Shared Random Effects as Links Between Outcome and Dropout

Why Link Repeated Outcomes and Dropout?

In a longitudinal study, each participant may contribute several outcomes before leaving follow-up. If participants with systematically high or low outcome trajectories also tend to drop out at different rates, the outcome process and dropout process are associated. A shared-parameter model represents that association using latent participant-level random effects. The central modeling choice is whether the same random effect enters both processes, or whether distinct random effects enter them and are allowed to be correlated.

The distinction is substantive, not just notational. A single shared effect says that a particular latent participant characteristic shifts both the outcome trajectory and dropout propensity. Two correlated effects allow the latent feature governing outcomes to differ from the feature governing dropout, while permitting them to vary together. Neither choice, by itself, establishes why dropout occurs or makes the modeling assumptions empirically certain.

Basic Notation and Assumptions

Let \(Y_{ij}\) denote the outcome for participant \(i\) at visit \(j\), and let \(B_i\) be a participant-level latent random effect. Let \(D_i\) indicate dropout during a specified interval in which participant \(i\) is at risk. A simple shared-effect specification has the form

Here \(\mu_{ij}\) is the part of the outcome mean determined by modeled covariates and visit, \(a\) is the outcome association with the latent effect, and \(\alpha+\gamma b_i\) is the conditional log-odds of dropout in the specified interval. The logistic link means that a log-odds value \(z\) corresponds to probability \(1/\{1+\exp(-z)\}\). The same \(B_i\) appears in both equations; that is what makes this a shared-effect specification.

A common working assumption is that the random effects have a specified distribution, often \(B_i\sim N(0,\sigma_B^2)\), and that their distribution does not vary across participants after accounting for modeled covariates. The conditional outcome model also needs an appropriate residual distribution and variance structure. For the dropout component, the link and the event interval must be clearly defined. Those choices are assumptions, not consequences of using a random effect. In the examples below, the stated distributions and conditional mean structure are treated as the model assumptions for the calculations.

These equations describe the components but do not, on their own, specify every detail of the joint data-generating process. In particular, this tutorial uses the shared effect to explain a link; it does not derive a joint likelihood or claim that all outcome and dropout variables are conditionally independent given that effect.

How a Shared Effect Creates Association

A short covariance calculation explains what the shared term accomplishes. Consider one outcome and a latent dropout propensity \(L_i\), where

Assume \(E(B_i)=0\), \(\operatorname{Var}(B_i)=\sigma_B^2\), \(E(\varepsilon_i\mid B_i)=E(\eta_i\mid B_i)=0\), and that the residual terms have zero covariance with each other and with \(B_i\). These conditions are sufficient for the following calculation; they are not a general claim about every shared-parameter model.

The constants have zero covariance with all terms, and the assumed residual covariances are zero, leaving only the product of the two loadings on \(B_i\) times its variance. Thus the common latent effect induces covariance between the outcome and the latent propensity \(L_i\). It is important not to confuse that result with a formula for covariance between the outcome and a binary dropout indicator: the latter also depends on the link function and the rest of the joint model.

One Effect Versus Two Correlated Effects

A distinct alternative assigns \(U_i\) to the repeated outcomes and \(V_i\) to dropout. For example,

with a joint distribution for \((U_i,V_i)\) that allows \(\operatorname{Cov}(U_i,V_i)\ne 0\). The two effects are distinct latent variables even though their distribution is related. Their correlation can represent a tendency for participants with higher outcome effects to have higher or lower dropout effects. In the shared-effect model, by contrast, the same random variable enters both components, with potentially different coefficients.

A nonzero correlation does not make two variables identical. For example, a bivariate normal pair with correlation \(0.6\) still has variation in each component that is not perfectly determined by the other. Conversely, a model that imposes perfect correlation and a deterministic relationship between the effects may collapse to a lower-dimensional shared-effect representation. The distinction should therefore be stated in terms of the model’s latent variables and restrictions, not just described loosely as “linking” the two processes.

Assumptions, Diagnostics, and Limitations

The model’s interpretation depends on its assumptions. The distribution chosen for the random effects, the form of their effect on repeated outcomes, the dropout link, and the way covariates enter each component all matter. A linear random-effect contribution to an outcome mean may be reasonable in one application and inadequate in another. Likewise, a logit-linear dropout association is a modeling choice rather than a universal law. A conditional dropout probability must refer to a defined interval and risk set; a single probability should not be interpreted as a complete model of an entire follow-up course.

Useful diagnostics can reveal some forms of misspecification. Check whether fitted outcome trajectories capture patterns across visits, whether residual variation changes over time, and whether observed dropout frequencies are broadly compatible with model-based predictions across relevant covariate and follow-up groups. Examine sensitivity to plausible alternative random-effect distributions, outcome structures, and dropout links. If estimates or conclusions change substantially under reasonable alternatives, report that dependence rather than presenting one specification as definitive.

These checks have limits. The latent effects are not directly observed, so a good fit to the observed data does not establish that the chosen shared-effect interpretation is true. Nor does a plausible fit prove that dropout assumptions are correct. Shared and correlated-effect formulations can also impose different structures on the relationship between outcomes and dropout; conclusions may depend on those structures. Pattern-mixture and selection models represent alternative modeling frameworks, not merely different names for a shared-effect specification.

A practical comparison is therefore about the scientific and statistical restriction being imposed. A shared effect is parsimonious and gives a direct common latent driver. Separate correlated effects allow the outcome and dropout propensities to have distinct latent sources, at the cost of additional distributional parameters and assumptions. A model with independent effects is a more restrictive alternative in which the latent outcome and dropout effects are unassociated. None should be selected solely because it yields a preferred treatment or outcome conclusion.

Common Mistake

Wrong reasoning: “The outcome and dropout effects are correlated, so this is automatically a shared-random-effect model.” Correlation is a relationship between two variables; it does not make them one variable.

Correction: Write down the latent terms in both components. If the same \(B_i\) enters each, the model has a shared effect. If \(U_i\) enters the outcome and \(V_i\) enters dropout, the effects are distinct, and any association is represented through their joint distribution. State the distinction explicitly, including the assumed distribution and the direction of each effect.