From One Joint Model to a Set of Conditional Models
When several variables have missing values, a completed data set must supply values for all those gaps while retaining the relationships among variables. Fully conditional specification (FCS) addresses this by modeling each incomplete variable conditionally on the others, then repeatedly using those models to update missing entries. The method is useful when variables have different types, such as a continuous laboratory measure, a binary indicator, and a count.
FCS is a modeling and computation strategy, not a guarantee that the specified models are correct or that the missingness process permits valid inference. Its key idea is to specify a conditional model for each variable that has missing values, rather than begin by specifying one joint probability model for all variables.
Suppose \(X_1,\ldots,X_p\) are variables, and some entries in each are missing. For an incomplete \(X_j\), FCS specifies a conditional model of the form
where \(X_{-j}\) denotes the other variables in the analysis. For example, if a continuous outcome, a binary indicator, and age are included, the model for a missing outcome might condition on age and the indicator; a model for a missing indicator might condition on age and the outcome.
The conditional model is chosen to suit the type and scale of the variable being imputed. A continuous variable may use a linear regression model, a binary variable may use logistic regression, and a count may use a count regression model. The predictor set should also reflect the substantive analysis: variables needed to preserve important associations should not be omitted merely because they are convenient to leave out.
How Iteration Makes the Models Work Together
At the start of a chain, missing entries have some starting values. A conditional model updates missing entries in one variable using current values of the other variables. As other variables are updated, those current values change, so a later update can use information that was not available at the start. Repetition allows the imputed values to respond to the changing set of predictors.
A useful schematic representation is to let \(X_j^{(t)}\) denote the current values of variable \(j\) at iteration \(t\). An update for an incomplete \(X_j\) uses the current values of the other variables and produces new values for its missing entries. Subsequent updates use those new values. The notation describes dependence between updates; it does not assert that the conditional models necessarily come from one common joint distribution.
In practical multiple imputation, updates must retain uncertainty rather than simply substitute a single best prediction everywhere. The examples below use conditional means or probabilities to make the conditioning logic transparent; those calculations illustrate model inputs and outputs, not a complete stochastic imputation procedure. Replacing missing values by conditional means alone generally suppresses variation and does not, by itself, provide proper multiple imputation.
Worked Example: Updating Two Incomplete Variables
Consider a hypothetical data set with age in years and a continuous biomarker score. A row for one person has age \(50\) and a missing biomarker score; another has a biomarker score of \(18\) and a missing age. Suppose that, in a larger hypothetical data set, fitted conditional mean models are
Assume for this illustration that these linear conditional mean forms are adequate over the relevant ranges, the units are as stated, and the observed cases contain enough information to fit them. For the first row, the missing score's conditional mean is
For the second row, the missing age's conditional mean is
These results show the direction of conditioning: age predicts a missing score, while score predicts a missing age. In an iterative chain, the current values are used as predictors, and updated values can affect later conditional predictions. The arithmetic here uses means to display that dependency; an actual multiple-imputation procedure must account for predictive uncertainty rather than treating these two means as certain replacements.
Conditional Models Must Fit Their Variables
Choosing FCS models is a substantive modeling task. Each conditional model needs an outcome distribution appropriate to its target variable, a defensible predictor set, and a functional form adequate for the observed relationships. If a variable is binary, a linear regression model can produce fitted values outside the interval from zero to one; that is a warning that the model's distributional form is inappropriate for a binary outcome. A logistic model instead represents the conditional probability through the log odds.
Worked Example: A Binary Conditional Model
Suppose a hypothetical study has a binary adherence indicator, coded \(1\) for adherent and \(0\) otherwise, with some adherence values missing. A fitted conditional logistic model uses age in years and a binary group indicator \(G\):
For a hypothetical person aged \(50\) years in group \(G=1\), the linear predictor is
The corresponding conditional probability is
Thus the fitted conditional probability of adherence is about \(69.0\%\) for these covariate values. This is a probability from the model, not a continuous imputed adherence value such as \(0.690\). An imputation procedure for the missing binary variable must produce a value on the binary scale while retaining uncertainty. The model is appropriate in form for a binary outcome, but its adequacy still depends on sufficient observed cases, a reasonable logit relationship with age, and a suitable predictor set.
Iteration in a Small Numerical Illustration
A simple pair of conditional mean equations makes the iterative dependence visible. They are not offered as a realistic fitted model or as a recipe for deterministic imputation; they isolate how one conditional prediction becomes an input to another. Suppose a hypothetical row has both \(X\) and \(Y\) missing, and the illustrative conditional means are
Worked Example: Following Values Across Iterations
Start with illustrative current values \(X=4\) units and \(Y=10\) units. First, use the current \(X\) to calculate the conditional mean for \(Y\):
Next, use this updated \(Y=4\) to calculate the conditional mean for \(X\):
On another pass, the corresponding conditional means are \(Y=2+0.5(2.6)=3.3\) units and \(X=1+0.4(3.3)=2.32\) units. The input values have changed, so the conditional predictions change as well. This arithmetic shows why FCS is iterative rather than a collection of one-time predictions. It does not establish that a chain converges, that these conditional equations describe a coherent joint distribution, or that deterministic conditional means are adequate imputations.
Assumptions, Diagnostics, and Alternatives
FCS does not itself specify why data are missing. Valid statistical inference requires an appropriate missing-data framework, along with suitable conditional models. In particular, an analyst must justify the missingness assumptions used for the intended inference; the iterative algorithm alone cannot make those assumptions true. A conditional model also needs enough observed information to estimate its parameters, appropriate handling of variable types, and adequate representation of important nonlinearities or interactions.
Useful checks include reviewing each model's fit on observed cases, checking predicted values for implausible ranges, and comparing observed and imputed distributions and associations. A model that repeatedly produces extreme predictions, fails to represent a strong nonlinear pattern, or yields imputed distributions unlike plausible values is a reason to revisit the specification. An apparent stable sequence of numerical summaries is not proof that the models are well specified or that inference is valid.
The main alternative is joint modeling: specify a joint distribution for the variables and derive the needed conditional distributions from it. When a suitable joint model is available and credible, this gives a unified probabilistic specification. FCS can be more convenient for mixed variable types because each variable receives a model tailored to its scale, but its separately specified conditionals are not automatically guaranteed to correspond to one joint distribution. This distinction matters when evaluating the assumptions and interpretation of the procedure.
| Approach | What is specified | Key consideration |
|---|---|---|
| FCS | A conditional model for each incomplete variable | Flexible for mixed types; each conditional model and the overall specification need scrutiny |
| Joint modeling | One joint distribution for the variables | Provides a unified model, but choosing a credible joint distribution can be difficult |
| Single imputation | One replacement value for each missing entry | Does not generally represent imputation uncertainty in subsequent inference |
A common mistake is to treat FCS as though choosing a plausible model for every variable automatically guarantees a valid joint analysis. The correction is to distinguish the computational strategy from its statistical assumptions: each conditional model needs to be adequate, and their combined use must be assessed rather than presumed valid. A second mistake is to equate the predicted conditional mean with the imputed value and then analyze the result as complete data. That discards uncertainty and can distort variability and associations.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A continuous measure and a binary indicator both have missing entries. What is the defining modeling choice in FCS?
- Why can a conditional probability of \(0.69\) for a missing binary value not simply be entered as the completed value \(0.69\)?
- In the illustrative models \(E(Y\mid X)=2+0.5X\) and \(E(X\mid Y)=1+0.4Y\), current \(X=6\). What is the conditional mean for \(Y\)?
- Does successful execution of an FCS algorithm establish that the missingness assumptions are appropriate? Explain briefly.
Solutions. (1) Specify a conditional model for each incomplete variable, selecting a model suited to its type and conditioning on the other variables. (2) The indicator must be imputed on its binary scale; \(0.69\) is a model probability, not a binary outcome, and using only a predicted mean omits uncertainty. (3) \(2+0.5(6)=5\) units. (4) No. FCS describes how conditional models are iteratively used; it does not establish the missingness assumptions or the adequacy of the conditional models.