BAYESIAN MODELS for CD4 COUNT, VIRAL LOAD, and SURVIVAL

                                Kate Cowles
                Department of Statistics and Actuarial Science
                             University of Iowa

   Quantitative measures of the amount of human immunodeficiency virus
   present in plasma (called "viral load") are increasingly being used
   as the primary endpoint in clinical trials of antiretroviral drugs.
   However, recent studies indicate that, even after controlling for
   viral load, CD4 count remains an important predictor of time to
   clinical disease progression or death.  In clinical trials of HIV
   therapies, CD4 count and viral load usually are measured
   repeatedly on each patient.  Both are subject to missingness and
   to measurement error and random biological fluctuation.  Thus,
   sophisticated models are needed to capture the relationships
   among time to clinical progression or death and the longitudinal
   trajectories of viral load and CD4 count.


   We fit fully Bayesian models in which bivariate longitudinal
   processes for CD4 count and viral load are embedded in accelerated
   failure time models for time to disease progression or death.
   Data obtained from assay kit manufacturers and from previous clinical
   trials are used to develop appropriate informative priors for the
   measurement error.  Markov chain Monte Carlo methods enable flexible
   choices of structures for the correlations within and between the two
   marker processes as well as of the form of the baseline survival
   distribution.   Sampling-based methods are used for model choice.
   We illustrate with data from a clinical trial.