Abstract

The HIV care cascade is a conceptual model that describes, for an individual patient, progression through stages of care. The cascade begins with HIV diagnosis; subsequent stages include linkage to care, initiation of treatment, retention in care, and eventually suppression of HIV viral load.

When individual-level longitudinal data are available, the care cascade can be operationalized as a state space or compartmental model in which patients transition between stages. From both a clinical care and epidemiologic perspective, it is of interest to summarize rates of transition over time and to understand which factors – at both the patient and program level – might affect transition rates.

In this talk I will describe a statistical model for characterizing the HIV care cascade and show how it can be used both for regression analysis of transition rates and causal inference to compare treatment policies. We use a multinomial model for longitudinal outcomes to characterize transitions between states, and use Bayesian g-estimation for quantifying causal effects. To mitigate potential biases from model misspecification, we use Bayesian additive regression trees (BART) for multinomial outcomes. We argue that the Bayesian approach facilitates a natural implementation of g-estimation because causal effects are quantified in terms of posterior distributions of potential outcomes.

The model is applied to data on over 30,000 individuals in HIV care in the AMPATH program in Eldoret, Kenya. We use the model to assess factors affecting engagement and retention in care, and to quantify the causal effect of immediate treatment upon diagnosis on progression through the cascade.

This work was led by Yizhen Xu, PhD, Postdoctoral Researcher at Johns Hopkins University

Watch Stream