Chained Equations for Partially Observed Modalities

Why Chain Conditional Models?

A study may measure several omics modalities on the same subjects while some subjects lack an entire modality block. A single joint distribution for every block can be difficult to specify or estimate. Chained equations offer a modular alternative: specify a conditional model for each modality in turn, then repeatedly update the blocks that are missing. The method is flexible, but its flexibility makes model specification and compatibility important.

Let \(Y_i^{(m)}\) denote the vector of measurements for subject \(i\) in modality \(m\), and let \(R_{im}=1\) if that block is observed and \(R_{im}=0\) if it is missing. Let \(C_i\) denote fully observed covariates used in the imputation models. A subject can have some blocks observed and other blocks missing. The procedure imputes the unobserved values of a block; it does not change the observed values.

These conditional models are usually estimated using subjects whose target block is observed. Their predictors can include observed covariates, other observed blocks, and current imputed values of blocks that are missing for the same subject. Thus, an imputation is conditional on the current state of the other blocks, not on their unknown true values.

Assumptions and Model Choices

Chained equations do not by themselves solve the identification problem created by missingness. For an interpretation as imputations of the unobserved values, a typical analysis assumes missing at random (MAR): conditional on the observed data included in the analysis, the probability that a block is missing does not depend on its unobserved value. The imputation models must also be adequately specified, with enough observed information to estimate the relevant conditional distributions. If the probability of observing a block is effectively zero for a covariate-defined subgroup, its conditional distribution in that subgroup cannot be learned from the observed data without additional assumptions.

The model for a target block should reflect its measurement scale and structure. A continuous scalar target might use a regression model with a residual draw; a binary feature might use logistic regression; and a count might use a count model. For a block with many features, analysts may use featurewise conditional models or a multivariate model, but the chosen approach must be estimable with the available sample size and overlap. Predictors should include variables associated with the target values and variables useful for explaining block availability, where those are observed and relevant.

When a modality is structurally absent because it was never defined or collected for a subject, imputing it as though it were an intermittently unmeasured value can change the scientific target. Chained equations are intended for quantities that are meaningful but unobserved under the analysis model. The distinction between structural absence and intermittent modality missingness therefore matters before fitting any conditional model.

A useful conditional model must generate plausible variation, not merely a fitted conditional mean. For example, replacing missing values only with regression predictions suppresses residual variation and can make the completed data too certain. Stochastic draws should reflect uncertainty in both the missing value and, when appropriate, the estimated model parameters. Whether the resulting imputations support a particular analysis also depends on the substantive variables and estimand; a convenient imputation model is not automatically suitable for every analysis.

One Iteration of the Algorithm

Choose an order for the modality blocks, initialize their missing values, and then update each block in turn. At a given update, the target model is fitted or used for that block, and only its missing entries are replaced. The newly updated values are available as predictors in later updates in the same sweep. A sweep through all blocks is one iteration. After an initial period in which the chain moves away from its starting values, retain draws separated or otherwise selected according to the implementation to form an imputed data set. Repeating the procedure from different starting values can help reveal instability.

Compatibility: Do the Conditionals Describe One Joint Model?

A key distinction is between having plausible individual conditionals and having conditionals that can all arise from one joint probability distribution. The latter property is called compatibility. A compatible collection has a joint distribution whose conditionals agree with the specified blockwise models. Chained equations can still be run when the specified conditionals are incompatible, but there may be no joint distribution that justifies the resulting chain as draws from a coherent joint model. In that case, the chain’s behavior and its implied distribution depend on the updating procedure.

A simple bivariate normal case gives a practical compatibility condition. Suppose the proposed conditionals are \(Y_1\mid Y_2\sim N(a_{12}+b_{12}Y_2,v_{1\mid2})\) and \(Y_2\mid Y_1\sim N(a_{21}+b_{21}Y_1,v_{2\mid1})\), with positive conditional variances. If they come from a nondegenerate bivariate normal distribution, both regression slopes are determined by the same covariance. Specifically, \(b_{12}=\operatorname{Cov}(Y_1,Y_2)/\operatorname{Var}(Y_2)\) and \(b_{21}=\operatorname{Cov}(Y_1,Y_2)/\operatorname{Var}(Y_1)\). The conditional variances are \(\operatorname{Var}(Y_1)(1-\rho^2)\) and \(\operatorname{Var}(Y_2)(1-\rho^2)\), where \(\rho\) is the correlation. These expressions imply the restriction below.

The restriction follows by substitution: the left side is \(\{\operatorname{Cov}(Y_1,Y_2)/\operatorname{Var}(Y_2)\} \{\operatorname{Var}(Y_2)(1-\rho^2)\} =\operatorname{Cov}(Y_1,Y_2)(1-\rho^2)\). The right side reduces to the same expression. Failure of this condition rules out a shared nondegenerate bivariate normal distribution with those conditionals. Satisfying it is not, by itself, a general proof of compatibility for more complex models or more blocks; it is a targeted check under the stated Gaussian assumptions.

Diagnostics, Alternatives, and Limitations

No single diagnostic certifies a chained-equation analysis. Inspect traces across iterations for summaries such as block means, variances, and selected cross-block summaries. Persistent drift, large oscillations, or chains that settle at different levels from different initial values indicate that the algorithm may not have stabilized. Also check the imputed values against observed ranges and scientific constraints, and inspect whether model fitting produces separation, implausible extrapolation, or unstable estimates. These checks help identify failures; stable traces alone do not prove that the conditional models are correctly specified, compatible, or identified by the data.

Overlap is especially important for whole-block missingness. If a target block is observed only in subjects whose other blocks or covariates occupy a narrow range, predictions for subjects outside that range rely on extrapolation. A conditional model may appear numerically stable while being unsupported for a substantial missingness pattern. Compare the observed predictors of subjects with and without each block, and verify that the data contain enough observed target blocks in the regions where imputations are required.

Chained equations are one alternative to a joint multiple-imputation model, which specifies a joint distribution directly, and to likelihood-based approaches such as EM for a specified model. They can be convenient when different blocks require different conditional families. In contrast, a joint model makes its joint assumptions explicit, while chained equations can leave their joint implications unclear if the conditionals are incompatible. Matrix-completion methods take a different route by imposing low-rank structure; they are not interchangeable with conditional stochastic imputation. The appropriate choice depends on the measurement scales, scientific model, missingness patterns, and the estimand.

A common mistake is to treat successful convergence as evidence that the imputations are valid: “the traces are flat, so the conditional models must describe the data.” Flat traces show only that selected summaries appear stable under the implemented updates. They do not establish MAR, adequate overlap, correct conditional forms, compatibility, or appropriate uncertainty. Assess those assumptions and diagnostics separately, and report important limitations rather than relying on convergence alone.