We conducted a retrospective analysis of longitudinal data of WHIV enrolled in WIHS between April 2014 to September 2019. WIHS is the oldest and largest prospective cohort of women living with or at risk for HIV [16, 17], which has now been combined with the Multicenter AIDS Cohort Study (MACS) to form the MACS/WIHS Combined Cohort Study (MWCCS). The main goal of WIHS is to investigate the natural history of HIV treatment and prevention in women in the United States. WIHS was founded in 1993, and women were recruited in four waves (1994-95, 2001–2002, 2011–2012, and 2013–2015). During the first three waves, participants were enrolled from the Bronx and Brooklyn, New York; Washington, DC; Los Angeles and San Francisco, California; and Chicago, Illinois. During the fourth wave, more participants were recruited from other research sites in Atlanta, Georgia; Chapel Hill, North Carolina; Miami, Florida; Birmingham, Alabama; and Jackson, Mississippi [18]. The Institutional Review Board approved this study at each of the study sites.
WIHS dataAt semi-annual study visits, data collection involved clinical exams, blood sample collection, and interviewer-administered questionnaires to collect basic sociodemographic, behavioral, and clinical data. This analysis included data on age (years), geographical region (Midwest, South, West, and Northeast), race (non-Hispanic white, non-Hispanic African American, and Hispanic of any race), educational level (below secondary, completed secondary, some college/completed college), household income (categorized here as <$24,000 vs. ≥$24,000), employment status (employed vs., unemployed), time since diagnosis with HIV (years), smoking status (women who never smoked, women who smoke, women who previously smoked), alcohol intake (abstainer, > 0–7 drinks/week, > 7–12 drinks/week, > 12 drinks/week), substance use at baseline (marijuana or hash, crack, cocaine, heroin, illicit methadone, methamphetamines, amphetamines, narcotics, hallucinogens, other drugs), and depression status (measured by the Center for Epidemiological Studies Depression (CES-D) Scale with a score of ≥ 16 indicating the presence of depressive symptoms and < 16 indicating no depression) [19]. It also included self-reported non-HIV and HIV medication use and adherence to ART. Self-reported ART adherence over the past month is categorized as “100% of the time”, “95–99% of the time”, “75–94% of the time”, “<75% of the time,” and “I have not taken any of my prescribed medications.”
Study participantsWe included WHIV enrolled in WIHS between April 2014 to September 2019 aged ≥ 18 years on ART, who had at least three self-reported adherence measurements and three visits with recorded data on non-HIV medications. Women with less than three adherence measurements and less than three visits with non-HIV medications recorded data were excluded due to the prerequisites for fitting GBTM. The first adherence visit between the above dates was designated as the “baseline” visit.
Data analysisPolypharmacyTo assess polypharmacy, we determined the number of non-HIV medications being taken at the time of each visit, including all reported prescribed and over-the-counter medications, herbal supplements, topical and ophthalmic, and as-needed medications. If a medication contained two or more pharmacologically active agents, each substance was counted individually in the analysis. Multivitamins that have multiple ingredients were counted as one. Herbals were counted as one regardless of the mixture. At each visit, polypharmacy was defined as the concomitant use of five or more non-HIV medications [6], and no polypharmacy was defined as zero to four non-HIV medications. Prescription-only polypharmacy was defined as five or more prescription-only medications, and no prescription polypharmacy was defined as zero to four non-HIV medications, excluding over-the-counter medications, herbal supplements, and as-needed medications. We defined polypharmacy as 5 or more medications because this threshold has been associated with poor treatment outcomes (e.g., disability, falls, frailty, and death) [20]. Furthermore, the definition has been used in most studies conducted among PWHIV to assess polypharmacy [21]. We assessed the pharmacologically active ingredients rather than the number of pills because the former is more comprehensive and can capture whether nonadherence to ART is due to pill burden or drug unintended effects.
Group-based trajectory modelingWe used GBTM to identify latent adherence and polypharmacy groups using a censored normal model and a logit model, respectively [22]. Each of the following steps was performed to identify the number and shape of trajectory groups for adherence and polypharmacy separately. Initially, the analysis procedure involved fitting several models sequentially to determine the appropriate number of trajectory groups. The second step entailed visual inspections and determining trajectory shapes considering constant, linear, quadratic, and cubic specifications. A set of criteria was considered to determine model fit namely, (1) Bayesian Information Criteria (BIC) with smaller values indicating better model fit, (2) the mean posterior probability of membership within each group (entropy) with values > 0.70 generally indicating acceptable classification, (3) the smallest group with at least 5% of the sample, (4) a tight confidence interval around estimated group membership probabilities and statistically significant groups (P < 0.05), and (5) parsimony in the model with few classes and parameters probabilities [13]. In addition, the model selection process was based on subject matter knowledge about the patterns of both variables and the interpretability of the model. The final step involved estimating a dual trajectory model using the univariate models that had been identified. We tested several models by varying the number of groups in each variable to ensure that the groups identified in univariate GBTM analysis in both adherence and polypharmacy were the best models for the joint analysis. The joint model summarized the interrelationships between adherence and polypharmacy trajectories as conditional probabilities of each variable on the other and their joint probabilities as well [13]. We assumed that the missingness was fully random, in which case GBTM would account for the missingness by fitting the model with maximum likelihood estimation and giving asymptotically unbiased parameter estimates [13].
Descriptive and comparative analysisFor each trajectory within adherence and polypharmacy, categorical variables were presented as numbers, and percentages and continuous variables were summarized as median and interquartile ranges (IQR). Comparisons were made across trajectories of adherence using the Chi-square test. We measured the association between the percentage of women who reported adherence at a level of ≥ 95% and the percentage of women on polypharmacy during the study period using the Pearson correlation method.
Predictors for membership in adherence trajectoriesWe fitted multinomial logistic regression analysis to identify predictors of the group membership of adherence, namely, age (years), race, educational level, annual income, alcohol intake, history of smoking, cumulative years in ART, depression, and substance use. We planned to use the group with the highest level of adherence probability as a reference group in the model. As an initial step, we performed univariable analyses, and covariates with P-values < 0.25 were selected to be included in the final model.
Sensitivity analysesIn a sensitivity analysis, we used the above-mentioned definition of prescription-only polypharmacy to classify women as having polypharmacy or not at every study visit. We assumed that the state of polypharmacy using prescription-only medications potentially affects ART adherence more than polypharmacy of a combination of prescription-only, over-the-counter, as-needed medications, and herbal supplements. We used GBTM to identify polypharmacy trajectories using the logit model, following the same statistical steps and criteria for identifying group numbers and trajectory shapes described above. Furthermore, we used GBTM to conduct another sensitivity analysis in which we considered the number of non-HIV medications as a continuous variable. We included all prescription-only, over-the-counter, as-needed medications, and herbal supplements in this analysis. We assumed that using a large number of non-HIV medications concurrently with ART at any given time would create a burden and, as a result, influence adherence. Using the same statistical procedure described above, we used the censored normal model to identify the number and shapes of trajectories.
Data analysis was conducted using Stata version 16, and the Stata Plugin was used to estimate GBTM parameters.
留言 (0)