Background:  Although depression is known to worsen the course of HIV, less is known about whether antidepressant medication treatment improves HIV outcomes. Prior studies have inadequately adjusted for time-varying confounding by indication.

Methods:  Participants were selected from a cohort of HIV⁺ homeless and marginally housed adults living in San Francisco who had been started on HAART during follow-up. To estimate the effect of antidepressant treatment on log₁₀ viral load, a marginal structural model (men who have sex with men [MSM]) was fitted using ordinary least squares (OLS) regression and weighted by inverse probability-of-treatment-and censoring weights. The treatment estimate was adjusted for demographic characteristics at baseline and these time-lagged confounders: substance abuse, homelessness, any mental health system contact, depression severity, and diagnosis of major depression. An equivalent MSM was fitted using logistic regression to examine probability of viral suppression. To determine whether the effect of antidepressants was mediated by adherence to HAART, 7-day self-reported adherence was entered into the MSM and the change in the treatment coefficient re-assessed. A sensitivity analysis used adherence as measured by unannounced pill counts (available for a subset of participants).

Results:  We included 418 participants (1330 person-years of follow-up). Unweighted OLS regression showed a ‑0.56 log₁₀ difference (95%CI, –1.04 to –0.08; p = 0.02) in viral load after antidepressant treatment, compared to no treatment. Inverse probability-of-treatment-and censoring-weighting showed a –0.83 log₁₀ difference (95%CI, –1.62 to –0.04; p = 0.04). When 7-day self-reported adherence was added to the MSM, the estimated log₁₀ viral load difference was –0.11 (95%CI, –0.92 to 0.71; p = 0.79). Pill count adherence data were available for 296 participants (258 person-years of follow-up). When this variable was used in a sensitivity analysis, the estimated log₁₀ difference was –0.15 (95%CI, –1.07 to 0.77; p = 0.75). Logistic regression showed a non-statistically significant change in the probability of viral suppression, in both unweighted (OR 1.29; 95%CI 0.96 to 1.74) and weighted (OR 1.44; 95%CI 0.86 to 2.39) analyses; these estimates were further reduced in statistical significance by the addition of either adherence measure to the MSM.

Conclusions:  Antidepressant medication treatment results in a statistically significant reduction in viral load. This effect is mediated largely through improved adherence to HAART.