A Bayesian mixture model for changepoint estimation using ordinal predictors.

Area of Science:

• Biostatistics
• Statistical Modeling
• Medical Informatics

Background:

• Ordinal predictor variables (e.g., ECOG performance status, biomarker levels) are common in medical regression.
• Determining the correct functional form for ordinal predictors is statistically challenging.
• Common approaches like dichotomization or linear treatment can lead to inaccurate inference.

Purpose of the Study:

• To propose a novel Bayesian mixture model for assessing the functional form of ordinal predictor variables.
• To simultaneously evaluate dichotomous and linear forms, identifying potential changepoints.
• To offer a flexible and accurate method for regression analysis with ordinal predictors.

Main Methods:

• Developed a Bayesian mixture model incorporating both dichotomous and linear predictor forms.
• Framed the problem as a threshold detection problem to identify changepoints.
• The model accommodates continuous, binary, and survival outcomes and integrates with penalized regression.

Main Results:

• The proposed Bayesian mixture model effectively assesses the functional form of ordinal predictors.
• Simulation studies demonstrated the method's validity and accuracy.
• Application to real-world datasets confirmed its practical utility.

Conclusions:

• The Bayesian mixture model provides a robust approach to handling ordinal predictor variables in regression.
• This method enhances statistical inference accuracy, leading to more reliable conclusions in medical research.
• The availability of JAGS code facilitates easy implementation and adoption of the proposed technique.