Covariates and Treatment Effects in Joint Outcome-Dropout Models

Why covariate placement matters

A joint outcome-dropout model has two related but distinct components: a model for the repeated outcome and a model for the time at which observation ends. Treatment and measured covariates can appear directly in either component. Shared random effects provide another route of association: a latent subject characteristic can influence both the outcome trajectory and dropout risk. Keeping these routes separate is essential. A treatment coefficient in the hazard is not the same parameter as a treatment coefficient in the outcome model, and neither is the shared-effect association parameter.

The focus here is model specification and interpretation. The examples use a continuous repeated outcome and a continuous-time dropout hazard, but the same distinction between direct covariate effects and shared-effect association applies with other outcome types and hazard links. A covariate’s appearance in one submodel does not automatically require its appearance in the other; inclusion should reflect the scientific model and the intended conditional comparisons.

A joint specification with treatment and covariates

Let \(i\) index participants and \(j\) index scheduled outcome measurements at times \(t_{ij}\). Let \(A_i\) denote treatment assignment, coded 0 for control and 1 for active treatment; let \(X_i\) be a baseline covariate centered at a meaningful reference value; and let \(Z_i(t)\) be a time-varying covariate available just before time \(t\). Let \(T_i\) be the dropout time. One useful specification is

Here \(f(t)\) describes the population time trend, while \(b_{0i}\) and \(b_{1i}\) are participant-specific random intercept and slope. The treatment coefficient \(\beta_A\) is the treatment contrast at \(X_i=0\), conditional on the other model terms. If the interaction \(\beta_{AX}\) is included, the treatment contrast at baseline value \(X_i=x\) is \(\beta_A+\beta_{AX}x\). The covariate coefficient \(\beta_Z\) describes the conditional outcome difference for a one-unit change in \(Z_i(t)\), when the model’s other terms are held fixed.

A corresponding dropout hazard might be

The baseline hazard \(\lambda_0(t)\) describes the hazard at the reference values of the included predictors and random effects. The coefficients \(\gamma_A,\gamma_X,\gamma_Z\) are direct hazard effects, conditional on the shared random effects and the other listed predictors. For example, \(\exp(\gamma_A)\) is the hazard ratio for active treatment versus control at the same covariate values and the same random effects. The terms involving \(\alpha_0\) and \(\alpha_1\) link dropout risk to the latent outcome trajectory. If \(\alpha_0>0\), a larger random intercept corresponds to a higher conditional dropout hazard, all else equal.

This formulation assumes that the random effects are drawn from a specified distribution, often a multivariate normal distribution, and that the residuals and random effects follow the stated outcome model. A common working assumption is that the random-effects distribution is the same across treatment groups and does not depend on baseline covariates; if that is not plausible, it must be modeled rather than silently assumed. The joint model also assumes that, conditional on the random effects and included predictors, the remaining outcome and dropout processes have the specified conditional structure. These are substantive assumptions, not consequences of including covariates.

Worked Example: Treatment Effects That Vary by Baseline Severity

Direct hazard terms and shared-effect association

A direct treatment term in the hazard allows dropout rates to differ between treatment groups even after conditioning on the shared effects and covariates. In contrast, a shared-effect term says that people with different latent outcome trajectories can have different dropout risks. In the specified hazard, the conditional hazard ratio comparing two people with random effects \((b_0,b_1)\) and \((b_0',b_1')\), at the same time and fixed covariates, is

This is an association conditional on the model; it does not say that changing a participant’s latent outcome causes dropout. Nor is it a direct effect of treatment. A fixed treatment shift in the outcome mean changes \(\beta_A\), but it does not itself change the distribution of \(b_i\).

The distinction is important when describing treatment-related dropout. If \(\gamma_A=0\), and the random-effects distribution is identical across treatment groups, treatment is independent of those effects, and the hazard depends on them as written, then the model does not imply a treatment difference in the marginal dropout pattern merely because \(\beta_A\ne 0\). A treatment-related difference could instead arise if treatment changes the distribution of shared effects, if a treatment-by-covariate term affects the hazard, or if another specified pathway—such as post-treatment outcomes entering the hazard—is included. Such pathways must be stated in the model.

Worked Example: A Direct Treatment Effect in the Hazard

Time-varying covariates: timing and interpretation

A time-varying covariate can enter the outcome model, the hazard model, or both. Its hazard coefficient \(\gamma_Z\) is a conditional association with the instantaneous dropout rate, comparing participants with different current \(Z_i(t)\) values at the same modeled time, treatment, baseline covariates, and shared effects. A covariate that changes after treatment begins may itself be affected by treatment, so its coefficient should not automatically be interpreted as a causal effect.

The value used to predict dropout at time \(t\) must be available immediately before \(t\). Using a value measured only after dropout, or a value whose measurement requires remaining under observation beyond \(t\), creates a temporal inconsistency or leakage. If a covariate is only intermittently measured, the model must state how its value between measurements is represented—for example, as the most recently observed value, if that is scientifically appropriate. Its measurement process and missingness may also need consideration.

Justification: when an outcome shift does not imply a dropout shift

The distinction between an outcome treatment effect and a dropout treatment effect can be seen directly by integrating over the shared effect. Consider the simpler hazard \(\lambda(t\mid b,A)=\lambda_0(t)\exp(\gamma_AA+\alpha b)\). Suppose \(b\) has the same distribution \(G\) in both treatment groups and there are no other treatment-dependent terms in the hazard or the distribution of \(b\). Conditional survival to time \(t\), for a given \(b\), is \(\exp\{-\Lambda_0(t)\exp(\gamma_AA+\alpha b)\}\), where \(\Lambda_0(t)=\int_0^t\lambda_0(s)\,ds\). Marginal survival is therefore

If \(\gamma_A=0\), the expression contains no \(A\), so it is the same in both treatment groups. The outcome model may have \(\beta_A\ne0\), but that coefficient is absent from this hazard and from the assumed common distribution \(G\). Consequently, a fixed shift in the outcome mean alone does not create a treatment difference in dropout under these conditions. A treatment difference requires a treatment-dependent hazard term or treatment-related change in the distribution of shared effects, or another explicitly modeled route.

Specification checks, alternatives, and limitations

Before interpreting estimates, check that each term has a clear role and a defensible measurement time. In particular, verify treatment coding and reference levels, centering of baseline covariates, the time scale and units, and whether interactions were intended. Examine whether the assumed covariate effects are plausible on the chosen scale: a linear effect in the outcome predictor or a constant log-hazard effect may be inadequate. A time-varying covariate entered in the hazard should be predictable from information available before the risk interval. Check whether treatment groups plausibly share the modeled random-effects distribution; randomization alone does not guarantee that a misspecified distribution is harmless in finite samples.

Warning signs include unstable estimates when treatment and covariate terms are highly correlated, a treatment estimate that changes sharply under modest changes to covariate coding, implausible predicted hazards for parts of the covariate range, or systematic residual and calibration patterns. Such patterns do not identify one unique error, but they prompt review of functional forms, interactions, omitted predictors, measurement timing, and the assumed shared-effect structure. Covariate adjustment cannot by itself validate the conditional-independence assumptions or determine the correct latent-effect distribution.

A shared-parameter model represents outcome-dropout association through common latent effects. Selection models instead express dropout conditional on outcomes, while pattern-mixture models describe outcome distributions by missingness pattern. These are different parameterizations and modeling strategies, not automatic substitutes that produce identical assumptions. Covariates may be included in any of them, but the interpretation of a coefficient depends on the model in which it appears. The present joint specification is limited by its assumed hazard form, outcome distribution, covariate histories, and shared-effect distribution; it does not remove the need to assess sensitivity to those assumptions.