Why Passive Imputation Is Needed
A derived variable is often included in an analysis because it represents a scientifically meaningful feature of other variables: a squared predictor, a product term, a total, or a rate. When one or more of its inputs are incomplete, the derived variable is incomplete too. Within Multiple Imputation by Chained Equations (MICE), the central choice is whether to model that derived variable as another stochastic variable or to calculate it from its inputs.
If the relationship is exact by definition, a separate model can generate values that contradict it. For example, if \(Z=X^2\), an independently imputed value of \(Z\) need not equal the square of the imputed \(X\). Passive imputation avoids that contradiction: it calculates the derived value from the source variables in each completed data set. The term is updated as a deterministic function, not sampled from a separate conditional distribution.
The earlier material on Derived Variables as Deterministic Functions establishes the key distinction: a derived term is not a new independent measurement when its value is fixed by its inputs. Here the focus is how to handle that distinction in a chained-equations workflow.
What Is Stochastic, and What Is Passive?
Suppose \(X_1,\ldots,X_k\) include incomplete variables. MICE uses conditional models to generate imputations for those source variables. The resulting completed values vary across imputations because they represent uncertainty about the missing data. Once a particular completed value of the sources is available, the value of \(Z\) is fixed by \(f\).
For imputation \(l\), where \(l=1,\ldots,m\), and record \(i\), the passive term is \(Z_i^{(l)}=f(X_{i1}^{(l)},\ldots,X_{ik}^{(l)})\). The superscript identifies the completed data set; it does not indicate an additional draw for \(Z\). Different completed values of the sources can produce different values of \(Z\) across imputations, so uncertainty in the sources is carried through the calculation.
The distinction is between uncertainty about the inputs and uncertainty about the definition. Passive calculation propagates the former and adds none of the latter. This is appropriate only when the definition is genuinely deterministic on the scale being used.
Worked Example: Calculating a Derived Value From an Imputed Input
Consider hypothetical records with weight \(W\) in kilograms and height \(H\) in metres. Define body mass index by \(Z=W/H^2\). Assume this is the exact definition intended for these recorded values, and that height is observed. For one record, suppose \(H=1.70\) m and an imputed completed-data value is \(W=68.0\) kg.
The passive value is calculated as \(Z=68.0/(1.70)^2\). First, \((1.70)^2=2.89\) square metres. Then \(Z=68.0/2.89=23.53\) kilograms per square metre, to four significant figures.
If a second imputation gives this record \(W=72.0\) kg while height remains \(1.70\) m, its derived value is \(72.0/2.89=24.91\) kilograms per square metre. The difference between these values comes from different imputed weights. Neither BMI value is separately sampled.
The calculation assumes the recorded height is positive and that the stated units and formula apply. If height were zero, the expression would be undefined; a passive formula does not resolve invalid inputs.
Why a Separate Conditional Model Can Break the Definition
A stochastic model for \(Z\) generally allows values that do not equal \(f(X_1,\ldots,X_k)\). That is not merely additional uncertainty: it changes the data structure. A direct justification follows from the conditional distribution implied by an exact definition. Given the source values, the derived variable has probability one at the value prescribed by the function.
Consequently, its conditional variance is zero. A separate imputation model with positive conditional variance assigns positive probability to values outside the deterministic relationship. It is therefore not a model for the same exact derived variable. A point-mass model that always returns \(f(X_1,\ldots,X_k)\) would agree with passive calculation, but it is simply the deterministic calculation expressed as a model.
Worked Example: Checking a Passively Calculated Square
Suppose the analysis uses \(X\) and \(Z=X^2\). In one hypothetical completed data set, a participant has imputed \(X=4.5\). The correct passive value is \(Z=(4.5)^2=20.25\).
Suppose instead that a separate model generates \(Z=19.0\). The pair \((4.5,19.0)\) violates the definition because \(19.0\ne20.25\). The discrepancy is \(20.25-19.0=1.25\), but the issue is not that the discrepancy is large or small: any nonzero discrepancy contradicts the exact definition.
The relevant assumptions are that \(Z\) is defined as the square of this same \(X\), with no rounding rule or measurement-error component that changes the relationship. Under those assumptions, the passive value is \(20.25\), not a separate draw near it.
Using Passive Terms in a Chained-Equations Workflow
In a MICE procedure, distinguish variables that receive conditional imputation models from terms that are determined by other variables. The incomplete source variables are handled by their specified conditional models. The passive term is represented by its defining formula and calculated from the sources in each completed data set. It is not treated as an additional incomplete outcome with an independently estimated conditional distribution.
This distinction matters when the analysis model includes the derived term. For example, if the analysis regresses an outcome on \(X\) and \(X^2\), each completed data set should contain pairs \((X,X^2)\) that obey the identity. Passive calculation preserves that identity. It does not, by itself, guarantee that the conditional models for the source variables are well specified, that the imputation and analysis models are compatible, or that the missingness assumptions needed for the intended inference are reasonable.
The source variables still need appropriate imputation models. A deterministic calculation cannot repair a poor source-variable model, create information that was not observed, or guarantee that the analysis is supported by the imputation procedure. It enforces the definition of the derived term; it does not settle every modeling question.
Worked Example: A Product Term From Two Incomplete Sources
Suppose an analysis includes \(Z=AB\), where \(A\) is a hypothetical dosage in milligrams and \(B\) is a hypothetical duration in days. Assume both variables are measured on the stated scales and the product is defined exactly. For one record, a completed data set contains imputed \(A=12.0\) mg and \(B=3.5\) days.
The passive product is \(Z=(12.0)(3.5)=42.0\) milligram-days. In another imputation, suppose the completed values are \(A=10.0\) mg and \(B=4.0\) days. Then \(Z=(10.0)(4.0)=40.0\) milligram-days.
Both products are consistent with their own source values. A separate model that produced \(Z=45.0\) milligram-days for the first completed pair would be inconsistent, because \(45.0\ne42.0\). The two source variables may be uncertain, but once their completed values are specified, their exact product is not uncertain.
Assumptions, Diagnostics, and Failure Modes
Passive calculation is justified when the defining relationship is exact and the inputs used in the calculation are the intended inputs. Before using it, check that the formula matches the scientific definition, units, and coding of the data. Also check that the completed source values are valid for the function. For example, a logarithm requires a positive input, and a ratio requires a nonzero denominator.
Observed values deserve particular attention. If a derived variable was directly recorded and can differ from the formula because of rounding, measurement error, or a different operational definition, it is not automatically interchangeable with an exact passive term. Overwriting such a measurement with a calculated value changes what the variable represents. Resolve the data definition first; do not force exact consistency when the relationship is not exact.
A useful diagnostic is to calculate the discrepancy between the stored term and its defining function wherever both are available: \(D_i=Z_i-f(X_{i1},\ldots,X_{ik})\). For an exact passive term, \(D_i\) should be zero, apart from documented numerical precision. A nonzero discrepancy can indicate that a value was generated separately, a formula was applied inconsistently, units were mixed, or a source value changed without the derived value being brought into agreement.
Inspect this consistency in every completed data set, not only in the original incomplete data. Also check for undefined or implausible calculated values. Exact arithmetic consistency is necessary for a passive term, but it does not establish that the imputed source values are plausible or that the analysis model is appropriate.
Write the intended function and verify that it is exact for the variables and scales being analyzed.
List the variables needed to calculate the term and confirm they have completed values in each imputed data set.
Use the completed source values to obtain the derived term; do not fit a separate stochastic model for an exact identity.
Verify the identity and confirm that the function is defined for the completed inputs.
Comparison With Other Approaches
| Approach | What it does | Main concern or use |
|---|---|---|
| Passive calculation | Calculates the term from completed source values. | Preserves an exact deterministic definition. |
| Separate conditional imputation | Draws the term from its own stochastic model. | Can generate values that contradict an exact definition. |
| Calculate only for analysis | Derives the term after the source variables have been completed. | Can preserve the identity if done consistently, but the term is unavailable during imputation-model fitting when needed there. |
| Specialized joint or substantive-model-compatible approach | Builds the imputation procedure around more complex relationships between the analysis and imputation models. | May be needed when simple chained conditional models do not adequately represent the required structure. |
The choice between passive calculation and calculating only at analysis time depends partly on whether the derived term is needed within the imputation procedure. In either case, the term used in the analysis must correspond to the completed source values. Separate stochastic imputation is appropriate only if the variable is not in fact an exact deterministic function, or if the scientific model explicitly includes variation beyond that function.
Common Mistakes and Limitations
- Modeling an exact term as if it were a new measurement. A separate conditional draw may look plausible marginally while violating the defining identity. Calculate the term from its sources instead.
- Assuming passive calculation solves compatibility. It enforces one deterministic relationship, but does not ensure that source-variable imputation models represent the analysis relationships well.
- Ignoring the meaning of recorded values. If a directly recorded quantity includes rounding or measurement error, an exact formula may not describe it. Establish whether the variable is truly derived before replacing values.
- Overlooking invalid inputs. A formula can be applied consistently yet produce undefined values or values outside the meaningful domain. Check the inputs and the function's domain.
Passive imputation is deliberately limited: it handles deterministic transformations, not every complicated relationship involving incomplete data. A correct formula does not compensate for misspecified source models, and an exact derived term can still be unhelpful if it is not justified by the analysis question.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A completed data set has \(X=6\) and \(Z=X^2\). What is the passive value of \(Z\), and why should it not be separately drawn?
- In one imputation, a source value changes but its stored derived term does not. What diagnostic should be used?
- A recorded variable is intended to equal a formula but differs slightly because it was rounded before being stored. Is exact passive replacement automatically justified? Explain.
- Across several imputations, a derived product varies because its source values vary. Does this mean the product received an independent stochastic draw?
Solutions. (1) \(Z=6^2=36\). Under the stated exact definition, \(Z\) is fixed once \(X\) is given, so a separate draw could violate the identity. (2) Recalculate the term from the current source values and check \(D=Z-f(X)\); for an exact term, the discrepancy should be zero apart from numerical precision. (3) No. Rounding means the stored variable may not be exactly the deterministic function on the recorded scale, so its definition and role must be clarified before replacement. (4) No. Different source values produce different deterministic products; variation across imputations reflects uncertainty in the sources, not a separate draw for the product.