Characterizing retention in HAART as a recurrent event process: insights into 'cascade churn'

Abstract

Objective: The benefits of HAART rely on continuous lifelong treatment retention. We used linked population-level health administrative data to characterize durations of HAART retention and nonretention.

Design: This is a retrospective cohort study.

Methods: We considered individuals initiating HAART in British Columbia (1996-2012). An HAART episode was considered discontinued if individuals had a gap of at least 30 days between days in which medication was prescribed. We considered durations of HAART retention and nonretention separately, and used Cox proportional hazards frailty models to identify demographic and treatment-related factors associated with durations of HAART retention and nonretention.

Results: Six thousand one hundred fifty-two individuals were included in the analysis; 81.2% were male, 40.6% were people who inject drugs, and 42.8% initiated treatment with CD4 cell count less than 200 cells/μl. Overall, 29% were continuously retained on HAART through the end of follow-up. HAART episodes were a median 6.8 months (25th, 75th percentile: 2.3, 19.5), whereas off-HAART episodes lasted a median 1.9 months (1.2, 4.5). In Cox proportional hazards frailty models, durations of HAART retention improved over time. Successive treatment episodes tended to decrease in duration among those with multiple attempts, whereas off-HAART episodes remained relatively stable. Younger age, earlier stages of disease progression, and injection drug use were all associated with shorter durations of HAART retention and longer off-HAART durations.

Conclusion: Metrics to monitor HAART retention, dropout, and reentry should be prioritized for HIV surveillance. Clinical strategies and public health policies are urgently needed to improve HAART retention, particularly among those at earlier stages of disease progression, the young, and people who inject drugs.

Fig. 1. Adjusted hazards of time to HAART and off-HAART episode discontinuation

Hazard ratios were estimated by including interaction terms between (multinomial) episode count and temporal era variables in a multiple regression model, controlling for all covariates noted in Table 3. 95% CI, 95% confidence interval; epi, episode.