Application of a hazard-based visual predictive check to evaluate parametric hazard models.

Area of Science:

• Biostatistics
• Pharmacometrics
• Survival Analysis

Background:

• Parametric time-to-event models are commonly assessed using survival-based visual predictive checks (VPCs).
• Interpreting hazard model deficiencies via survival-based VPCs can be indirect and challenging.
• Hazard is the primary quantity modeled in parametric survival analysis.

Purpose of the Study:

• To evaluate the performance of nonparametric hazard estimators as diagnostic tools for VPCs.
• To assess the viability of hazard-based VPCs for improving model evaluation.

Main Methods:

• Assessed bias of histogram-based and kernel-smoothing nonparametric hazard estimators.
• Evaluated estimators using Weibull and bathtub-shaped hazard scenarios.
• Examined the utility of these estimators for VPC evaluation of hazard models.

Main Results:

• Nonparametric hazard estimators performed reasonably well across sample sizes, with expected bias near time boundaries.
• Flexible bandwidth and boundary correction methods effectively reduced bias.
• Nonparametric estimators successfully identified misfit when a Weibull model was applied to bathtub-shaped hazard data.

Conclusions:

• Hazard-based VPCs provide a more direct and interpretable assessment of hazard models compared to survival-based VPCs.
• Nonparametric hazard estimators are viable and useful diagnostics for VPCs in time-to-event analyses.