Visual predictive check of longitudinal models and dropout.

Area of Science:

• Pharmacometrics
• Clinical Trial Methodology
• Statistical Modeling

Background:

• Visual predictive checks (VPCs) are standard for evaluating pharmacometric models.
• Patient dropout, common in clinical trials (especially oncology), can bias VPC performance.
• Existing dropout-handling methods for VPCs lack clear understanding of their differences and applicability.

Purpose of the Study:

• To clarify methods for handling dropout in VPCs and their appropriate use.
• To introduce an improved VPC approach using confidence intervals constructed from observed data.
• To compare the performance of different VPC methods with dropout.

Main Methods:

• Developed a theoretical framework for incorporating dropout into VPCs.
• Proposed and implemented two approaches: full (parametric time-to-event) and conditional (parametric or Cox proportional-hazard models).
• Applied methods to tumor growth dynamics (TGD) modeling using data from two cancer trials (nivolumab, docetaxel) with 3504 measurements from 855 subjects.

Main Results:

• The full approach showed limited improvement over naive VPCs (no dropout adjustment) for TGD model evaluation.
• The conditional approach, using Weibull or Cox proportional-hazard models, outperformed the full approach.
• Confidence intervals enhance VPC interpretation; the conditional approach is more generally applicable with dropout.

Conclusions:

• The conditional VPC approach, incorporating confidence intervals from observed data, provides a more robust evaluation when patient dropout occurs.
• This method is particularly valuable for complex models like TGD in oncology trials.
• Nonparametric methods may offer additional robustness in VPC analysis with dropouts.