Dose finding with longitudinal data: simpler models, richer outcomes.

Area of Science:

• Clinical Oncology
• Biostatistics
• Pharmaceutical Research

Background:

• Phase I oncology trials aim to determine optimal doses for Phase II studies.
• Traditional methods often focus on toxicity at cycle 1, despite repeated measurements.
• Ordinal toxicity scales are common, but modeling repeated events requires advanced statistical approaches.

Purpose of the Study:

• To evaluate a proportional odds mixed-effect model for ordinal outcomes in dose-finding.
• To compare this model against methods using repeated binary variables and under-parameterized models.
• To assess the utility of repeated measurements for identifying optimal doses and time trends in toxicity.

Main Methods:

• Comparison of statistical models: proportional odds mixed-effect model for ordinal outcomes versus repeated binary variables and under-parameterized models.
• Analysis of simulated and real-world phase I oncology clinical trial data.
• Evaluation of model performance based on accuracy of dose-finding and detection of time trends.

Main Results:

• Both repeated binary and ordinal outcome models significantly improve dose-finding accuracy.
• Ordinal outcome models demonstrate superiority in detecting toxicity time trends, even with nonproportional odds.
• Less parameterized models exhibited the best operating characteristics for dose-finding.

Conclusions:

• Integrating repeated toxicity measurements enhances the precision of phase I dose-finding.
• Ordinal outcome models offer a robust framework for analyzing longitudinal toxicity data in oncology trials.
• The proposed statistical approaches provide valuable tools for optimizing dose selection and understanding toxicity patterns over time.