Two Models, Two Different Jobs
In a MICE analysis, a regression model may appear both in the scientific analysis and in the imputation procedure. The shared use of regression does not make these the same model. The analysis model defines the scientific relationship to be estimated. An imputation model describes a variable with missing values conditionally on information available for the cases being imputed. Its job is to generate plausible values while representing uncertainty about those values.
This distinction matters especially when a predictor is missing. The analysis may describe an outcome given that predictor, while the imputation step needs a model for the predictor given the other recorded information. The response variable, predictors, coefficient meanings, and even the direction of conditioning can therefore differ.
A Regression Analysis With a Missing Predictor
Suppose the scientific question concerns the association between a continuous outcome \(Y\) and a continuous exposure \(X\), adjusting for a recorded covariate \(Z\). A possible analysis model is
Here \(\beta_1\) is the adjusted mean difference in \(Y\) associated with a one-unit increase in \(X\), holding \(Z\) fixed, under this linear model. Its units are outcome units per unit of \(X\). The analysis model is chosen to represent the target and the intended scientific analysis; it is not automatically the model used to fill in \(X\).
If \(X\) is missing for some cases, an imputation model must instead describe \(X\) conditionally on available information. For example, with an auxiliary variable \(W\), one might use
This equation has \(X\) on the left, because the quantity being imputed is \(X\). Its coefficients describe conditional prediction of \(X\), not the adjusted association of \(X\) with \(Y\) in the analysis. In a MICE iteration, model parameters are drawn to reflect their uncertainty, and missing \(X\) values are drawn conditionally using the fitted model. The resulting completed data sets are each analyzed using the scientific analysis model.
Worked Example: Naming the Models and Their Coefficients
Consider a hypothetical study measuring \(Y\), a symptom score in points, \(X\), a treatment exposure in days, and \(Z\), age in years. Some exposure values are missing. The scientific target is the adjusted association of exposure with symptom score. Assume the analysis relationship is adequately represented by a linear conditional mean, and assume the chosen imputation model is an adequate conditional description of exposure given its predictors.
The analysis model is \(Y_i=\beta_0+\beta_1X_i+\beta_2Z_i+\varepsilon_i\). If \(\beta_1=0.25\), its interpretation is an estimated 0.25-point difference in mean symptom score per additional exposure day, at a fixed age. The units are points per day.
Suppose the imputation model for missing exposure is \(X_i=\alpha_0+\alpha_1Y_i+\alpha_2Z_i+u_i\). If \(\alpha_1=0.40\), this coefficient describes a 0.40-day difference in the conditional mean exposure per additional symptom-score point, at a fixed age, under that imputation model. Its units are days per point.
The numbers \(0.25\) and \(0.40\) are not competing estimates of one parameter: they belong to different conditional models and have different units. The first addresses the scientific analysis target. The second contributes to predicting missing exposure values.
What the Imputation Step Actually Does
The imputation model is more than a formula for a predicted value. To represent uncertainty, an appropriate MICE procedure draws model parameters and then draws missing values conditional on those parameters and the available predictors. For a continuous predictor under a normal linear imputation model, this involves uncertainty both in the fitted regression and in the residual variation. A fitted conditional mean alone is not a completed-data draw.
The models must be considered together as parts of a specified analysis plan, but they are not interchangeable. The analysis model states which relationship will be reported. The imputation model supplies conditional distributions needed to fill in missing entries. Variables that are not analysis predictors can still be useful in imputation, for example if they predict the missing variable or its missingness. Conversely, the analysis formula does not by itself specify a complete imputation procedure.
Standard multiple-imputation inference relies on assumptions that include an adequately specified imputation model and a missingness mechanism that is ignorable under the variables included in the imputation procedure, commonly expressed as missing at random conditional on observed information. There must also be sufficient observed-data support for the conditional predictions in the regions where values are missing. These assumptions are not established merely by obtaining stable chains or plausible-looking draws. They require substantive assessment; diagnostics can reveal problems but cannot prove the missingness assumption.
Worked Example: Drawing a Missing Exposure Value
Suppose a hypothetical imputation-model fit for exposure \(X\), measured in days, is \(X_i=\alpha_0+\alpha_1Y_i+\alpha_2Z_i+\alpha_3W_i+u_i\). For one illustrative parameter draw, let the coefficients be \(1.0\) days, \(0.40\) days per symptom point, \(0.05\) days per year, and \(0.50\) days per unit of \(W\), respectively. Let the residual standard deviation be \(1.2\) days. Assume this linear conditional model is a reasonable approximation for the missing case and that the case's predictor values lie within the range supported by observed cases.
For a case with symptom score \(Y=10\), age \(Z=50\), and auxiliary value \(W=2\), the conditional mean under this parameter draw is
This is the conditional mean, not the imputed value itself. If the illustrative residual draw is \(u=0.7\) days, the imputed value is \(8.5+0.7=9.2\) days. In actual multiple imputation, parameter uncertainty is also represented through parameter draws; the displayed arithmetic isolates one such draw to make the roles clear.
The completed data set, including this drawn exposure, is then analyzed using the specified outcome regression \(Y\) on \(X\) and \(Z\). The imputation coefficient \(0.40\) is not reported as the treatment-exposure effect. It was used to generate a plausible missing \(X\).
Why Reversing the Analysis Regression Is Not Enough
A frequent mistake is to take a fitted analysis regression of \(Y\) on \(X\), rearrange it algebraically, and use the result to impute \(X\). The rearrangement does not generally give the conditional distribution of \(X\) given \(Y\). The distinction can be seen directly from the error term.
Suppose the analysis relationship is \(Y=a+bX+\varepsilon\), where \(E(\varepsilon\mid X)=0\). Rearranging gives
This identity does not imply that \(E(X\mid Y)=(Y-a)/b\). Taking the conditional expectation given \(Y\) yields
The analysis assumption \(E(\varepsilon\mid X)=0\) does not generally imply \(E(\varepsilon\mid Y)=0\). Therefore, simply reversing the analysis line need not give the conditional mean needed for imputation, and it does not supply the residual distribution or parameter uncertainty needed to generate draws.
Worked Example: The Naive Inverse Gives a Different Prediction
Suppose a hypothetical outcome analysis is \(Y=2+3X+\varepsilon\), with \(Y\) in score points and \(X\) in exposure units. For a case with \(Y=14\), the naive inverse gives
Now suppose a separately fitted conditional imputation model for \(X\) given \(Y\) and other predictors has conditional mean \(3.6\) units for this case and conditional residual standard deviation \(1.0\) unit. Assume that this conditional model is suitable for the case and that its fitted uncertainty is also represented in the imputation procedure. Its mean prediction is \(3.6\), not \(4\); a residual draw of \(0.5\) would give an illustrative imputed value of \(4.1\) units.
The difference is not an arithmetic inconsistency. The inverse calculation treats the analysis regression as though it were a deterministic equation and ignores the conditional distribution of \(X\) given \(Y\). The imputation model is fitted for that conditional prediction task. It must be assessed on its own terms rather than obtained by reversing the analysis equation.
Practical Checks and Common Failure Modes
Keep the two roles explicit when specifying and reviewing an analysis. For each model, record the response, predictors, target, and coefficient units. This simple bookkeeping makes it easier to detect a model that has been copied into the wrong role or interpreted with the wrong direction of conditioning.
- Check the response of each conditional model. For a missing \(X\), the imputation model must generate \(X\), not generate \(Y\) and then treat the fitted outcome as a replacement for \(X\).
- Check that uncertainty is represented. Replacing every missing value by a fitted conditional mean is single imputation, not proper multiple imputation, and generally fails to represent uncertainty from missing values.
- Check plausible support and fit. Implausible imputed values, weak overlap between observed and missing cases, or poorly supported predictor combinations can signal extrapolation or a misspecified conditional model. Distributional and association diagnostics can identify warning signs, but a good-looking diagnostic does not verify all modeling assumptions.
- Check interpretation and units. An imputation coefficient for \(X\) conditional on \(Y\) is not the scientific coefficient for \(Y\) conditional on \(X\). Reporting one as the other reverses both the conditioning statement and often the units.
Complete-case analysis avoids specifying imputation models but discards cases with missing analysis variables and can target a biased association when the observed cases are selected in a relevant way. Single imputation may produce a complete table but does not carry missing-value uncertainty through to inference. Multiple imputation instead repeats the analysis over multiple completed data sets and combines the resulting estimates and uncertainty. None of these approaches removes the need to define the scientific target or assess assumptions.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- An analysis model relates blood pressure \(Y\) in mmHg to a dose \(X\) in mg. What does its slope describe, and what does the slope in a model for missing \(X\) given \(Y\) describe?
- A researcher rearranges a fitted regression of \(Y\) on \(X\) and uses the result as the imputed \(X\) for every missing case. Identify two problems with this procedure.
- A missing predictor's imputation model has a fitted conditional mean of 12 units. Is 12 automatically the imputed value in a multiple-imputation draw? Explain.
- Why can an auxiliary variable be included in an imputation model even if it is not in the scientific analysis formula?
Solutions. (1) The analysis slope describes the adjusted mean difference in blood pressure per mg increase in dose, with units of mmHg per mg. A slope in the imputation model describes the conditional change in the predicted dose per unit of blood pressure, with units of mg per mmHg, given its other predictors. (2) The algebraic inverse is not generally the conditional model for \(X\) given \(Y\), and using a fitted value alone omits residual and parameter uncertainty. (3) No. It is the conditional mean for the fitted parameters; a proper draw also represents model uncertainty and residual variation. (4) An auxiliary variable may help predict the missing predictor or explain observed information relevant to missingness, improving the conditional imputation model even when it is not part of the stated scientific target.