The University of California San Francisco and Committee on Human Subjects Research and the Partner's Human Research Committee approved all procedures.

Participants enrolled in the REACH cohort were invited to participate in a substudy focused on intensive adherence monitoring, as previously described [25], [28], [29]. Briefly, subjects had unannounced visits by research staff at their usual place of residence every three to six weeks, over a one year period. Percent adherence was determined from the number of remaining or unused antiretroviral pills and number of pills refilled between visits [2]. Three hundred and fifty seven HIV-positive subjects were monitored for medication adherence. The earliest monitoring period started in March 1998 and the latest monitoring date was in October 2007.

Confounders of the effect of current adherence on virologic failure included the following: prior adherence, prior duration of HAART, prior exposure to mono/dual nucleoside therapy, recent CD4+ T cell count (lagged 2 months), CD4+ T cell nadir (lagged 2 months), demographics (sex, ethnicity, age), years of education, past and current antiretroviral treatment characteristics, crack cocaine and alcohol use, calendar time, and homelessness. Current antiretroviral treatment was classified as one of five potential regimens: ritonavir boosted protease inhibitor containing regimen, unboosted protease inhibitor containing regimen, non-nucleoside reverse transcriptase inhibitor containing regimen (NNRTI), combined non-nucleoside reverse transcriptase inhibitor-protease inhibitor containing regimen (PI/NNRTI), and nucleoside reverse transcriptase inhibitor (NRTI) only containing regimen.

Blood was drawn monthly for HIV RNA levels (viral loads) and quarterly for CD4+ T cell levels. HIV-1 viral load was performed using the HIV-1 Amplicor Monitor Version 1.0 ultra sensitive assay (Roche Molecular Systems, Alameda, CA).

For every month of observation, each REACH cohort participant was classified as experiencing either virologic suppression (defined as HIV RNA less than 50 copies/ml) or virologic failure (defined as HIV RNA at least 50 copies/ml). Each participant in the REACH cohort who was suppressed for at least two consecutive months during adherence monitoring contributed data to the analysis; there were 221 participants who met this criterion. For each such participant, his/her data was included starting at the first month of viral suppression during adherence monitoring and continuing until virologic failure occurred or the adherence monitoring period ended.

Duration of suppression at any month was defined as the number of prior consecutive months of viral suppression. In calculating the duration of viral suppression, we used data on viral loads prior to the initiation of adherence monitoring.

We used marginal structural models to estimate the effect of adherence during a given month on the probability of virologic failure at the end of that month. These effects were estimated within subpopulations defined by duration of prior viral suppression. For example, among participants who had maintained viral suppression for exactly 4 consecutive months, we estimated the effect of adherence during month 5 on the probability of virologic failure at the end of month 5. For our primary analysis, adherence was stratified into the following four categories: 0−49%, 50−74%, 75−89%, or 90−100% pills taken. The marginal structural model was fit using a targeted maximum likelihood estimator (TMLE) [27], in order to adjust for the potential confounders listed above. The TMLE, in this application, relies on 1) a multinomial logistic regression model for predicting medication adherence given potential confounders and number of consecutive months virally suppressed and 2) a logistic regression model for the probability of virologic failure given adherence, confounders, and number of consecutive months virally suppressed. The TMLE is robust to model misspecification, in that it gives asymptotically unbiased estimates whenever at least one of the models (1) or (2) is correct. It also has several advantages over alternative estimation techniques, such as inverse probability of treatment weighting [30], g-computation [31], [32], and doubly robust estimators [33]–[36]; these advantages have been previously described [27].

Based on the marginal structural model fit using targeted maximum likelihood estimation, we estimated the probabilities of virologic failure at each set stratum of adherence, for duration of continuous suppression ranging from 1 to 12 months. Confidence intervals for the difference between the probability of failure after 1 month of continuous suppression and the probability of failure after 12 months of continuous suppression, at each adherence stratum, were calculated using 10,000 iterations of the nonparametric bootstrap (BCa method) [37].

We tested for effect modification based on the estimated coefficients of the marginal structural models. More precisely, we tested whether the effect of the proportion of medication taken on the probability of virologic failure in the current month, is impacted by the number of consecutive months one has been virally suppressed. This test involved first computing the logarithm of the causal relative risk of virologic failure, comparing high adherence (90–100%) vs. low adherence (0–49%), conditional on number of consecutive months virally suppressed. This causal relative risk was based on the fit of the marginal structural model discussed above. Next, we used a linear regression model to test the null hypothesis that the causal relative risk does not depend on the number of consecutive months virally suppressed. Standard errors were computed using the nonparametric bootstrap as above.

Time-lagged confounder measurements were used to ensure that confounders occurred before, and thus could not be influenced by, medication adherence. We imputed missing confounder values by carrying the most recent observation forward. Because subjects could contribute data at multiple time points, bootstrap resampling was based on subject rather than data-point.

Of the 357 subjects in the REACH cohort who received unannounced pill counts, 221 met our criteria for inclusion in the analysis. The median age at start of adherence monitoring was 44.1 years (IQR 10.5). The median CD4+ T cell count was 390 cells/mm³ and the median CD4+ T cell nadir was 206 cells/mm³. Fifty-six of the 221 participants (25%) were receiving an unboosted protease inhibitor based regimen. The mean observed duration of continuous viral suppression was 6.7 months. Of the 1201 subject-months of observation, 95 (8%) were missing adherence measurements and 159 (13%) were missing viral load measurements. See Tables 1 and 2 for a more extensive list of participant characteristics.

Unannounced pill count adherence ranged from 0 to 100%, with a median of 92%. In univariate analyses, higher adherence was associated with higher past adherence, longer duration of viral suppression, higher CD4+ nadir, and white/Caucasian ethnicity, while lower adherence was associated with use of crack, intravenous drug use, and Black/African American ethnicity. Comparing observed adherence proportions among subjects who maintained viral suppression for different durations of time, there was a gradual decrease in the proportion with 0–49% adherence (from 11% of subject-months at 1 month of suppression to 5% of subject-months at 12 months of continuous suppression) and 50–74% adherence (from 19% of subject-months at 1 month of suppression to 15% of subject-months at 12 months of continuous suppression). There was a corresponding gradual increase in the proportion with 75–89% adherence (from 19% of subject-months at 1 month of suppression to 25% of subject-months at 12 months of continuous suppression) and 90–100% adherence (from 50% of subject-months at 1 month of suppression to 55% of subject-months at 12 months of continuous suppression). None of these changes in the proportion of participants at a given adherence level was statistically significant at the 0.05 level.

Of the 221 participants who achieved viral suppression during adherence monitoring, 108 (49%) subsequently experienced virologic failure during adherence monitoring. Based on univariate regression analyses, participants who had longer duration of suppression also had higher CD4+ counts, higher CD4+ nadir, and higher cumulative average adherence proportions. Participants who had longer duration of suppression were also more likely to be on an NNRTI or PI/NNRTI regimen than an unboosted protease inhibitor regimen. Virologic failure (defined as a plasma HIV RNA greater than 50 copies/ml) was associated with low recent CD4+ T cell count, low past adherence, shorter duration of prior suppression, and intravenous drug use.

A logistic regression of the probability of virologic failure on adherence, confounders, and number of months since initial suppression was used in the targeted maximum likelihood analysis (see Table 3). Based on this logistic regression, a lower risk of virologic failure was associated with longer duration of continuous suppression, lower past viral load, and being on an NNRTI-based regimen.

We first estimated the causal effect of different adherence proportions on the probability of virologic failure, conditioned on duration of continuous viral suppression (Figure 1). We stratified subject-months by corresponding duration of suppression and then estimated the probability of virologic failure setting the adherence proportion to be in the ranges (0–49%, 50–74%, 75–89%, and 90–100%). Because adherence proportion cannot be randomly assigned as an intervention, we relied on a marginal structural model and targeted maximum likelihood estimation of the causal effects of adherence on viral suppression at each time point.

At one month of suppression, the estimated probabilities of virologic failure at all four adherence categories were relatively high: 0.36 for 0–49% adherence, 0.49 for 50–74% adherence, 0.34 for 75–89% adherence, and 0.37 for 90–100%. The widths of the confidence intervals for these estimates were large (ranging from 0.24 to 0.43), and there was a substantial region of mutual overlap among all four confidence intervals.

The point estimates for the probability of virologic failure decreased with longer duration of suppression (Figure 1). The probabilities of virologic failure after 12 months of viral suppression were 0.24, 0.02, 0.06, and 0.01 for adherence categories of 0–49%, 50–74%, 75–89%, and 90–100%, respectively. Comparing the probability of failure just after achieving suppression versus after 12 consecutive months of suppression, there was a statistically significant decrease in the probability of virologic failure for adherence set above 50%. The estimated risk difference, comparing the probability of virologic failure after 1 month vs. after 12 months of continuous viral suppression was 0.47 (95% CI 0.23–0.63) for the 50–74% adherence stratum, 0.29 (95% CI 0.03–0.50) for 75–90% adherence stratum, and 0.36 (95% CI 0.23–0.48) for the 90–100% adherence stratum. The estimated risk of failure decreased by only 0.13 (95% CI −0.74–0.39) for the 0–49% adherence stratum, which was not statistically significant at the 0.05 level.

We carried out a test for effect modification by number of consecutive months virally suppressed. We considered the causal relative risk, comparing the effect of different adherence proportions on the probability of virologic failure. This causal relative risk was calculated conditioning on the number of months of viral suppression, using a marginal structural model. The null hypothesis that this causal relative risk does not depend on the number of consecutive months virally suppressed was rejected (p-value 0.001). Thus, the data provide evidence for duration of suppression being an effect modifier of the adherence-suppression relationship.