What the Conditioning Set Specifies
In fully conditional specification (FCS), each variable with missing values has its own conditional model. In an update for one variable, that variable is the target; the other variables chosen to predict it form its conditioning set. The set is a modeling decision: it determines which information appears on the right-hand side of that target’s conditional model. It does not, by itself, determine which records can be used or establish that a missing-data assumption is satisfied.
For example, if the variables are age, treatment, baseline severity, and a follow-up laboratory measurement, an update for the follow-up measurement might use the other three variables as predictors. An update for treatment would have a different target and could have a different conditioning set. The same variable can be a predictor in one update and the target in another.
A conditioning set identifies which variables enter the model, but not necessarily every detail of how they enter. A continuous predictor might be represented linearly, with a nonlinear term, or in an interaction. A categorical predictor needs an appropriate categorical representation. Thus, specifying a set and specifying a model form are connected decisions, but they are not identical.
Choosing Predictors for a Target
A useful starting point is to ask which variables contain information relevant to predicting the target, to its missingness, or to the later analysis. Include variables needed to preserve important relationships among the variables being imputed and the analysis variables. A variable can be helpful even if it is not itself part of the final analysis: such a variable may provide auxiliary information about the target.
The conditioning set should be considered separately for each target. There is no requirement that every update use the same predictors. For a continuous target, a linear regression model may be a reasonable conditional model if its mean relationship and residual behavior are adequate. For a binary target, a logistic regression model may be appropriate if the binary outcome is modeled with a suitable logit relationship. Other target types require models suited to their scale and support.
Predictor selection also needs substantive and statistical checks. The proposed model should make sense for the variables, have enough information to estimate its terms, and represent important patterns plausibly. A predictor with a strong relationship to the target may be useful, but a large collection of weak or redundant terms can make estimation unstable. An interaction or nonlinear term can represent a relationship that a main-effects-only model cannot. Adding such terms should be motivated by the subject matter or by a credible pattern, not by indiscriminate searching.
Worked Example: Setting Predictors for a Continuous Target
Consider a hypothetical clinical data set with follow-up biomarker concentration in units per liter, age in years, treatment group, and baseline severity score. The follow-up concentration is the target \(X_j\). A proposed conditioning set is age, treatment, and baseline severity; the target itself is excluded.
Suppose, for illustration, that the conditional mean model is \(\operatorname{E}(X_j\mid X_{P_j})=4.2+0.03(\text{age})+1.1(\text{treatment})+0.5(\text{baseline severity})\), where treatment is coded 1 for the active group and 0 otherwise. For a hypothetical participant aged 60 years, in the active group, with baseline severity score 2, substitution gives \(4.2+0.03(60)+1.1(1)+0.5(2)=4.2+1.8+1.1+1.0=8.1\). The model’s conditional mean prediction is 8.1 units per liter for these predictor values.
This calculation clarifies what the conditioning set does: it defines the information used in the conditional mean. The prediction is not a completed value, and the example does not establish that a linear model is adequate. The analyst would need to assess whether a linear mean relationship and the chosen predictors are credible for this hypothetical application.
Worked Example: A Binary Target and Its Predictors
In a hypothetical study, the target is whether a participant experiences a specified binary symptom at follow-up. Let \(X_j=1\) denote symptom present and \(X_j=0\) denote symptom absent. Suppose the conditioning set contains age and treatment, and the proposed logistic model is \(\operatorname{logit}\{P(X_j=1\mid X_{P_j})\}=-2+0.04(\text{age}-50)+0.7(\text{treatment})\), with treatment coded 1 for active treatment and 0 for comparison.
For a hypothetical 60-year-old participant in the active group, the linear predictor is \(-2+0.04(60-50)+0.7(1)=-2+0.4+0.7=-0.9\). The corresponding conditional probability is \(1/(1+\exp(0.9))\approx 0.2891\). Thus this specified model assigns a conditional probability of about 0.289, or 28.9%, to symptom presence for those predictor values. This is an implication of the stated model, not evidence that the model fits well or that age and treatment are sufficient predictors.
Here the conditioning set consists of age and treatment, while the symptom indicator is the target. Entering the target as a predictor would not answer the intended question: it would use the value being modeled to predict itself.
Terms, Redundancy, and a Short Justification
Adding a variable to a conditioning set is not guaranteed to improve an imputation model. A useful way to see why is to consider the conditional mean. If \(X_k\) is conditionally independent of \(X_j\) given \(X_{P_j}\), then learning \(X_k\) in addition to \(X_{P_j}\) does not change the conditional distribution of \(X_j\). In particular, it does not change its conditional mean:
This follows from the definition of conditional independence: given \(X_{P_j}\), the conditional distribution of \(X_j\) is the same whether or not \(X_k\) is also known. Integrating \(X_j\) with respect to those identical conditional distributions gives the same conditional mean. If \(X_k\) is not conditionally independent of the target, it may contribute predictive information; whether it does so in a way that can be estimated reliably depends on the data and model.
This justification is about information in a conditional model. It does not say that every statistically predictive variable must be included, or that conditional independence is known from a finite data set. It explains why “include everything” is not a universal rule: extra terms can contribute little while increasing estimation demands. Conversely, excluding an important predictor can make the conditional model fail to represent relevant associations.
Worked Example: Representing a Treatment-Specific Relationship
Suppose, in a hypothetical study, a follow-up symptom score is the target, baseline symptom score is \(B\), and treatment indicator \(T\) is 0 or 1. Consider the model \(\operatorname{E}(X_j\mid B,T)=2+0.6B+1.0T-0.2BT\). The conditioning set contains \(B\) and \(T\); the product \(BT\) is an interaction term formed from those predictors.
For a participant with baseline score \(B=5\) in the comparison group (\(T=0\)), the predicted conditional mean is \(2+0.6(5)+1.0(0)-0.2(5)(0)=2+3+0-0=5\) score points.
For a participant with the same baseline score in the active group (\(T=1\)), it is \(2+0.6(5)+1.0(1)-0.2(5)(1)=2+3+1-1=5\) score points.
The equal predictions at \(B=5\) do not mean the interaction is redundant at every baseline value. The baseline-score slope is 0.6 when \(T=0\), and \(0.6-0.2=0.4\) when \(T=1\). The interaction allows those slopes to differ. Whether that flexibility is warranted requires subject-matter justification and adequate information for estimating the terms.
Common Strategies and Their Trade-Offs
One strategy is to use a common predictor set for every target, often including all other available variables. It is simple to describe and can retain broad associations, but may be too ambitious when there are many variables, sparse categories, or strong redundancy. Another strategy is to choose a separate set for each target. This can better match the conditional model to the target and its data, but requires careful documentation and risks omitting useful predictors if the choices are made casually.
| Strategy | Potential advantage | Potential concern |
|---|---|---|
| Common set of predictors | Consistent and easy to audit across updates | May include unnecessary or unstable terms for some targets |
| Target-specific sets | Can reflect each target’s scale and relevant predictors | Requires a reasoned choice for every conditional model |
| Include every available variable | May preserve a wide range of associations | Can create redundancy, sparse model terms, or estimation problems |
FCS uses conditional models one target at a time; a joint-modeling approach instead specifies a joint distribution for the variables. These are different modeling strategies, not interchangeable guarantees of validity. Within FCS, choosing a conditioning set does not prove that all the conditional models are mutually compatible or that the resulting procedure has desirable properties. Those are separate questions.
Diagnostics and Limitations
Assess each target model rather than treating a list of predictors as a sufficient specification. Useful warning signs include unstable coefficient estimates, very large standard errors, sparse combinations of categorical predictors, separation in a binary model, and implausible fitted values. Strong dependence among predictors can also make individual terms difficult to estimate. These symptoms can indicate that the chosen set or model form is too complex for the available information.
- Check the target’s scale. The model should respect whether the target is continuous, binary, categorical, or another type.
- Check the terms. Consider whether key nonlinearities or interactions are omitted, and whether included terms have a defensible purpose.
- Check support. Look for predictor combinations with few or no data points, especially combinations created by interactions or multiple categories.
- Check sensitivity. If there is uncertainty about a predictor or term, compare reasonable specifications and note whether the substantive conclusions are sensitive to that choice.
A conditioning set is not a substitute for considering the missing-data mechanism. Choosing predictors does not establish that missingness is conditionally independent of unobserved values, nor does it guarantee that the conditional models are correctly specified. Missingness assumptions and their implications belong to the broader analysis; here, the practical aim is to make the predictor role in each FCS model explicit and defensible.
A common mistake is to say, “All variables are in the data set, so every variable should predict every other variable.” The correction is to evaluate the set target by target. Include variables with a clear role in predicting the target or preserving relevant associations, but check whether the model can support the resulting terms. Also, do not confuse a target with a predictor: when modeling \(X_j\), the conditioning set contains other variables, not \(X_j\) itself.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- For an update modeling \(X_2\) from \(X_1\) and \(X_3\), identify the target and the conditioning set. Is \(X_2\) part of its own conditioning set?
- A binary target has a proposed logistic model with age and treatment as predictors. For a 50-year-old participant with treatment coded 1, the linear predictor is \(-1.5+0.02(\text{age}-50)+0.5(\text{treatment})\). Calculate the probability of target value 1.
- In a continuous-target model, when can adding \(X_k\) leave the conditional mean unchanged, given the existing conditioning set?
- Give one reason an interaction might be included and one diagnostic warning that could indicate the resulting model is too complex.
- Does choosing a conditioning set establish that the missingness assumption is satisfied? Explain briefly.
Solutions. (1) \(X_2\) is the target and the conditioning set is \(\{1,3\}\); \(X_2\) is excluded from that set. (2) The linear predictor is \(-1.5+0.02(0)+0.5=-1.0\), so the probability is \(1/(1+\exp(1))\approx 0.2689\), or 26.9%. (3) If \(X_j\) and \(X_k\) are conditionally independent given the existing predictors, then their conditional distribution for \(X_j\), and hence its conditional mean, is unchanged by adding \(X_k\). (4) An interaction can represent a predictor effect that differs across groups; sparse combinations or unstable coefficient estimates are warnings of excessive complexity. (5) No. The conditioning set specifies model predictors; it does not establish a missingness assumption.