Statistical Analysis

For this analysis, women were excluded if they were HIV-infected at baseline (n=6), had <24 weeks of electronic monitoring data, including those with drug stops (n=94), or never initiated PrEP (n=1), leaving 233 HIV uninfected women with 24 weeks of complete data. Group-based trajectory modeling was used to identify patterns (or trajectories) of PrEP adherence and, separately, HIV risk behavior over the first 24 weeks of PrEP use. These models have been used to find adherence patterns in many health fields, including HIV [18,19]; detailed guides are available for these models as well [20,21] (Supplement). Using a specified number of trajectories and their shape (linear, quadratic, or cubic), the model estimates the proportion of the population belonging to each trajectory and calculates the posterior probability of an individual belonging to each trajectory based on their observed data; the individual is then assigned to the trajectory with the highest posterior probability [20,22–24].

Adherence was modeled on PrEP doses per week (range 0–7), using a censored normal (tobit) distribution [20], and HIV risk behavior was modeled on any sex (regardless of condom use or type of sexual partner) over the prior month, as reported at quarterly study visits, using a binomial distribution, over the first six months of PrEP use. For each outcome, models with 2–6 trajectories were compared using the Bayes test (recommended to be >10 to indicate good fit) and average posterior probabilities (recommended to be >0.7 to indicate good fit)[20,25]; when needed, content expertise was used to determine the most meaningful model. Once the number of trajectories was selected, the same criteria were used to compare linear, quadratic, and cubic terms for each trajectory. The distribution of covariates by assigned trajectory is presented.

Next, baseline covariates were added independently to the models; these are interpreted as the odds of being in each trajectory compared to a reference trajectory. Wald tests were computed to test for significance [22]. Covariates of interest as reported at enrollment were: perception of HIV risk from their study partner (dichotomized as high/moderate/unknown vs low/none); any report of problem drinking behaviors; marital status; parity; couple HIV risk score; and age (</≥25 years old). Likewise, time-varying covariates (also known as turning points) were added independently to the model; these are interpreted as the change in outcome when the covariate was present, within each trajectory [22]. Again, Wald tests were computed to test for significance. Pregnancy, partner’s use of ART, and frequency of sex over the prior month (with their study partner or outside partners) were recorded at follow-up study visits and included as possible covariates. Any significant covariates were included in the final model. We compared mean adherence across groups, using the assigned group as the predictor in a mixed effects model to account for repeated observations per woman.

Finally, we modeled the joint trajectories, using the final models for adherence and risk behavior. Joint trajectories allow the calculation of both the population and individual probabilities of belonging to each adherence trajectory, given a women’s risk behavior trajectory. We used chi-square tests to assess whether the population probability of each adherence trajectory differed by risk behavior trajectory. We also modeled the log₁₀ individual probability of each adherence trajectory in separate linear regressions, with the individual’s probability of the steady risk behavior trajectory as the predictor (rather than assigned risk behavior trajectory, to account for possible error in trajectory assignment). As a sensitivity analysis, we included women missing adherence data (n=45); women who had a drug stop were still excluded as inability to use PrEP is a separate issue from adherence. All analyses were conducted using SAS 9.4 (SAS Institute, Cary NC).