Identifying and interpreting subgroups in health care utilization data with count mixture regression models.

Area of Science:

• Health Services Research
• Biostatistics
• Health Economics

Background:

• Inpatient care constitutes a significant portion of healthcare expenditure.
• Analyzing inpatient utilization patterns is crucial for understanding healthcare spending growth.
• Common utilization measures (e.g., length of stay) exhibit statistical complexities like zero inflation, overdispersion, and skewness, complicating analysis.

Purpose of the Study:

• To apply and compare likelihood-based and parametric Bayesian mixture models for analyzing inpatient utilization.
• To identify latent patient subgroups with distinct utilization patterns and covariate relationships.
• To evaluate the accuracy of Bayesian approaches versus information criteria for finite mixture model selection.

Main Methods:

• Application of likelihood-based finite mixture models.
• Application of parametric Bayesian mixtures of negative binomial and zero-inflated negative binomial regression models.
• Simulation study to compare model performance in identifying mixture components.
• Analysis of hospital length of stay data for lung cancer patients.

Main Results:

• The Bayesian approach demonstrated superior accuracy in identifying the true number of mixture components compared to information criteria.
• Distinct patient subgroups were identified within the lung cancer patient cohort.
• These subgroups differed in their mean and variance of hospital days, health and treatment covariates, and the relationships between covariates and length of stay.

Conclusions:

• Parametric Bayesian mixture models offer a robust approach for analyzing complex inpatient utilization data.
• The identified patient subgroups provide valuable insights into heterogeneous healthcare utilization patterns.
• These findings can inform more targeted healthcare resource allocation and cost management strategies.