Methods
Source Population
The analysis used longitudinal pharmacy refill data collected prospectively from HIV-positive persons on HAART and followed in the Veterans Aging Cohort Study Virtual Cohort from October 1, 2000 to September 30, 2010. Details of the Veterans Aging Cohort Study Virtual Cohort have been previously described. Laboratory and clinical data and outpatient prescriptions for each subject were obtained by linking Immunology Case Registry and Pharmacy Benefits Management Registry records, respectively. HAART was defined using the Department of Health and Human Services (DHHS) guidelines. Only person-years in which HAART was used for at least 180 days in the year were included.
For each person-year, we used the regimen most frequently refilled to classify HAART as NNRTI-based, PI-based (including users of PIs, and both NNRTIs and PIs), INSTI-based, or 3 nucleoside reverse transcriptase inhibitors containing abacavir or tenofovir. We classified regimens as being single versus multi-pill and whether administered once-daily versus twice-daily.
Outcomes and Exposures
Because HIV RNA levels were determined using assays with varying detection limits, we used values of <400 copies per milliliter as nondetectable viral load and used the last HIV RNA test of the year for analyses. Sustained suppression was examined among those with multiple viral load measurements in a year and was defined as having undetectable levels following their first measurement if suppressed.
We calculated adherence to HAART using the medication possession ratio defined by Steiner and Prochanska, which measures the duration of time the patient had the medications available, relative to the total number of days between refills. This was calculated for each person-year that contained at least 1 refill as follows:
We excluded stockpilers (20.2% of study population), defined as person-years with a refill frequency exceeding the scheduled dosing interval by more than 5%, because the Steiner algorithm was not validated in this subgroup.
Potential confounders of viral load suppression and adherence included sociodemographic, behavioral, disease, and treatment characteristics. Fixed characteristics included race, smoking, and geographical location obtained at the first time seen after October 1, 2000 (baseline). Time-varying factors for each year included alcohol abuse, drug abuse, and major depression obtained using the International Classification of Diseases, Ninth Revision (ICD-9) diagnosis codes recorded at 1 inpatient visit and 2 outpatient visits, the number of antiretrovirals used, number of days in possession of HAART regimens, regimen type, time since first HAART initiation, and mean CD4 cell count.
Statistical Methods
We graphically depicted temporal trends of adherence, suppression, regimen type, and dosing frequency from 2001 to 2010. The change in adherence over time was determined using linear mixed-effects models with adherence percent as outcome, accounting for repeated measures over time, and adjusting for confounders. In sensitivity analysis, we restricted the entire population to (1) those who were in follow-up after January 1, 2009 (ie, including those starting before or after 2009, but in follow-up between 2009 and 2010) to avoid a biased temporal trend because of earlier attrition of those with worse outcomes from low adherence and (2) person-years on the first HAART regimen because switching regimens may not be random and may result from lower adherence and drug resistance.
We defined the minimum optimal adherence as the level of adherence at which the odds of suppression were not statistically different from that observed among those with ≥95% adherence. To focus on newer HAART regimens, we restricted this analysis to data from 2006 onward and used logistic regression with viral load suppression as the outcome and adherence percent as the primary exposure controlling for repeated measures over time and adjusting for confounders. Because characteristics informing prescribing patterns may affect adherence and HIV RNA suppression, we adjusted for this possible confounding by indication using propensity scores to weight the repeated-measures logistic regression model. The propensity score for using an NNRTI-based regimen was determined by logistic regression, which included age, race, geographical location, time since first HAART initiation, and CD4 count, HIV RNA suppression, drug abuse, alcohol abuse, and major depression diagnosis lagged to the previous year. Using the propensity score, weights were generated as the average treatment effect for the treated (ATT) and included in the repeated-measures logistic regression model as a covariate.
In sensitivity analyses, we varied the restriction on the number of days on HAART in the year to 270 and 330 days. All analyses were performed using SAS 9.2 (SAS Institute, Inc., Cary, NC) and STATA 12.1 (StataCorp. 2011, Stata Statistical Software, College Station, TX); P < 0.05 was used to define statistical significance.