Data for these analyses were collected through the AIDS Care Cohort to Evaluate access to Survival Services (ACCESS), an ongoing community-recruited prospective cohort study of HIV-positive IDU which has been described in detail previously (15
). In brief, beginning in May 1996, participants were recruited through self-referral and street outreach from Vancouver’s Downtown Eastside, the local epicenter of drug-related transmission of HIV. At baseline and semi-annually, all HIV-positive participants provided blood samples and completed an interviewer-administered questionnaire. The questionnaire elicits demographic data as well as information about participants’ drug use, including information about type of drug, frequency of drug use, involvement in drug treatment and periods of abstinence. All participants provide informed consent and are remunerated $20 CDN for each study visit. The study is somewhat unique in that the province of British Columbia not only delivers all HIV care free of charge through the province’s universal healthcare system but also has a centralized HIV treatment registry. This allows for the confidential linkage of participant survey data to the Drug Treatment Program at the BC Centre for Excellence in HIV/AIDS to a complete prospective profile of all HIV-related clinical monitoring and antiretroviral dispensation records. The Providence Health Care/University of British Columbia Research Ethics Board reviewed and approved the ACCESS study.
Participants were eligible for the present analysis if they initiated antiretroviral therapy between May 1996 and December 2009. The primary outcome in this study was adherence to antiretroviral therapy based on a previously-validated measure of prescription refill compliance (22
). Specifically, using data from the centralized ART dispensary, we defined adherence as the number of days for which ART was dispensed over the number of days an individual was eligible for ART in the year after ART was initiated. This calculation was restricted to each patient’s first year on therapy to limit the potential for reverse causation that could occur among patients who cease antiretroviral therapy after they have become too sick to take medication (17
). We have previously shown this measure of adherence to reliably predict both virological suppression (19
) and mortality (22
). As in previous studies, adherence was dichotomized as ≥95% versus <95% (19
). As an initial analysis, we calculated the proportion of individuals achieving at least 95% adherence to prescribed therapy in the year following initiation of ART during each year from 1996 to 2009 and used the Cochrane-Armitage test for trend to assess if rates of adherence changed over time.
We then examined factors independently associated with 95% adherence using logistic regression modeling and were specifically interested if year of ART initiation was associated with adherence after adjustment for potential confounders. We considered explanatory variables potentially associated with 95% adherence including: gender (female vs. male); age (<24 yrs. vs. ≥24 yrs.); ethnicity (Aboriginal ancestry vs. other); daily heroin injection (yes vs. no), daily cocaine injection (yes vs. no); daily crack cocaine smoking (yes vs. no); methadone use (yes vs. no); any other addiction treatment use (yes vs. no); and unstable housing (yes vs. no). Age was defined as a dichotomous variable according to the World Health Organization’s definition of a ‘young person’, using the upper age limit of 24 as the cut-off (25
). All dichotomous behavioural variables referred to the six-month period prior to the interview. As in our previous work (26
), we defined unstable housing as living in a single-room occupancy hotel, shelter or being homeless. Clinical variables included baseline HIV-1 RNA level (per log10
copies/mL) and CD4 cell count (per 100 cells/mm3
To estimate the independent relationship between calendar year and likelihood of 95% adherence to prescribed ART, we fit a multivariate logistic regression model using an a-priori defined protocol suggested by Greenland et al (27
). First, we fit a full model including the primary explanatory variable and all secondary variables with p < 0.20 in univariate analyses. In a manual stepwise approach we fit a series of reduced models by removing one secondary explanatory variable, noting the change in the value of the coefficient for the primary explanatory variable. We then removed the secondary explanatory variable associated with the smallest absolute change in the primary explanatory coefficient. We continued this process until the maximum change from the full model exceeded 5%. This technique has been used in a number of studies to best estimate the relationship between an outcome of interest and a primary explanatory variable (28
). All statistical procedures were performed using SAS version 9.1 (SAS, Cary, NC, USA). All p
-values are two-sided.