Stimulant Use and AIDS Progression After HAART
Study Design and Procedures
Participants were enrolled in the MACS, an ongoing prospective study of HIV infection among gay and bisexual men as well as other men who have sex with men (MSM) in the United States. Enzyme-linked immunosorbent assays with confirmatory Western blot tests were performed on all participants at enrollment and every semiannual visit thereafter for initially HIV-negative participants. T-lymphocyte subsets were quantified using standardized flow cytometry, and HIV viral load was measured using standardized polymerase chain reaction methods. MACS protocols were approved by the institutional review boards of each of the participating centers. Informed consent was obtained from all participants.
This study included all MACS participants who initiated HAART, had at least 1 follow-up visit with assessment of stimulant use after initiation, and had data for covariates available within 2 years before stimulant use was assessed. This study used the data collected prospectively from MACS visits 26–57 (October 1996 to September 2012). Baseline was defined as the first visit after initiating HAART. The characterization of HAART regimens was guided by the DHHS/Kaiser Panel guidelines and defined as 3 or more antiretroviral drugs consisting of (1) one or more protease inhibitors, or (2) one nonnucleoside reverse transcriptase inhibitor, or (3) the nucleoside reverse transcriptase inhibitors—abacavir or tenofovir, or (4) an integrase or an entry inhibitor.
Outcomes: AIDS and All-cause Mortality
Using the 1993 Centers for Disease Control classification system, participants were assessed for an ADI during MACS visits. Participants met the criteria for AIDS if they were diagnosed with an ADI or had a CD4 cell count of less than 200 cells per microliter or CD4 cell percentage of <14. The date of AIDS diagnosis was confirmed through medical chart abstraction and interviews with medical providers. Using the National Death Index—Plus, final mortality information (including date and cause) was obtained for enrolled participants over the follow-up period.
Primary Predictor: Stimulant Use
Participants reported whether they had used methamphetamine, cocaine, crack cocaine, or ecstasy since their last MACS visit. Participants were categorized as reporting any stimulant use (1) or no stimulant use (0) at each visit. The time-varying cumulative proportion of MACS visits with any self-reported stimulant use was calculated. Compared with a reference group that reported no stimulant use (0%), patterns of intermittent (ie, 1%–49%, 50%–99%) and persistent (ie, 100%) stimulant use were characterized to investigate an expected dose–response association.
Demographics and HIV disease markers were included to adjust for possible confounding. Age at each MACS study visit was calculated using self-reported date of birth and treated as a time-varying continuous covariate (centered at 50 years). Self-reported race/ethnicity was categorized as white (reference group), African American/black, and Hispanic/Latino or other ethnic minority. Self-reported highest level of education completed at enrollment was categorized as high school or less (reference group), some college (grades 13–15), and college graduate or greater (grade 16 or more). CD4 cell count before HAART initiation was measured using peripheral venous blood samples for the MACS visit before starting HAART. For those who began HAART before enrolling in the MACS, pre-HAART CD4 cell count was measured using medical chart abstraction. Participants with a pre-HAART CD4 cell count of greater than or equal to 500 cells per microliter (reference group) were compared to those with 499–350, 349–200, and <200 cells per microliter. Time-varying CD4 cell count, log10HIV viral load, and self-reported HAART adherence were lagged (approximately 6 months) to adjust for time-dependent confounding with stimulant use. Self-reported HAART adherence was measured using a single item where participants indicated how often they took HAART medications as prescribed by selecting one of the following options: 100%, 95%–99%, 75%–94%, and <75%.
Health status indicators and behavioral factors were measured as possible confounders. Hepatitis C virus (HCV) coinfection was defined as antibody or RNA positive at baseline. Other lagged time-varying health status indicators measured at each visit included body mass index (weight in kilograms/height in square meters), high blood pressure (ie, systolic blood pressure ≥140 mm Hg or diastolic blood pressure ≥90 mm Hg or diagnosed with hypertension and use of antihypertensive medications), and dyslipidemia (ie, fasting total cholesterol ≥200 mg/dL or low-density lipoprotein ≥130 mg/dL or high-density lipoprotein <40 mg/dL or triglycerides ≥150 mg/dL or use of lipid-lowering medications with self-report of clinical diagnosis in the past). Lagged, time-varying self-reported physical health and mental health were assessed using the SF-36 Physical Component Summary and Mental Component Summary scores. Participants with scores of 16 or greater on Centers for the Epidemiologic Study of Depression scale were categorized as reporting clinically significant distress, which was examined as a lagged time-varying covariate. Finally, binge drinking (ie, ≥5 alcoholic drinks per day for at least once a month) and cigarette smoking in the last 6 months were included as time-varying covariates.
Marginal Structural Model Analyses
Because we were concerned that declining health might lead to a subsequent reduction in the use of stimulants (ie, a "sick quitter" effect), marginal structural modeling was used to address time-dependent confounding. This requires an initial, longitudinal unordered multinomial logistic model with time-varying stimulant use [4 categories: 0% (reference), 1%–49%, 50%–99%, and 100%] as the outcome to obtain stabilized weights for the final weighed models. The logistic model for determining the numerator of the weights for stimulant use included all time-fixed covariates (ie, site, race/ethnicity, education, baseline HCV status, and pre-HAART CD4 cell count), number of visits from baseline, and cumulative percent of the previous 3 visits where any stimulant use was reported. To obtain the denominator of the weights for stimulant use, all time-varying covariates were also included (ie, age, CD4 cell count, HIV viral load, self-reported HAART adherence, body mass index, high blood pressure, dyslipidemia, self-reported physical health, self-reported mental health, Centers for the Epidemiologic Study of Depression scores ≥16, binge drinking, and cigarette smoking). Similarly, a second logistic model determined the weights of remaining uncensored to control for informative dropout. The final stabilized weights were calculated by multiplying the weights of stimulant use and weights of remaining uncensored. If the weights were >4, they were set to 4.
The primary analyses consisted of separate weighted pooled logistic regression models for time to all-cause mortality and time to AIDS or all-cause mortality. Pooled logistic regression is a standard method for the analysis of discrete-time survival data, involving expansion of the binary outcome data to reflect a time-to-event outcome. A weighted competing risk analysis was also performed with a pooled multinomial logistic model (ie, alive, AIDS-related mortality, and non-AIDS mortality) to examine the association of stimulant use with AIDS-related and non-AIDS mortality separately. This was a discrete version of a competing risk analysis based on cumulative incidence functions.