Comparison of different parametric proportional hazards models for interval-censored data: a simulation study

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Survival Analysis

Background:

• Interval censoring is common in clinical trials but often analyzed as right-censored data, leading to potential bias.
• Existing statistical methods for interval-censored data are underdeveloped, particularly for estimating hazard ratios (HR) in proportional hazards (PH) models.
• While parametric PH models are implementable, their performance with mis-specified baseline hazard functions was not well understood.

Purpose of the Study:

• To evaluate the performance of parametric PH models for interval-censored data.
• To assess the impact of mis-specifying baseline hazard functions on HR estimates.
• To determine the most robust parametric PH models for clinical trial data analysis.

Main Methods:

• Conducted an extensive simulation study using data generated from 6 different models.
• Examined parametric PH models with exponential, Weibull, and 10-piece exponential baseline hazard functions.
• Analyzed performance under various data distributions and censoring schemes.

Main Results:

• Mis-specifying the baseline hazard function had minimal impact on hazard ratio estimates.
• Parametric PH models using Weibull or 10-piece exponential functions provided robust HR estimates with low bias and mean squared errors (MSE).
• Using a simple exponential function to approximate complex baseline hazards resulted in increased bias and MSE.

Conclusions:

• Parametric proportional hazards models are effective and practical for analyzing interval-censored data in clinical trials.
• The choice of baseline hazard function approximation (Weibull or 10-piece exponential) is crucial for robust HR estimation.
• These findings support the wider adoption of parametric PH models for interval-censored survival data analysis.