Shared Latent-Variable Models For Joint Outcomes

Why Introduce a Shared Latent Variable?

At a fixed dose, patients can differ in underlying characteristics that affect both whether treatment is effective and whether toxicity occurs. If those characteristics are not fully measured, modeling efficacy and toxicity as unrelated outcomes may miss an association across patients. A shared latent variable represents this unobserved patient-to-patient variation. The central distinction is that outcomes can be independent conditional on the latent variable and still be associated marginally, after that variable is averaged over.

Here, \(Y_E=1\) indicates efficacy and \(Y_T=1\) indicates toxicity within prespecified assessment windows; zero denotes absence of that outcome. Dose is denoted by \(d\). The model concerns the joint distribution of these two binary outcomes for a patient at dose \(d\). It does not, by itself, establish that efficacy causes toxicity or that a latent characteristic is a biological mechanism.

A General Binary-Outcome Model

One standard specification uses logistic regression for each outcome. Let \(U\) be a shared random effect with mean zero and variance one, and let \(\operatorname{expit}(x)=1/(1+\exp(-x))\). The model is

The dose-specific terms \(\eta_E(d)\) and \(\eta_T(d)\) describe the conditional log odds when \(U=0\); the loadings \(\lambda_E\) and \(\lambda_T\) describe how susceptibility shifts each outcome's conditional log odds. Setting the variance of \(U\) fixes its scale: without a scale constraint, multiplying \(U\) by a constant and dividing both loadings by that constant would leave the linear predictors unchanged.

The standard normal distribution is one possible assumption, not a requirement. A finite latent-class model instead assigns patients to unobserved classes, each with a specified probability and class-specific outcome probabilities. That simpler version makes the association mechanism transparent and allows exact calculations by weighted sums.

For either specification, the model assumes the latent-variable distribution is appropriate for the population at the dose being considered; the same patient's \(U\) affects both outcomes; and no residual association remains after conditioning on \(U\) and dose. If measured patient factors are included, the model can condition on them as well. In a randomized dose study, comparisons should still respect the trial's assignment and analysis plan; the latent variable is not a substitute for accounting for design features or missing outcome data.

How the Marginal Association Arises

The mechanism follows from the law of total covariance. Given dose \(d\), write \(m_E(U)=E(Y_E\mid U,d)\) and \(m_T(U)=E(Y_T\mid U,d)\). Decomposing covariance into its conditional and between-latent-variable parts gives

Under conditional independence, the first term is zero. The remaining covariance reflects how the two conditional probabilities vary together across values of \(U\). When both probabilities increase with \(U\), patients with larger \(U\) tend to have higher probabilities of both outcomes, producing positive marginal covariance. When one increases and the other decreases, the induced covariance can be negative. With nonlinear links and more general latent distributions, the signs of the loadings are informative but the magnitude must be obtained by integrating over \(U\).

Conditional and Marginal Interpretations

In the logistic specification, \(\lambda_E\) and \(\lambda_T\) are conditional log-odds changes per one-unit increase in \(U\), holding dose and other modeled covariates fixed. They are not marginal effects for the trial population. The marginal probabilities require integration:

The marginal joint probability is the integral of the product of the two conditional probabilities under conditional independence. In general, it is not the product of the two marginal probabilities. Because the logistic function is nonlinear, averaging conditional probabilities also does not generally give the probability obtained by setting \(U=0\). Consequently, a conditional odds ratio or coefficient should not be reported as though it were a population-averaged effect.

Assumptions, Diagnostics, and Alternatives

The model's conditional-independence assumption is substantive: it says that the shared latent factor accounts for the residual association between outcomes after conditioning on dose and included covariates. It may be implausible if, for example, an unmodeled patient feature affects only one outcome in a way related to another residual process. A shared effect can also be too restrictive when different subgroups have different dependence patterns.

Assess fit by comparing model-implied and observed outcome frequencies at each dose, including all four cells of the binary joint table: both outcomes, efficacy only, toxicity only, and neither. Also inspect whether predicted marginal rates and conditional patterns are plausible across dose and relevant covariates. Sparse cells, unstable loading estimates, or substantial changes under reasonable latent-distribution choices are warnings that the available data may not support a detailed dependence model. Since \(U\) is unobserved, its distribution and interpretation cannot be checked directly from the data alone.

A shared random-effect model is one way to represent dependence, not the only one. The earlier material on copula models describes a different approach: specify marginal outcome distributions and use a dependence structure to join them. Shared-effect models make the conditional patient-heterogeneity story explicit, but their marginal distributions and dependence arise together from the chosen latent distribution and loadings. A direct joint binary model can instead parameterize joint cell probabilities more directly, though it may offer less of a latent patient interpretation. Model choice should be guided by the scientific question, estimand, data support, and plausibility of assumptions.

A single pair of binary outcomes per patient may provide limited information about the shape of the latent distribution and its loadings. Different latent specifications can fit similar observed joint tables, so an estimated shared effect should not automatically be interpreted as a uniquely identified biological susceptibility. Repeated outcomes, informative covariates, or carefully justified prior information may help, but do not eliminate the need to assess assumptions. The model also does not resolve issues such as unobserved outcomes, unequal follow-up, or changing outcome definitions; those require their own design and analysis choices.

Common Pitfalls

• Equating conditional independence with no association. The mistaken reasoning is that if efficacy and toxicity are independent given \(U\), they must be independent overall. The correction is to integrate over \(U\); variation in both conditional probabilities can create nonzero marginal covariance.
• Assuming every shared effect creates positive association. The direction depends on how \(U\) affects each outcome. Effects in opposite directions can induce negative dependence.
• Calling a latent association causal. A shared factor is a model component, not proof of a biological pathway or of a causal effect of efficacy on toxicity.
• Reporting conditional coefficients as population-average effects. Marginal probabilities require integration over the latent distribution, and nonlinear links make conditional and marginal summaries generally different.