Combining epidemiologic and biostatistical tools to enhance variable selection in HIV cohort analyses

Area of Science:

• Biostatistics
• Epidemiology
• Survival Analysis

Background:

• Variable selection is crucial for multivariate regression models.
• Complex HIV cohort analyses require robust variable selection strategies.
• Integrating epidemiological and biostatistical approaches enhances model building.

Purpose of the Study:

• To explore comprehensive variable selection methods for multivariate regression in HIV cohorts.
• To compare different statistical approaches for identifying significant covariates.
• To develop a parsimonious and well-fitting survival model.

Main Methods:

• Utilized three variable selection methods: Score test (significance-based stepwise), Akaike Information Criterion (information theory-based stepwise), and Bayesian Model Averaging.
• Applied methods to survival data from the HIV Atlanta VA Cohort Study and the Department of Defense's National History Study.
• Compared resulting models for parsimony and goodness-of-fit.

Main Results:

• All three methods converged on a similar parsimonious survival model.
• Three previously included covariates were excluded from the final models.
• The parsimonious model demonstrated reduced variance in main survival estimates compared to the prior model.

Conclusions:

• Variable selection approaches including significance tests, information criteria, and Bayesian model averaging are effective.
• A balanced, parsimonious model integrating these methods offers improved fit and reliability.
• The findings support a comprehensive strategy for variable selection in HIV research.