Common Specification Errors and Reporting Limits

Why Specification and Reporting Require Care

A shared-parameter model links a longitudinal outcome process and a dropout process through a latent subject-level quantity. That link can be useful: subjects with different latent outcome tendencies may also have different dropout risks. But a fitted association parameter is meaningful only relative to how the latent effect is defined, scaled, and entered into both parts of the model. Likewise, a dropout likelihood contribution is meaningful only if it represents the event or continued observation that actually occurred.

The central reporting limit is equally important. A joint model can estimate an association under its assumed structure, but observed data alone generally do not establish that the chosen latent distribution, link functions, and outcome-dropout dependence are correct. Good fit checks can reveal some problems; they cannot turn untestable assumptions into observed facts.

Specify the Shared Effect Before Interpreting It

For subject \(i\), let \(Y_{ij}\) denote an outcome at visit \(j\), and let \(b_i\) be a latent subject-level effect. A simple shared-intercept specification might use a conditional outcome model with linear predictor \(\eta_{ij}=\boldsymbol{x}_{ij}^{\mathsf T}\boldsymbol{\beta}+b_i\), and a continuous-time dropout hazard \(\lambda_i(t\mid b_i)=\lambda_0(t)\exp(\boldsymbol{z}_i^{\mathsf T}\boldsymbol{\xi}+\gamma b_i)\). Here \(\lambda_0(t)\) is a baseline hazard, \(\boldsymbol{x}_{ij}\) and \(\boldsymbol{z}_i\) are covariates, and \(\gamma\) is an association coefficient on the log-hazard scale. More elaborate models may share a random slope or another function of the latent effects; the interpretation changes accordingly.

Several common specification errors follow from overlooking those qualifications:

• Leaving the shared quantity ambiguous. If the outcome model has a random intercept and a random slope, stating only “the random effect predicts dropout” does not say which effect, or what combination of effects, enters the hazard.
• Changing the latent-effect scale without changing the coefficient. A coefficient per one unit of \(b_i\) is not a coefficient per one standard deviation unless that standard deviation is one.
• Reversing the effect’s orientation. Replacing \(b_i\) by \(-b_i\) reverses the sign of its association coefficient. The fitted model can be equivalent if the outcome and dropout parts are both transformed consistently, but the coefficient cannot be interpreted using the old sign convention.
• Confusing conditional and marginal statements. The hazard ratio from \(\gamma\) compares hazards conditional on the specified latent effect and covariates. It is not generally the hazard ratio comparing observed groups defined by their measured outcomes.

Code the Observed Dropout History

A dropout contribution must encode both what happened and how long the subject remained at risk. In continuous time, if \(T_i\) is the observed dropout or censoring time, \(\delta_i=1\) denotes dropout and \(\delta_i=0\) denotes right censoring, the hazard contribution conditional on the latent effect is

The integral represents survival without dropout over the entire observed at-risk period. The hazard factor appears at \(T_i\) only for an observed dropout event. For visits indexed by discrete intervals, define \(D_{ij}=1\) if dropout occurs in interval \(j\), and \(D_{ij}=0\) if the subject is still observed through that interval. If \(h_{ij}(b_i)\) is the conditional dropout probability in an interval, the contribution through the last observed interval \(J_i\) is

This form assumes the interval indicators and hazards are defined consistently with the recorded observation schedule. A frequent coding mistake is to record a dropout event as \(D_{ij}=1\) in every subsequent interval, even though the subject is no longer at risk. Another is to omit the non-event factors for intervals in which the subject was observed to remain in follow-up. Either changes the modeled history.

In a joint model, the outcome and dropout contributions are combined conditional on the shared effects and then integrated over their distribution. The observed outcome contribution must include only outcomes that were actually observed; the dropout contribution accounts for the observed follow-up and event history. A missing post-dropout outcome is not a recorded value to insert into the outcome likelihood.

Conditional Association Is Not a Causal or Marginal Effect

In the hazard model above, increasing \(b_i\) by one unit multiplies the conditional hazard by \(\exp(\gamma)\), provided the model uses the log-linear hazard form shown. If \(\gamma=0\), the hazard has no direct dependence on \(b_i\) through that term. Neither statement by itself says that changing an individual’s outcome would change dropout. The latent effect is a model quantity, and its association with dropout may represent shared unobserved subject characteristics rather than an intervention.

For a binary repeated outcome, the outcome link also matters. If the conditional outcome model is logistic, \(b_i\) shifts the conditional log odds of the outcome; \(\gamma\) still describes its role in the dropout log hazard. Those are different scales. A coefficient in one model cannot be read as an odds ratio or a probability difference in the other.

Justification and Limits of the Joint Likelihood

The conditional dropout expression follows from the definition of a hazard. Given survival without dropout up to time \(t\), the chance of dropout in a short interval of width \(dt\) is approximately \(\lambda_i(t\mid b_i)dt\). The conditional probability of remaining event-free through \(T_i\) is the survival function

For an observed event, the event-time density is the hazard at the event time multiplied by survival to that time, yielding \(\lambda_i(T_i\mid b_i)S_i(T_i\mid b_i)\). For right censoring, the observed contribution is survival through the censoring time, \(S_i(T_i\mid b_i)\), with no event-hazard factor. These are exactly the two cases expressed by \(\delta_i\) in the likelihood contribution above.

To obtain an observed-data likelihood in a shared-parameter model, one integrates the conditional product of the observed outcome contribution and dropout contribution over the latent-effect distribution. This construction relies on the chosen outcome model, dropout model, latent-effect distribution, and conditional dependence structure. For example, conditional independence of outcomes and dropout given the shared effects is an assumption of a common construction, not a fact guaranteed by fitting it.

Consequently, standard fit diagnostics have a limited but valuable role. They can flag poor fit to observed outcome distributions, dropout timing, or covariate patterns; numerical diagnostics can flag unstable estimation. They cannot, using observed data alone, generally prove the assumed latent-effect distribution or the adequacy of the shared-effect dependence mechanism for unobserved outcomes. Conclusions that change markedly across plausible specifications should be reported as sensitive to those assumptions.

Compare the Model and Report Its Conclusions

These alternatives are not interchangeable parameterizations with automatically equivalent interpretations. A responsible report should state the outcome and dropout links, the shared effect and its scale, the latent-effect distribution, the dropout coding and censoring rule, and whether the reported contrast is conditional or marginal. It should give estimates and uncertainty with units or an explicit scale, and distinguish the primary analysis from sensitivity specifications.

For example, say “under the specified shared-random-intercept model, a one-standard-deviation increase in the latent intercept was associated with an estimated conditional dropout hazard ratio of 2.23.” Do not say “higher symptoms caused dropout” unless a separate causal argument supports that claim. Also do not present a single model-based estimate as if the missing outcomes’ distribution had been verified by the observed data.

Common Mistakes and Diagnostics

• Reporting a bare association coefficient. Give the latent effect’s units or standardization and the link scale; otherwise the coefficient has no unambiguous magnitude.
• Using an event contribution for a censored subject. Check that a dropout indicator is one only when dropout was observed and that survival is counted for every interval actually at risk.
• Reading conditional parameters as population-average contrasts. State what is held fixed and avoid interpreting latent-effect coefficients as observed-group effects.
• Treating convergence as validation. Convergence indicates that the numerical procedure met its criteria; it does not verify the model’s untestable dependence assumptions.
• Reporting only the preferred specification. If plausible latent distributions, links, or shared-effect definitions produce different conclusions, show that sensitivity rather than implying robustness.

A useful audit is to trace one subject’s record from entry through the last observed time: identify the outcome values available, the intervals at risk, the dropout event or censoring status, and the exact latent quantity entering each model component. Then check whether every coefficient statement matches the chosen scale and conditioning.