Joint modeling of multiple longitudinal cost outcomes using multivariate generalized linear mixed models

Area of Science:

• Biostatistics
• Health Economics
• Longitudinal Data Analysis

Background:

• Traditional healthcare cost modeling aggregates data, potentially obscuring covariate impacts on specific cost categories.
• Analyzing cost categories separately risks ignoring interdependencies, leading to inaccurate conclusions.
• Longitudinal cost data presents unique challenges in accurately modeling multiple, correlated outcomes.

Purpose of the Study:

• To propose and evaluate a multivariate generalized linear mixed model (mGLMM) for joint modeling of longitudinal healthcare costs from multiple sources.
• To compare four distinct mGLMM approaches varying in how they account for correlations among cost outcomes.
• To identify the most suitable modeling strategy for analyzing complex healthcare expenditure data.

Main Methods:

• Developed and assessed four mGLMM configurations: shared random intercept, shared random intercept and slope, separate random intercepts (joint distribution), and separate random intercepts and slopes (joint distribution).
• Employed goodness-of-fit measures (AIC/BIC) and residual plots for model comparison.
• Utilized longitudinal data from 740,195 US veterans with diabetes (2002-2006) to demonstrate the joint modeling approaches.

Main Results:

• The separate random intercepts approach demonstrated the best fit (lowest AIC/BIC) for both log-normal and gamma generalized linear mixed models (GLMMs).
• Despite superior fit, more complex models did not yield qualitatively different conclusions compared to the simpler shared random intercept model in this specific dataset.
• Model selection depends on balancing statistical fit with the practical interpretation of results for longitudinal healthcare cost data.

Conclusions:

• Multivariate generalized linear mixed models (mGLMMs) provide a robust framework for jointly analyzing longitudinal healthcare costs from diverse sources.
• While separate random intercepts may offer superior statistical fit, a shared random intercept model can be sufficient when conclusions remain consistent.
• The choice of mGLMM approach should consider the specific dataset characteristics and the need for interpretable results in health economic research.