelihood when forming the conditional target for missing \(X\), then recalculate \(XZ\) from each resulting draw. Another mistake is to treat the conditional mean as the imputed value; doing so suppresses individual-level imputation variation and fails to represent uncertainty.

The normal closed-form update is convenient but limited to its stated model structure. It should not be carried over unchanged when the outcome model has a nonlinear effect, a non-normal outcome, a categorical incomplete covariate, or several incomplete covariates. Those settings require appropriate substantive and covariate models and may require a different method for drawing from the target distribution. The central principle remains: the conditional draw for an incomplete covariate must reflect the substantive model, and every deterministic product must be regenerated from the completed inputs.