Extending the latent variable model for extra correlated longitudinal dichotomous responses.

Area of Science:

• Pharmacometrics
• Pharmacokinetics/Pharmacodynamics (PK/PD) modeling
• Statistical modeling of categorical data

Background:

• Generalized nonlinear mixed-effects modeling for ordered categorical data has been available for over a decade.
• Pharmacometricians observe fewer transitions or greater correlations in response values than models predict.
• Existing latent variable (LV) approaches offer convenience for incorporating pharmacological concepts.

Purpose of the Study:

• Introduce the multivariate latent variable (MLV) approach for PK/PD analysis.
• Address the issue of extra correlation or fewer transitions in ordered categorical data.
• Evaluate the accuracy and computational efficiency of MLV approximation methods.

Main Methods:

• Developed the multivariate latent variable (MLV) approach, extending the latent variable (LV) method.
• Utilized correlations between latent residuals (LR) to account for extra correlation.
• Formulated and evaluated four approximation methods for dichotomous MLV data via simulation.
• Presented analytical results for models linear in subject-specific random effects.
• Revisited a case study using an MLV approximation method.

Main Results:

• The MLV approach is flexible and accurately predicts transitions by handling diverse correlated data structures.
• Incorrect modeling of population covariances in LV models with extra correlation may not impact marginal mean predictions.
• Accurate prediction of population mean probabilities leads to adequate population variance predictions, even with misspecified covariances.
• Between-subject random effects describe marginal covariances, not marginal variances, for response data.

Conclusions:

• The MLV approach enhances the accuracy of predicting transitions in ordered categorical PK/PD data.
• For population mean predictions, the simpler LV model may suffice even with extra correlation.
• The MLV approach is recommended when individual-level data prediction or simulation is the primary objective.