Why Product Terms Need Special Handling
An interaction in a regression model is often represented by a product of predictors. If \(X\) and \(Z\) are numerical variables, define \(W=XZ\). When either source variable is incomplete, the product is generally treated as missing too, unless its value is already determined by the available information. For example, if \(Z=0\) is known, then \(XZ=0\) regardless of the value of \(X\).
In multiple imputation by chained equations, the key requirement is that every completed data set obey the definition of the product term. Once an imputed value of \(X\) or \(Z\) changes, the value of \(W\) must be updated from the current values of both factors. Drawing or retaining \(W\) separately can produce completed records for which \(W\ne XZ\), so the analysis would no longer use the stated interaction.
The product definition presumes that multiplication is meaningful for the variables and coding being used. For continuous predictors, \(W\) has units equal to the product of the units of \(X\) and \(Z\). For categorical predictors, an interaction may be constructed from indicator variables, but the precise coding and resulting terms must be specified; the examples here use numerical predictors.
How the Product Is Updated in Chained Equations
A chained-equations procedure draws values for incomplete variables from specified conditional models in turn. A passive product is not another uncertain source variable: it is a deterministic calculation based on the current source values. Accordingly, whenever a draw changes \(X\) or \(Z\), update \(W\) using those current values before using \(W\) in a subsequent imputation model or in the analysis.
Record that \(W=XZ\), and confirm that the coding and units of \(X\) and \(Z\) make this product the term intended for the analysis.
Use the specified conditional imputation models to update incomplete \(X\) and \(Z\), in the chosen chained-equations order.
Set \(W\) equal to the current \(X\) multiplied by the current \(Z\). Keep a product whose value is already determined, such as zero from a known zero factor.
Pass the updated product to any conditional imputation model that includes it, and use it in the analysis model. Repeat the calculation whenever either source changes.
This procedure guarantees algebraic consistency, but not by itself statistical compatibility between the imputation models and the substantive analysis. For example, if the analysis includes an interaction, the conditional models used to impute its source variables must be considered in light of that structure. Including a passively calculated product can help represent the intended relationship, but it does not guarantee that all conditional models jointly match the analysis model.
Why Recalculation Is Necessary
The justification follows directly from the deterministic definition. In any completed data set intended to represent \(W=XZ\), every record must satisfy the equation. Thus, after a source draw produces \(X=x\) and \(Z=z\), the corresponding value is fixed at \(w=xz\); there is no additional random draw for \(W\).
If \(W\) is instead drawn independently, it can differ from the product of its sources. With continuous-valued sources, the equality \(W=XZ\) defines a restricted set of possible triples \((X,Z,W)\). An independently drawn \(W\) will generally not land on that set. The discrepancy is not ordinary imputation uncertainty: it contradicts the variable's definition.
An interaction also changes how the regression relationship is read. In a model with an identity link,
the conditional change in the fitted mean for a one-unit increase in \(X\), holding \(Z\) fixed, is \(\beta_1+\beta_3Z\). Thus the interaction term is not merely a convenient extra column: it allows the association of \(X\) with the outcome to vary with \(Z\). For other link functions, the interpretation is on the scale of the linear predictor unless translated to the response scale.
Worked Example: Recomputing a Product After Imputation
Consider a hypothetical data set with \(W=XZ\). For one record, \(X\) is missing and \(Z=4\). A chained-equations draw imputes \(X=1.5\). For another record, \(X\) is missing but \(Z=0\) is observed. Find the product values where possible.
For the first record, after the draw, the current source values are \(X=1.5\) and \(Z=4\). Therefore,
The product is 6 in the completed record. Before \(X\) was imputed, the product was generally unavailable because the unknown \(X\) could have taken different values.
For the second record, \(X\) remains unknown, but the known factor is \(Z=0\). For every possible value of \(X\),
So \(W\) is already determined as zero; it need not wait for the imputation of \(X\). The source variable \(X\) is still missing and still requires its own imputation. The known product does not supply the missing value of \(X\).
Worked Interaction With Two Incomplete Predictors
Suppose a hypothetical analysis uses two continuous predictors, baseline score \(X\) in points and treatment exposure \(Z\) in hours, and includes \(W=XZ\). Both predictors have missing values for some records. The substantive model is
where \(Y\) is an outcome measured in outcome units. The coefficient values are hypothetical. The product \(W\) has units of point-hours, and \(\beta_3=0.8\) is the change in the fitted mean per one point-hour, conditional on the other terms in this model.
Worked Example: Updating an Interaction After Two Draws
For one hypothetical record, suppose the current imputation iteration draws \(X=2.4\) points and \(Z=3.0\) hours. Calculate \(W\), then calculate the fitted mean under the model above.
First passively update the product:
Substitute the current values into the model:
The interaction contribution is \(5.76\) outcome units. If a later draw changes either source value, 7.2 is no longer automatically valid: calculate the new product from both updated sources before using it again.
The model also shows why the interaction can matter when imputing a source. The fitted-mean change for a one-point increase in \(X\) at \(Z=3.0\) hours is
This is a model-based conditional slope, not a causal effect. Whether it is appropriate depends on the study design and the adequacy of the model.
Checking the Implementation and Recognizing Failure
The most direct diagnostic is an algebraic check in every completed data set: for records with both source values available, calculate \(XZ\) and compare it with the stored \(W\). Their difference should be zero, up to numerical rounding. Also verify that missingness in the stored product follows the chosen deterministic rule: it should ordinarily be unavailable when a source is unknown, but a known zero factor determines a zero product.
- Stale product values: \(W\) was calculated before a source draw and not refreshed afterward. Check the product after each relevant update, not just at the start of an iteration.
- Independent product imputations: \(W\) was assigned values from its own conditional model. This can break the defining identity even if its marginal distribution looks plausible.
- Wrong coding or units: the product was formed from a recoded or scaled variable that is not the source specified by the analysis. Verify the definitions and units of both factors.
- Misleading treatment of zero: software may propagate a missing value through multiplication even when the other factor is known to be zero. Apply the mathematical definition explicitly if that product is genuinely determined to be zero.
- Inadequate imputation structure: exact equality \(W=XZ\) can hold while the conditional imputation models still fail to represent important relationships with the outcome or other variables. Algebraic consistency is necessary, not sufficient.
When checking distributions across imputations, look at the source variables and the product together. A product distribution can change as the source draws change; that variation is expected. The concern is not that \(W\) varies between imputations, but that its values do not correspond to the \(X\) and \(Z\) values in the same completed data set.
Alternatives and Limitations
Passive construction is the direct way to maintain a deterministic product in completed data. It differs from independently imputing \(W\), which treats a value fixed by the sources as if it were a separate uncertain quantity. It also differs from omitting the interaction: omission may simplify imputation but changes the substantive model if that interaction is genuinely part of the scientific analysis.
A more tailored imputation approach may be needed when ordinary chained conditional models do not adequately preserve the substantive model's structure. The earlier material on imputation-model compatibility explains why plausible marginal distributions alone do not guarantee a suitable joint relationship. Passive recomputation solves the deterministic bookkeeping problem; it does not resolve all compatibility problems. In particular, it cannot make misspecified conditional models correct, recover information absent from the observed data, or remove the assumptions needed for valid inference under missingness.
The product rule is also sensitive to how predictors are represented. If \(X\) or \(Z\) is a category represented by indicators, the interaction generally involves products of the relevant indicators, with interpretation tied to the reference coding. If either source is transformed, the term must follow the analysis definition exactly. These issues should be settled before producing completed data, not repaired by changing \(W\) independently afterward.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- For a record with \(X\) missing and observed \(Z=5\), a draw gives \(X=2.2\). What is the updated product \(W=XZ\), and what are its units if \(X\) is measured in kilograms and \(Z\) in hours?
- If \(X\) is missing but \(Z=0\) is observed, is \(W\) missing? Does knowing \(W\) determine \(X\)? Explain.
- A completed record has \(X=3\), \(Z=4\), and stored \(W=11\). What check fails, and what should the product be?
- In \(E(Y\mid X,Z)=\beta_0+\beta_1X+\beta_2Z+\beta_3XZ\), what is the conditional change in fitted mean for a one-unit increase in \(X\) at \(Z=z\)?
Solutions. (1) \(W=(2.2)(5)=11\) kilogram-hours. (2) \(W=X(0)=0\), so the product is determined, but it does not determine \(X\): many possible values of \(X\) yield the same product. (3) The stored product violates the identity because \((3)(4)=12\), not 11; update \(W\) to 12. (4) The conditional change is \(\beta_1+\beta_3z\) units of \(Y\) per one unit of \(X\), holding \(Z=z\) fixed.