Methods for handling model misspecification in pharmacometric model-based approaches for the analysis of longitudinal clinical trial data

Sammanfattning: Pharmacometrics model-based analyses proved to be increasingly useful in drug discovery and development. This usefulness was made evident by the formalisation of the model-integrated drug discovery and development framework (MID3). The overall objective of this thesis was to assess and improve the robustness of longitudinal pharmacometric-model-based approaches toward model misspecification. IMA is a new approach using mixture models to test for a drug effect. It was presented and assessed together with other approaches, including model averaging, in the context of controlled balanced two-armed parallel trials. The assessment (type I error, power, and accuracy) carried out over real and simulated data demonstrated problems with the existing method, caused by model misspecification and selection bias, and the robustness of IMA towards these issues. In the same context another novel approach presented and assessed among the alternatives, rcLRT, also showed robustness and good performance. IMA was generalised to unbalanced designs and dose-response scenarios where it confirmed its good performance. A new extension of the visual predictive checks diagnostic tool to the specific case of mixture models was presented. It was able to diagnose the mixture proportion aspect, enabling the diagnostic of model misspecification in that aspect. PACS is a new approach balancing data information and prior information in the selection process between correlated covariates, in which the prior was introduced as a penalty to the objective function. PACS was able to influence to some extent the final covariate selection process and resulted in a strategy not previously described as the best under several conditions This approach is an interesting tool against data-driven misspecification when strong prior knowledge is available. A new methodology revolving around "digital twins" in pharmacometrics was presented as another alternative to making analysis results robust with respect to model approximations. The method filtered the simulations based on observed data. The selected avatars are both insurance that some of the simulated data are mimicking reasonably well the observations at clinical endpoints of interest, and a diagnostic tool for model misspecification when the avatar subset is too different from the raw model simulations. In conclusion, the new pharmacometric approaches provide robust methods towards model misspecification, which opens up new possibilities for the use of pharmacometrics models in drug development.