Effects of drug and hazardous alcohol use on having a detectable HIV viral load: An adherence mediation analysis

Highlights • People living with HIV (PWH) with substance use or alcohol use often have detectable plasma HIV viral loads that are predictive of negative clinical outcomes.• Among PWH with substance or alcohol use, it is not known what proportion of the observed detectable HIV viremia is mediated through medication nonadherence.• Use of amphetamines, opioids, and multiple substances but not alcohol predicted having a detectable HIV viral load.• Up to 40% of the substance use effects were mediated by self-reported adherence.


Introduction
The potential negative effects of active alcohol and substance use on achieving and maintaining an undetectable HIV plasma HIV load (pVL) have been a concern since the advent of antiretroviral therapy. A common assumption is that any putative effects of alcohol or substance use on HIV viral control are mediated through effects on antiretroviral medication nonadherence (Amico et al., 2006;Chander, Lau, & Moore, 2006;Hendershot, Stoner, Pantalone, & Simoni, 2009). Regimens with higher resistance barriers, simplified dosing, and long half-lives may attenuate the detrimental effects of alcohol and substance use on adherence-mediated and adherence-independent causal pathways toward HIV viral load suppression (Christopoulos, Grochowski, Mayorga-Munoz, Hickey, Imbert, Szumowski, Dilworth, Oskarsson, Shiels, Havlir, & Gandhi, 2022).
Few studies have examined adherence-mediated and direct effects of hazardous alcohol and substance use on HIV viral suppression (Cook et al., 2017;Feelemyer et al., 2020;Kahler et al., 2017;Parsons, Rosof, & Mustanski, 2008). We hypothesized that the effects of alcohol and substance use on viral suppression would be largely mediated through effects on antiretroviral nonadherence. The primary study aim was to determine the proportion of the total effect of substance use (including alcohol) on having a detectable viral load that was mediated through antiretroviral adherence. Our secondary aim was to determine the proportion of patients having an undetectable pVL according to the frequency of substance use and self-reported adherence category.

Study design and population
We conducted a retrospective cross-sectional analysis of persons with HIV (PWH) receiving care in an urban academic HIV clinic (Owen Clinic) at the University of California, San Diego (UCSD), in the United States, between 2014 and 2018. Our cohort included PWH 18 years of age or older with prescribed antiretroviral therapy who completed patient-reported outcome (PRO) questionnaires at least annually and had available the most recent, subsequent HIV pVL collected as part of routine care. Patients' characteristics included sex at birth, race, HIV risk factor, age, and study year and were collected using de-identified electronic medical records information. All participants signed a written consent form before enrolment in the study. The study protocol was approved by the UCSD Human Subjects Research Protection Program (UCSD HRRP #171338).

Data collected domains and procedures
The PRO questionnaire contains self-report items to ascertain alcohol use, substance use, and HIV antiretroviral adherence. Substance use measures included AUDIT-C and NIDA-ASSIST as measures of hazardous drinking over the past year (AUDIT-C ≥ 4 in men and ≥3 in women) and recent (within the past 3 months) substance use, respectively (Bradley et al., 2007;WHO Assist Working Group, 2002). Categories of prescription and recreational drugs surveyed in the PRO battery included cannabis, amphetamines (including methamphetamine), opioids (both legal and illegal), sedative-hypnotics, cocaine, inhalants, and hallucinogens. Self-report of adherence was assessed in four ways: (1) the number of doses missed in the last 7 days (1 = 0 missed, 6 = >4 missed); (2) Likert scaled adherence (1-6) where 1 = very poor and 6 = excellent; (3) visual analog scale (VAS) of percent adherence over the past month; and (4) last missed dose (1 = within the past week, 6 = never skip medications) (Amico et al., 2006;Feldman et al., 2013;Simoni et al., 2006;Stirratt et al., 2015).

Statistical methods
To select a single measure of medication adherence that most discriminated between those with detectable and undetectable followup HIV viral load measurements, we conducted receiver operating characteristic (ROC) analysis of the four candidate measures using binary coded viral load (≤200 copies/ml vs. >200 copies/ml) as the reference criterion. We chose 200 copies/ml as reference criteria as this value matches the threshold for lack of HIV transmission to uninfected partners (Eisinger, Dieffenbach, & Fauci, 2019). To evaluate whether the four adherence measures could jointly enhance discrimination beyond any individual measure, a logistic regression model was fit, and the predicted probabilities of viral non-suppression were estimated. ROC areas and Hosmer-Lemeshow tests were used to evaluate model discrimination and calibration, respectively. The optimal cut-point of the most discriminating predictor of viral non-suppression was estimated using Liu's method implemented in Stata cutpoint.ado (Liu, 2012), StataCorp. 2016. Stata Statistical Software: release 15.1. College Station, TX, USA.
The most discriminating adherence measure was then dichotomized at the optimal cut-point and used in subsequent analyses as the operationally defined measure of non-adherence. Next, we examined bivariate associations: (1) between alcohol and specific substance use and the most discriminating binary measure of adherence and (2) between alcohol and specific substance use and HIV viral non-suppression. We then examined the effects of alcohol and specific substance use on having a detectable pVL using unadjusted and mutually adjusted logistic regression models. Substances (including alcohol) that were independent predictors of having a detectable pVL in logistic models were subsequently selected as independent predictors of HIV viral nonsuppression in the mediation analyses. Separate mediation models were fit for each selected substance using the Stata paramed.ado program to estimate the natural direct effect (NDE), natural indirect effect (NIE), and marginal total effect (MTE) on the odds ratio scale. Percent mediation was estimated as the ratio of the natural logarithms of odds ratios for the NIE/MTE (Emsley, Liu, & Dunn, 2013;Lee et al., 2021;Vanderweele & Vansteelandt, 2010). Sensitivity analysis controlling our mediation models for sex at birth, race, HIV transmission risk factor, and age was also conducted. We determined, based on the observed prevalence of unsuppressed pVL (10%), the prevalence of excellent adherence (58%), and the correlation between the drug or alcohol exposure measures and adherence varying from 0.1 to 0.5, that with type I error = 0.05 and power = 0.8, we could detect a mediation odds ratio from 1.4 to 1.49, respectively, with a sample size of 3000 (Vittinghoff, Sen, & McCulloch, 2009). Table 1 shows the characteristics of the study participants. During the study period, 3125 patients met the eligibility criteria. PWH had a median age of 51 years, of whom 54% were non-white, 12% were female, and 72% were men who have sex with men (MSM) as self-reported HIV risk factors. The prevalence of hazardous alcohol consumption was 26%, and use of prescription and recreational substances was common: marijuana (27%), amphetamine (13%), inhalants (12%), sedativehypnotics (5%), and opioids (3%). Multiple substance use was reported by 17% of participating PWH. The mean time from PRO to HIV viral load collection was 89 days (SD: 155 days). Despite prevalent substance use, most (90%) patients had undetectable HIV pVL (<200 copies/ml), and more than half (58%) reported excellent adherence to antiretroviral therapy. Table 2 presents the measurement properties of the four single-item adherence measures and the composite predicted adherence proportion based on logistic regression of the four single items on HIV pVL. Because the fewest non-missing values (n = 3110) were with item 2 and because the pVL discrimination (ROC area) of this item was highest among the single-item adherence measures, we chose this item for use in mediation analysis. The optimal antiretroviral adherence cut-point in predicting having a detectable pVL was at excellent (6) vs. not excellent (≤5). An adherence score ≤5 on this item was operationally defined as nonadherence. Table 3 presents the substance and pVL distributions according to the self-reported adherence category. Nonadherence was significantly associated with detectable pVL and the use of amphetamines, cocaine, opioids, marijuana, hallucinogens, and inhalants. Table 4 shows substance use and self-reported adherence according to pVL category. Having a detectable pVL was significantly associated with nonadherence and the use of amphetamines and opioids. In contrast, hazardous alcohol consumption and the use of marijuana, cocaine, inhalants, sedative-hypnotics, and hallucinogens were not predictive of detectable pVL.

Results
In both unadjusted and mutually adjusted logistic regression models, the independent use of amphetamines and opioids was significantly associated with having a detectable pVL (Table 5). PWH reporting amphetamine or opioid use were more than twice as likely to have a detectable pVL. The inability of binary-coded hazardous alcohol use to predict having a detectable pVL was confirmed in ROC analysis of the numeric AUDIT-C score evaluated against pVL as reference criterion (ROC 0.492, 95% C. I. 0.457-0.527). Table 6 presents mediation model results for substances that were found to be significant predictors of having a detectable pVL. For single substances, the percent mediation was highest (36%) for amphetamine use, while for opioids, the percent mediated was 27%. The percentage mediation for multiple substance use was 40%. After controlling for sex at birth, race, HIV transmission risk factor, and age, the percent mediation for amphetamine, opioids, and multiple substances did not change substantially and were 38%, 28%, and 40%, respectively (Supplementary Table). Examining missing value patterns across all variables included in any mediation analyses presented in Table 6, those with no missing values (complete cases, n = 2936) did not differ from those with one or more missing values (incomplete cases, n = 189) by year of measurement, sex, or age (p > 0.05). However, complete cases were more likely to identify as white versus non-white (47% vs. 37%, p =   0.002) and to report male sex with men and not injection drug use (MSM, not-IDU) as their HIV transmission risk factor (66% vs. 56%, p = 0.01).

Discussion
Our research found that recent use of amphetamines, cocaine, opioids, marijuana, hallucinogens, inhalants, hazardous alcohol use, and multiple substance use were significantly associated with less than excellent self-reported antiretroviral adherence. Although with variability in their conclusions, many studies have previously documented the detrimental effects of the use of specific recreational substances and alcohol on various adherence measures, including self-report and electronic monitoring approaches (Chander et al., 2006;Hendershot et al., 2009;Gonzalez, Barinas, & O'Cleirigh, 2011;Lai et al., 2020;Rosen et al., 2013;Socias & Milloy, 2018;Velloza et al., 2020). Most robust conclusions pertain to the use of amphetamines, cocaine, opioids, and alcohol. Whereas for cannabis use, the intensity and regularity of use may modify the effects on adherence (Bonn-Miller, Oser, Bucossi, & Trafton, 2014; Vidot, Lerner, & Gonzalez, 2017).

Strengths and limitations
Our study focus was on distal and more impactful outcomes of antiretroviral therapy, specifically HIV viral suppression. Self-reported use of amphetamines, opioids as individual substances and multiple substance use predicted having a detectable pVL. However, surprisingly hazardous alcohol use was not predictive.
Our study found that hazardous alcohol use was associated with worse ART adherence but not with having detectable HIV pVL as others have noted (Nolan et al., 2017). A large meta-analysis also found a significant and reliable association between alcohol use and antiretroviral medication nonadherence; however, it did not address the effects of alcohol on viral suppression (Hendershot et al., 2009). Important differences in study design, secular trends in the type of antiretroviral therapy, and instruments used to measure alcohol use could account for differences in our findings from other studies that observed hazardous alcohol use to be associated with detectable viral loads. Some were based on cross-sectional assessments (Conen et al., 2009). The ones with the longitudinal design were conducted in the late 1990s to early 2000s when antiretroviral therapy was poorly tolerated, requiring a significant pill burden, and many PWH with hazardous alcohol use avoided taking them while drinking (Chander et al., 2006;Conigliaro, Gordon, McGinnis, Rabeneck, & Justice, 2003). Our study, conducted in a more contemporary antiretroviral era, suggests that, despite imperfect adherence, PWH with hazardous alcohol use can take enough medications to keep their pVL undetectable. Nonetheless, we did not assess patterns of alcohol intake which have been shown to impact pVL, with those with binge drinking having more likelihood of having detectable pVL (Conigliaro et al., 2003). Like our findings, a prospective randomized controlled trial failed to show the same association between hazardous alcohol use and detectable pVL (Samet et al., 2005). Noteworthy, in many of these cohorts, either exposure was defined as any alcohol use or instruments other than AUDIT-C were used (Palepu et al., 2003). We believe that our inability to detect an effect of hazardous alcohol use on pVL was not attributable to the specific AUDIT-C cutoff score used. We assessed, using ROC analysis, the full numeric range of AUDIT-C scores and found that the measure did not discriminate between those with and without HIV viral suppression.
This study found that up to 40% of the effects of having a detectable pVL appeared to be mediated by our measure of medication adherence. Thus, mediation analyses restricted to substances that were associated with having a detectable pVL (amphetamines, opioids, and multiple substance use) did not support our a priori hypothesis that their effects on viral suppression would be largely mediated through antiretroviral adherence. There are several potential explanations for this observation. 1. Composite predicted adherence proportion based on logistic regression of items 1-4 on viral load indicator. 2. For items 2-4, undetectable viral load (<200 copies/ml) was coded 1. For items 1 and 5, undetectable viral load was coded 0.
This could represent a true finding, pointing to the effect of these substances not mediated through antiretroviral adherence. Another potential explanation is the occurrence of pharmacokinetic or pharmacodynamic drug-drug interactions that adversely influence antiretroviral effects independent of adherence behavior (Desai et al., 2020). Other possible explanations are related to measurement issues: the specific metric chosen for substance use (any past 3-month use on NIDA ASSIST), the timing of the substance use recall period with respect to subsequent viral load measurement (median 24 days), and use of an insensitive measure of medication adherence (although we selected the most discriminating of the candidate measures available (Table 2). Residual confounding and collider effects resulting in selection bias owing to missing values must also be considered. These issues merit further exploration and replication. Among the few published reports of mediation analyses examining the effects of alcohol or substance use on viral non-suppression, we identified 3 concerning effects of alcohol or methamphetamine use. All three reported that the predominant effects were direct and not mediated through selected measures of adherence. Kahler et al. used a longitudinal cohort design (n = 533 patients) to examine the extent to which the effect of 30-day heavy drinking on viral non-suppression was mediated by antiretroviral adherence (measured by the number of missed doses over a 3-day period) (Kahler et al., 2017). These investigators found a significant natural indirect effect (NIE) on the risk ratio scale (1.03 [95% CI: 1.00-1.05]), corresponding to a 20 percent mediation effect. Cook et al. reported results of a mediation analysis  using baseline data (n = 619 patients) for which the adherence measure was based on a 30-day number of days with missing doses and drinking measures included specifications for low, binge, and heavy drinking according to AUDIT-C criteria (Cook et al., 2017). For heavy drinking, the predominant effect on viral non-suppression was direct, with only 9.1% mediated through adherence. We were able to identify only one mediation analysis for the effects of methamphetamine use and antiretroviral adherence on HIV viral non-suppression (Feelemyer et al., 2020). Employing a cohort design (n = 645 patients), the investigators reported that the effect of methamphetamine use was predominantly direct. Unfortunately, the indirect effect was not reported. So, our findings of the predominant direct effects of amphetamine, opioid, and polysubstance use on viral non-suppression are consistent with previous reports.
Limitations of our findings are that they represent a single-center experience, a cross-sectional rather than a longitudinal design, and are based on a data set with missing values that limited complete case analysis for multivariable components of the analysis. The self-report nature of substance use and adherence might have potentially impacted the observed associations. Often, self-report substance use underestimates the use (Khalili et al., 2021), and self-reported adherence tends to overestimate adherence (Stirratt et al., 2015). Also, participant characteristics linked to social determinants of health, such as being unsheltered, mental health, and poverty, were not included in the analysis; hence, their contribution to observed effects could not be explored in this study. Our analysis mainly included white individuals and MSM, and our results might not be generalizable to HIV cohorts with more diverse characteristics. Nonetheless, we conclude that the mediation analysis findings are provocative and merit examination in expanded longitudinal cohorts with longitudinal mediation modeling (O'Laughlin, Martin, & Ferrer, 2018).

Conclusion
In summary, in a cohort of 3125 PWH, the use of amphetamines, opioids, and multiple substances predicted a detectable plasma HIV viral load. Up to 40% of the effects were mediated by self-reported adherence.

Funding
This study is unfunded. The Clinical Investigation Core of the San Diego Center for AIDS Research (CFAR) provided administrative assistance in retrieving data (AI036214).

Data Access Statement
Data supporting this study are not publicly available but can be accessed upon request. Please contact the corresponding author: eca-chay@health.ucsd.edu.

Role of the sponsor
The sponsor had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Previous presentation
None.

Declaration of Competing Interest
The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: EC received research grant funding paid to the Regents of the University of California from Gilead and Merck, and consulting fees for an advisory board from Gilead for unrelated projects. The other authors have no reported conflicts of interest.

Appendix A. Supplementary material
Supplementary data to this article can be found online at https://doi. org/10.1016/j.abrep.2023.100486.