Joint Modeling And Multilevel Chained Equations

Two Ways to Represent Clustered Variables

When several variables are incomplete in clustered data, an imputation procedure must represent both their relationships with one another and their dependence within clusters. A joint model specifies a probability distribution for the variables together. Fully conditional specification (FCS), also called chained equations, instead specifies a conditional model for each incomplete variable given the others, then cycles through those models to update imputations. In either approach, ignoring clustering can distort both within-cluster relationships and uncertainty.

The choice is not simply between a “correct” and an “incorrect” method. A joint model offers a coherent probability model when its distributional assumptions are reasonable and it can represent the variables involved. FCS can accommodate different types of variables through different conditional models, but the collection of conditional models may not correspond to any single joint distribution. For clustered data, either approach also needs a defensible representation of cluster structure.

For notation, let \(i=1,\ldots,m\) index clusters and \(j=1,\ldots,n_i\) index individuals within cluster \(i\). Cluster sizes \(n_i\) need not be equal. The word “multilevel” describes the dependence structure; it does not, by itself, identify a particular imputation algorithm.

A Joint Model and Its Implied Dependence

A simple joint model for a vector of continuous variables measured on each individual is a multivariate random-intercept model. Write the vector as \(Y_{ij}\), its population mean vector as \(\mu\), and the cluster-specific random vector as \(u_i\):

Here \(T\) is the between-cluster covariance matrix and \(S\) is the within-cluster residual covariance matrix. This specification assumes random effects and residual vectors are independent, clusters are independent of one another, and—conditional on \(u_i\)—individual residual vectors are independent with the same covariance \(S\). It also assumes the stated normal distributions and a common covariance structure across clusters. Additional predictors or more complex random effects may be needed in an application.

The model implies that two different individuals in the same cluster share covariance \(T\), while the covariance for two individuals in different clusters is zero under this model. For one individual, the marginal covariance is \(T+S\). These are implications of the model, not facts guaranteed by the mere presence of clusters. If observed data show markedly different cluster variances, non-normal variables, or dependence beyond the random intercept, this simple joint model may be inadequate.

Compatibility: A Concrete Check

Compatibility matters because chained models are often described as though they necessarily define a single joint probability distribution. They need not. A useful small illustration comes from two jointly normal variables \(X\) and \(Y\), with variances \(\sigma_X^2\) and \(\sigma_Y^2\), covariance \(\sigma_{XY}\), and \(|\rho|<1\). Their conditional regression slopes are determined by the same covariance:

The conditional residual variances are \(v_{X\mid Y}=\sigma_X^2-\sigma_{XY}^2/\sigma_Y^2\) and \(v_{Y\mid X}=\sigma_Y^2-\sigma_{XY}^2/\sigma_X^2\). These paired conditional models are compatible because both follow from the same bivariate normal distribution. Their slopes must satisfy the covariance identity below. It avoids dividing by a slope, which would fail when the covariance is zero.

If \(\sigma_{XY}=0\), both slopes are zero; the identity remains valid. Compatibility in a multilevel setting is more demanding: the conditional models must also agree about cluster dependence, relevant predictors, and the distributions of random effects. In practice, a set of plausible conditional regressions is not automatically a compatible set.

How FCS Differs in Practice

In FCS, each incomplete variable is updated in turn from a model conditional on the other relevant variables. For clustered data, those conditional models must reflect the hierarchy in a way appropriate to the variable and the scientific analysis. The sequence of updates is an algorithm; it is not itself proof that the conditionals correspond to a coherent joint distribution. If they are incompatible, the procedure may still settle into a stable simulation behavior, but that behavior is not automatically the distribution intended by a joint model.

The attraction of FCS is flexibility. A conditional model can be chosen to suit a variable’s scale—for example, a continuous, binary, ordinal, nominal, or count variable—rather than forcing all variables into one multivariate normal distribution. That flexibility comes with a responsibility: examine whether each conditional model is sensible and whether the collection represents the important relationships needed for the analysis. The earlier material on multilevel imputation covers particular variable types and cluster structures; the present distinction is about how the models fit together.

A joint model can be attractive when a coherent distribution for all incomplete variables is scientifically defensible and can be estimated with adequate information. Its parameters and dependence assumptions are explicit. But a simple joint normal model may be awkward for mixed variable types, nonlinear relationships, or distinctive cluster structures. FCS may be more adaptable, but it can be harder to establish global coherence, and one poorly specified conditional model can affect multiple subsequent updates.

Assumptions, Diagnostics, and Failure Modes

Both approaches rely on assumptions about the missingness process and the observed-data models. For the usual multiple-imputation justification under missing at random, missingness may depend on observed information included in the imputation model, but not on the missing values themselves after conditioning on that information. The imputation model must also represent the distributions and dependencies relevant to the analysis. Proper imputation requires uncertainty in model parameters to be reflected, rather than treating fitted values as known. These conditions do not make one modeling strategy universally preferable.

In multilevel data, check whether the imputation model preserves cluster membership and represents the relevant between- and within-cluster variation. Compare observed and imputed distributions overall and by cluster; inspect cluster summaries, ranges, category frequencies, and relationships among variables. For FCS, examine convergence and stability across iterations, and check whether conditional models produce plausible values and preserve important associations. A stable trace is useful evidence about algorithmic behavior, but it does not prove compatibility or correct specification. For joint models, check fit of the assumed distributions and covariance structure, including whether the model allows the observed heterogeneity and variable types.

Common warning signs include imputations outside a variable’s allowable range, category frequencies that are implausible within clusters, a loss of observed associations, or estimates that change greatly under modestly different reasonable specifications. These are prompts to investigate, not automatic proof of a particular failure. The missingness assumptions cannot generally be verified from observed data alone, so substantive knowledge and sensitivity analysis remain important.

Common Mistake and Practical Limits

Common mistake: “Each FCS regression looks reasonable, so together they must define a coherent joint model.” This conclusion does not follow. Conditional models can each be plausible in isolation yet be incompatible—for example, their slopes or variance assumptions may not be jointly attainable from one distribution. The correction is to distinguish a convenient iterative imputation algorithm from a fully specified joint probability model, and to assess compatibility where possible rather than assume it.

A second practical limit is information. A complex joint model may require more parameters than the observed clusters can support; highly flexible conditional models can also be unstable when cluster-specific information is sparse. Neither method automatically repairs weak identification, an unrepresented cluster structure, or departures from the missingness assumptions. The main alternative is not “always use the other method,” but to choose a simpler defensible model, improve the predictor set where possible, and assess sensitivity to plausible specifications.