Using VA registry data to estimate age-related risk trajectories for HIV patients

Ut Southwestern Medical Center, Dallas TX

Investigators

Linked publications & trials

Abstract

DESCRIPTION (provided by applicant): Specifically, a Bayesian survival model, which uses the nonparametric smoothing spline approach, will be constructed to estimate an age-related risk trajectory and assess the interaction between aging and long-term exposure to Highly Active Antiretroviral Therapy (HAART). The VA CCR data contains diagnoses, pharmaceutical usage, laboratory, and resource utilization information from more than 64,000 HIV infected veterans. This study will exploit the rich information contained in CCR data to construct a population-level risk trajectory over a wide range of age, say 30 to 75. For each subject, the segment of the risk trajectory within the range of his/her age during the follow-up period forms a unique individual baseline hazard curve. This approach is more flexible than the traditional Cox survival model, where the proportional hazard assumption implies that the hazard curves of all patients have exactly the same shape. The Bayesian survival model also includes a varying coefficient model (VCM) component to assess age- HAART interaction. Specifically, an age-varying regression coefficient is assumed for the cumulative exposure to HAART. The VCM enables researchers to assess whether HAART received at a younger age has a different effect on CVD risk than that received at an older age. Both the age-related risk trajectory and the age-varying coefficient of HAART are modeled by smoothing spline functions, which can accommodate any unknown pattern of change during the course of aging. The smoothness assumption imposed by the smoothing spline model is intuitively appealing because most biological changes take place gradually as patients grow older. Important research questions, such as whether there exists significant age-HAART interaction or whether the CVD risk increase monotonically with age, will be answered through model selection based on the deviance information criterion (DIC). Furthermore, cross-validation procedures will be performed to assess the predictive power of the Bayesian survival model. The full data set will be randomly split into a derivation and a validation set. The Bayesian model will be fitted on the derivation st and then used to predict survival probabilities for subjects in the validation set. The Harrell¿s C statistic will be calculated to measure the concordance between observed survival outcomes and estimated survival probabilities. Finally, an R package and webpages will be developed to help disseminate the knowledge achieved from this study. The R package is more oriented toward biostatisticians who are versed in statistical programming and are interested in applying the Bayesian survival model to their own data. The webpages is constructed to help clinicians utilize the knowledge about CVD risk in HIV patients learned from CCR data.