Why Cross-Omics Dependence Can Disappear
When a subject is missing one omics modality, imputation should represent uncertainty about that block while retaining its relationship with the blocks that are observed. A procedure can produce plausible values for each modality separately and still fail at this task: if it imputes each block independently, the completed data may have much weaker cross-block associations than the data-generating process.
Let \(U\) and \(V\) denote two modality measurements for the same subject. They may be single measurements in a simple illustration, or vectors of features in an application. A cross-omics association is a feature of their joint distribution, such as \(\operatorname{Cov}(U,V)\), a correlation, or a conditional relationship. The objective is not to force every imputed value to match a particular observed value. It is to draw missing values from a model that reflects their uncertainty and their dependence on available information.
The distinction matters when a downstream integrated analysis uses relationships between modalities. If those relationships are weakened during imputation, the completed data can suggest a different multivariate structure than the one supported by the observed overlap. Imputing each modality’s marginal distribution correctly is therefore not enough.
Separate Marginal Models and a Gaussian Benchmark
Consider a centered bivariate normal model for two scalar features:
Here \(\sigma_{UV}=\operatorname{Cov}(U,V)\). Suppose the covariance matrix is positive definite, so neither variable is an exact linear function of the other. The conditional Gaussian distributions are
These distributions show what an imputation of \(V\) should use when \(U\) is observed: the conditional mean depends on \(U\), and the conditional variance represents remaining uncertainty. Drawing \(V\) from its marginal distribution instead discards this conditioning information.
Worked Example: Independent Marginal Draws Weaken an Association
Suppose, hypothetically, that \(U\) and \(V\) are standardized, jointly normal measurements with correlation \(0.8\). For three subjects, the observed \(U\) values are \(-1,0,1\), while \(V\) is missing for all three. Compare the expected cross-product association under a conditional model and under independent marginal draws.
Under the joint model, \(E(V\mid U)=0.8U\). The conditional expected \(V\) values are therefore \(-0.8,0,0.8\). The sample covariance based on these conditional expectations is \( [(-1)(-0.8)+(0)(0)+(1)(0.8)]/(3-1)=1.6/2=0.8\). This is the association represented by the conditional means in this small illustration.
Under independent marginal imputation, each \(V\) is drawn from its standard normal marginal distribution, independently of \(U\). Thus \(E(V\mid U)=0\), and for each subject \(E(UV)=U E(V\mid U)=0\). The expected cross-product association is zero, not \(0.8\). A particular random completed sample need not have covariance exactly zero, but the imputation model does not encode the true positive dependence.
This calculation isolates the issue: correct marginal distributions do not ensure a correct joint distribution. Conditioning the missing block on observed modalities is essential when those modalities carry information about it.
Joint Imputation and Conditional Compatibility
A joint imputation model specifies a distribution for all modalities together, then draws missing components from the appropriate conditional distribution given the components observed for that subject. In the Gaussian example, this uses the conditional formulas above. This approach makes the cross-block covariance part of one coherent model. Its validity still depends on assumptions such as an adequate joint distribution, sufficiently informative overlap, and a suitable missingness assumption for the inferential target. Under a missing-at-random analysis, for example, the model must condition on the observed information relevant to missingness; a joint model does not make a wrong missingness assumption correct.
Chained equations instead specify a sequence of conditional models, one for each block, and cycle through them. Such a procedure can preserve dependence if the conditional models include relevant other modalities and generate draws that reflect their relationships. But merely including cross-block predictors is not a guarantee: the specified conditionals may fail to correspond to any single joint distribution. The following compatibility calculation makes that concern concrete for two Gaussian conditionals.
To justify the slope condition, write \(b_{V\mid U}=\sigma_{UV}/\sigma_U^2\) and \(b_{U\mid V}=\sigma_{UV}/\sigma_V^2\). Their product is
A valid covariance matrix has \(\sigma_U^2>0\), \(\sigma_V^2>0\), and \(\sigma_{UV}^2<\sigma_U^2\sigma_V^2\), so \(0\le\rho^2<1\). If \(\sigma_{UV}=0\), both slopes are zero and the variables are independent under the bivariate normal model. This explains why the lower bound is inclusive.
Worked Example: Checking Gaussian Conditional Compatibility
Suppose two hypothetical centered conditional models are \(E(V\mid U)=0.8U\), with conditional variance \(0.96\), and \(E(U\mid V)=0.5V\), with conditional variance \(0.60\). Check whether these specifications can arise from a common nonsingular bivariate normal distribution.
The product of slopes is \(0.8(0.5)=0.40\), which is in the required range \(0\le 0.40<1\). For a direct construction, set \(\operatorname{Var}(U)=1\). The first slope implies \(\operatorname{Cov}(U,V)=0.8(1)=0.8\). The second slope then implies \(\operatorname{Var}(V)=0.8/0.5=1.6\).
The variance of \(V\mid U\) implied by this joint model is \(1.6-0.8^2/1=0.96\). The variance of \(U\mid V\) is \(1-0.8^2/1.6=0.60\). Both match the stated conditional variances. The covariance determinant is \(1(1.6)-0.8^2=0.96>0\), so the covariance matrix is positive definite. With zero means, these conditionals are compatible with the joint normal distribution having variances \(1\) and \(1.6\) and covariance \(0.8\).
If the same slopes had instead been paired with conditional variances that contradicted these identities, passing the slope-product check would not rescue them: the full conditional specifications would not describe this common Gaussian model.
Choosing and Checking an Imputation Strategy
The main options differ in how directly they encode cross-block dependence. Their suitability depends on the pattern of available data, the number of features, and whether the model can be estimated adequately from the subjects with overlapping modalities.
| Approach | How dependence is represented | Main concern |
|---|---|---|
| Independent marginal models | Each modality is modeled without conditioning on the others; cross-block dependence is generally absent from the draws. | Can erase associations even when each marginal model fits well. |
| Joint model | A single model specifies the joint distribution and yields conditional draws for missing blocks. | Requires an adequate, estimable joint model; high-dimensional blocks can make this difficult. |
| Chained conditional models | Each block is imputed conditional on available predictors, including other modalities where appropriate. | Conditionals may be incompatible, poorly specified, or unstable; convergence of the cycle does not prove compatibility. |
In practice, assess whether the imputation model has a basis for learning each important cross-block relationship. Subjects observed in both modalities provide direct information about their joint behavior. If overlap is sparse or concentrated in a narrow part of the covariate space, a sophisticated joint model cannot manufacture the missing information without stronger assumptions. This links dependence preservation to the overlap structure and identifiability considerations in the earlier material on those subjects.
Useful checks include comparing observed cross-block relationships with the corresponding relationships in completed data, reviewing whether conditional models include scientifically relevant predictors, and examining whether the model’s distributions are plausible in the observed ranges. These checks are diagnostic, not proofs: a completed-data association may look reasonable while still reflecting misspecification, and a noisy estimate can vary across imputations. Detailed masking-based and distributional diagnostics are addressed in the later material on those subjects.
Identify which cross-block associations the integrated analysis needs to retain, rather than treating each modality’s marginal fit as sufficient.
Use a joint model or conditional models that include relevant cross-block information and, where possible, are compatible with a coherent joint distribution.
Review overlap, model assumptions, conditional predictions, and whether imputed relationships are plausible. Interpret these checks alongside the limitations of the observed data.
Common Pitfalls and Limitations
- Equating good marginal fit with good joint fit. Correct-looking distributions within each modality do not establish that their dependence is represented. Check cross-block relationships as well.
- Assuming predictors guarantee compatibility. Including another modality in each chained equation can help, but incompatible conditional models need not correspond to one joint distribution.
- Confusing random variation with preserved dependence. An independent draw may happen to produce a nonzero sample covariance. The relevant question is whether the model’s expected conditional behavior encodes the association.
- Imputing a structurally absent modality as if it were intermittently missing. A block that was not defined or collected for a subject is not automatically a missing value from the same measurement process. The distinction requires substantive judgment.
- Overstating what imputation can recover. Sparse overlap, weak predictors, a misspecified distribution, or nonignorable missingness can undermine dependence estimates. Imputation does not replace information or remove assumptions.
The practical trade-off is between coherence and flexibility. A joint model makes dependence explicit but may be restrictive or computationally demanding. Chained equations can accommodate different variable types and tailored conditional models, but their flexibility requires careful attention to compatibility and diagnostics. Independent modality-specific models may be reasonable only when cross-block independence is scientifically defensible for the purpose at hand, not simply because separate fitting is convenient.
Worked Example: Conditional Imputation for a Missing Block
Suppose, hypothetically, that standardized features \(U\) and \(V\) follow a bivariate normal distribution with correlation \(0.6\). For a subject with observed \(U=1\) and missing \(V\), calculate the conditional distribution and one possible imputed value.
For standardized variables, \(\sigma_U^2=\sigma_V^2=1\) and \(\sigma_{UV}=0.6\). Therefore, \(E(V\mid U=1)=0.6(1)=0.6\), and \(\operatorname{Var}(V\mid U)=1-0.6^2=1-0.36=0.64\). The conditional standard deviation is \(\sqrt{0.64}=0.8\), so \(V\mid U=1\sim N(0.6,0.64)\).
If a hypothetical normal draw uses \(z=0.5\), the imputed value is \(0.6+0.8(0.5)=1.0\). Another draw could differ; the conditional mean is not a fixed replacement value. Repeating the draw across imputations represents uncertainty while keeping the observed \(U=1\) informative about \(V\).
This calculation assumes the bivariate normal model is adequate and its parameters are known for illustration. Real analyses must estimate model parameters, account for their uncertainty through the imputation procedure, and assess whether the overlap and missingness assumptions support the intended use.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- For standardized Gaussian modalities with correlation \(0.5\), what is the mean and variance of \(V\mid U=2\)?
- Can a nonsingular bivariate normal model have conditional slope product zero? What does that imply?
- Two conditional slopes are \(0.7\) and \(0.6\). Is their product by itself enough to establish compatibility?
- Why can independent marginal imputations weaken a cross-block association even if each marginal distribution is correctly modeled?
Solutions. (1) The conditional mean is \(0.5(2)=1\), and the variance is \(1-0.5^2=0.75\); hence \(V\mid U=2\sim N(1,0.75)\). (2) Yes. Zero covariance gives both slopes zero and, in the bivariate normal model, independence; zero is included in \(0\le b_{V\mid U}b_{U\mid V}<1\). (3) The product is \(0.42\), within the necessary range, but compatibility is not established: intercepts and conditional variances must also agree with a common joint distribution. (4) Independent draws do not condition the missing value on the observed modality, so their expected cross-product can omit the population association even when the marginal draws are individually correct.