Start With the Target, Not the Imputation Algorithm
An imputation model is not an end in itself. Its role is to produce completed data on which a specified analysis can estimate a defined target. If imputation changes the information that distinguishes values in that target, the resulting analysis may no longer estimate what was intended—even if the imputed values look plausible one variable at a time.
The target regression estimand might be a slope, an adjusted contrast, or a collection of coefficients defined by a particular model and population. Before deciding how to impute a variable, identify which distinctions in that variable the estimand uses. A continuous slope uses more than a coarse category; a log-linear effect uses the log scale; a factor with separate contrasts uses more than a single numeric score.
For example, suppose the intended analysis is a regression of a continuous outcome \(Y\) on a continuous predictor \(X\), with other predictors \(Z\):
Here, \(\beta_1\) is the change in the conditional mean of \(Y\) associated with a one-unit increase in \(X\), holding \(Z\) fixed, under the stated linear model. If the imputation process retains only a coarse grouping of \(X\), it may lose distinctions needed to estimate that per-unit slope. The analysis formula and the target must therefore be considered before choosing a representation for imputation.
Why the Representation Matters
A useful way to formalize the issue is to distinguish the original variable \(X\) from the representation \(g(X)\) retained by an imputation model. The analysis may use a design vector \(d(X)\), containing the functions of \(X\) that define its regression terms. For example, \(d(X)\) might contain \(X\), \(\log(X)\), or separate indicators for factor levels.
If the imputation representation is \(g(X)\), then the analysis-relevant design \(d(X)\) can be preserved using that representation alone only if there is a function \(a\) such that \(d(X)=a(g(X))\). In practical terms, two values treated as identical by the imputation representation must not require different analysis contributions.
This condition is necessary when the analysis relies on \(d(X)\). If it fails, information discarded by \(g\) cannot be recovered merely by fitting the final regression after imputation. The condition does not, by itself, establish that the imputation model is correctly specified, that the missingness assumptions hold, or that valid uncertainty is propagated.
Worked Example: Coarsening a Continuous Predictor
Suppose a hypothetical analysis targets the slope of a continuous predictor \(X\) in a linear regression. In a portion of the conditional mean, its contribution is \(\beta_1X\). An imputation representation instead retains only \(C=I(X\geq50)\), where \(I(\cdot)\) is 1 when the condition is true and 0 otherwise. Let \(\beta_1=0.08\) outcome units per unit of \(X\), and consider \(X=51\) and \(X=75\), holding all other terms fixed.
Both values give \(C=1\), so the category-only representation treats them alike. But their linear contributions are:
The difference is \(6.00-4.08=1.92\) outcome units, equivalently \((75-51)(0.08)=24(0.08)=1.92\). Because the values lie on the same side of the threshold, the category does not retain the distinction that defines the continuous slope. If the analysis target is the per-unit linear effect, imputing only the category cannot represent both contributions. This does not mean categorization is always wrong; it means the category alone does not preserve this particular continuous-slope target.
Preserving Functional Form and Category Structure
Preservation applies not only to which variables are included, but also to the functions and distinctions used to define the target. A transformed predictor may be essential to the analysis scale. A categorical predictor may require separate level contrasts rather than a numeric trend. The appropriate representation depends on the estimand, not on whether a particular coding is convenient.
Worked Example: A Log-Scale Predictor
Consider a hypothetical regression in which the conditional mean includes \(4\log(X)\), with \(X>0\). The coefficient 4 describes change in outcome units per one-unit increase in \(\log(X)\), not per one-unit increase in \(X\). Compare \(X=1\) with \(X=2\), holding other terms fixed.
The difference in the analysis contribution is:
Since \(\log(1)=0\) and \(\log(2)\approx0.693147\), direct calculation gives \(4\log(2)\approx2.7725887\), which rounds to 2.773 outcome units. Using the rounded intermediate value gives \(4(0.6931)=2.7724\), a lower-precision approximation. An imputation representation that retains only a linear contribution in \(X\) does not automatically preserve the log-scale relationship. The relevant check is whether the completed data retain the distinctions required by the analysis term \(\log(X)\).
Worked Example: Separate Contrasts for a Categorical Predictor
Suppose a hypothetical analysis compares three treatment groups, A, B, and C, using A as the reference group. At fixed values of the other predictors, assume the target model’s conditional mean is 10 units for A, 12 units for B, and 17 units for C. Thus, the target contrasts relative to A are 2 units for B and 7 units for C.
If an imputation representation replaces group membership by the numeric scores 0, 1, and 2 and uses one linear score effect \(\delta\), it imposes equal successive increments. Matching B to A requires \(\delta=12-10=2\). That choice predicts a C mean of \(10+2(2)=14\), not 17. Conversely, matching C to A requires \(2\delta=17-10=7\), or \(\delta=3.5\); that choice predicts B at \(10+3.5=13.5\), not 12.
No single \(\delta\) matches both target contrasts. The numeric score has imposed a structure that differs from the separate category effects in the target. When those contrasts define the estimand, retaining the categories as distinct levels is necessary; whether a numeric trend is appropriate depends on the scientific target and model assumptions.
Assumptions and a Practical Preservation Check
Preserving the analysis-relevant variables and functional relationships is necessary, but it is only one part of a defensible analysis. The target should be stated clearly, the analysis model should represent the intended estimand, and the imputation models should be adequate for the observed data and missingness pattern. Standard missing-data assumptions—including an appropriate missing-at-random assumption conditional on the variables used—must also be considered. The required support must be present: for example, relevant predictor combinations cannot be wholly absent if the intended relationships are to be learned from the data.
In MICE, the variables being imputed have conditional models within an iterative procedure. The preservation question is therefore practical as well as conceptual: do those conditional models retain the information and relationships needed by the planned analysis? Compatibility and congeniality are related concerns, but preserving an analysis term alone does not establish either property. Nor does it ensure that imputations are proper or that the final inference is unbiased.
Name the target coefficient or contrast, the population, and the analysis model that defines it.
Identify which variables, scales, factor levels, and derived terms determine the target.
Ask whether values treated alike by the imputation model can have different contributions to the target model.
Review distributions and relationships relevant to the target, then assess model fit, support, missingness assumptions, and uncertainty separately.
Diagnostics can reveal warning signs. Compare the distributions of variables and analysis-relevant terms across observed and imputed values, remembering that their marginals need not match exactly. Check whether imputed values preserve plausible relationships with other observed variables, whether factor levels or important ranges disappear, and whether derived analysis terms are computed consistently. Use masked-value checks where feasible, and inspect chain behavior for instability. Such checks can identify problems but cannot prove that the estimand has been preserved or that missingness assumptions are true.
Alternatives, Limitations, and a Common Mistake
| Approach | Potential advantage | Key limitation |
|---|---|---|
| Impute the original variable, then construct analysis terms | Can retain the original scale and make analysis transformations explicit. | Validity still depends on adequate imputation models and correct handling of derived terms. |
| Represent analysis-relevant categories or transformations directly | Makes the target’s distinctions visible in the imputation design. | Does not by itself ensure appropriate conditional distributions or uncertainty. |
| Use a model designed around the substantive analysis structure | Can align the imputation process more closely with the intended analysis. | Requires suitable modeling assumptions and careful implementation; it is not a substitute for checking the target. |
| Use complete cases or a simpler imputation model | May be easier to implement and describe. | May change the analyzed population or fail to retain relationships needed for the target. |
A common mistake is to say, “The imputed variable has a plausible distribution, so the regression estimand is preserved.” This does not follow. A variable can have a plausible marginal distribution while losing a relationship, a scale distinction, or a category contrast that defines the regression target. The correction is to assess not just the variable’s values, but also whether its representation retains the analysis-relevant design and relationships.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A regression targets a linear slope for a continuous \(X\). An imputation model retains only whether \(X\) is above a threshold. What must be checked to decide whether this preserves the target?
- In an analysis with a \(\log(X)\) term, why does retaining \(X\) on a linear scale not automatically preserve the analysis relationship?
- A three-level categorical predictor has separate target contrasts. What restriction is introduced by coding the levels 0, 1, and 2 with one linear coefficient?
- Does preserving all analysis terms prove that the MICE analysis is valid? Name two additional issues to assess.
- Why can plausible marginal distributions of imputed values fail to establish preservation of a regression estimand?
Solutions. (1) Check whether distinct values of \(X\) that have different contributions to the target slope are collapsed into the same threshold category. If they are, the category alone cannot retain the continuous distinction. (2) The analysis contribution depends on \(\log(X)\), so a linear representation does not necessarily retain that transformed relationship; the imputation and analysis representations must be assessed against the target term. (3) One linear coefficient forces equal increments between successive scores, whereas separate category contrasts need not be equally spaced. (4) No. Adequacy of the conditional imputation models, missingness assumptions, support, proper uncertainty, and analysis-model specification remain relevant. (5) Marginal plausibility says little about whether the imputed values retain the conditional relationships, scales, or category contrasts that determine the target coefficient or contrast.