Abstract Body

We propose a state space representation of the HIV care cascade and corresponding statistical framework that can describe the longitudinal dynamics along the phases in the cascade. For this representation, each phase in the HIV care cascade is viewed as a ‘state’. Then various types of patient behaviors such as the cyclical process of engagement, disengagement, and re-entry into care (a.k.a. ‘churn’) in the cascade are described in terms of transition from one state to another. We illustrate the proposed framework using data on 57,596 patients enrolled in AMPATH (a partnership between Kenya and North American institutions) between 2008 and 2012. As testing, linkage, and viral load data was highly limited, we focused on retention aspects of care. 

We operationalized a 5-state care cascade including engaged in care, disengaged from care, transferred-out, lost-to-follow-up (LTFU), and deceased. Transition probabilities were defined by Pjk(t)=Pr(S(t)=k | S(t-1)=j) where S(t) denotes the state at time t. For example, P11(t) represents probability of transition from engaged to engaged, which is retention; P21(t) is probability of transition from disengaged to engaged, which is re-entry into care. We used a multinomial regression for longitudinal data to capture the effects of covariates on state transitions.

On average, in a given 200-day interval, 78% of individuals were engaged in care, 16% disengaged, 0.4% transferred, 5% LTFU, and 1% died. Among those engaged at a given time, probability of retention, disengagement, transfer-out, and death were .84, .13, .001, and .01. Once disengaged, probability of return to care, continued disengagement, LTFU, and death were .10, .54, .35, and .004. Regression modeling identified that among those engaged, patients with CD4>350, age>35 and patients on ARV were more likely to remain engaged in care, while male patients and those whose CD4 was not recently measured were substantially less likely. Among those disengaged, older patients, those previously on ART, and those with CD4<350 were more likely to return to care, and less likely to be permanently LTFU.

Our representation of the HIV care cascade is a novel application of state space models, and includes regression formulations.  It provides a unified approach to modeling individual-level longitudinal data from a clinical cohort. A simple version of the model is illustrated here, but an extended version includes testing, initial linkage, and viral suppression.