Hendershot, Christian S PhD*; Stoner, Susan A PhD†; Pantalone, David W PhD^‡; Simoni, Jane M PhD*

INTRODUCTION

Since its widespread introduction in 1996, highly active antiretroviral therapy (HAART) has led to marked improvements in immunologic and virologic outcomes, quality of life and longevity among individuals living with HIV.^1,2 However, the optimal benefits of HAART are closely tied to adherence.³ The implications of nonadherence for disease progression are well documented and include adverse consequences for the individual and public health.^4-6 Improving the ability to identify and remediate barriers to HAART adherence is therefore a clear priority of behavioral HIV/AIDS research.⁷

Among the most frequently studied correlates of HAART nonadherence is substance use.⁸ Historically, most attention has focused on injection drug use (IDU),^9-11 but studies focusing on alcohol have been emerging with increasing frequency. A focus on drinking behavior in the context of HAART is warranted for several reasons. Alcohol use is prevalent among HIV-positive individuals,^10,12-14 is cited by patients as a reason for nonadherence;^15-18 and is associated with nonadherence in numerous studies. Evidence also suggests a deleterious influence of alcohol use on markers of immunological functioning and viral suppression,^9,19-22 effects that could be partly mediated by nonadherence.⁹ Given these findings and the fact that alcohol use is a modifiable behavior, interventions targeting alcohol use among those living with HIV/AIDS have the potential to improve disease management and perhaps to delay disease progression.^16,18,22,23

Although alcohol use has been associated with HAART nonadherence in numerous studies, the nature, strength, and consistency of this association remain unclear. A number of null findings have been reported, leading to a somewhat equivocal literature.^8,24 Additionally, the question of whether adherence is progressively compromised as drinking levels increase (ie, a “dose-response” effect) has not been systematically evaluated, in part due to the use of dichotomous drinking variables in most studies. Of the few studies to examine multiple levels of alcohol use, some have found a positive and linear relationship between drinking and medication nonadherence,^9,25,26 whereas others have found similar adherence rates across moderate and high drinking levels.^14,22,27 Finally, some evidence suggests that the relation of alcohol use and nonadherence could be moderated by demographic or methodological variables, including gender^28,29 and how adherence is defined.^16,30-32 There have not been systematic efforts to evaluate hypothesized moderators of the alcohol-adherence association.

There are several potential explanations for heterogeneous findings across studies, including insufficient power in smaller studies and variable methodological and measurement approaches across investigations.²⁴ Discrepant findings could also reflect the influence of moderating variables that have yet to be systematically evaluated. It is also noteworthy that existing studies are largely based on retrospective survey data³³ and have often used alcohol and adherence measures that were not temporally concurrent. These methodological features are significant barriers to inferring causal associations³⁴ and leave open the possibility that some reported findings reflect the influence of third variables.¹⁶ In sum, a detrimental influence of alcohol use on HAART adherence, though plausible and supported by some empirical data, is yet to be clearly, convincingly, or reliably characterized. A causal effect of alcohol consumption on HAART adherence could be better substantiated if this relationship proved consistent and robust across studies, especially given the varied methods reported in the literature.

The current study provides a meta-analytic review of published studies examining the association of alcohol use and antiretroviral adherence. Meta-analytic techniques are well suited for maximizing statistical power by aggregating effect sizes across studies and allowing examination of effect moderators at the aggregate level. This study had the following specific aims: (1) to provide a descriptive account of studies examining alcohol-adherence associations; (2) to provide quantitative estimates of the magnitude and stability of alcohol-adherence associations across studies; (3) to provide an aggregate evaluation of “dose-response” effects by examining effect sizes across objectively defined categories of consumption; and (4) to evaluate methodological and demographic variables as moderators of the alcohol-adherence association.

METHODS

Study Selection

We aimed to identify published, peer-reviewed studies that reported on the association of alcohol consumption and HAART nonadherence. Both electronic and manual literature searches were conducted to identify candidate studies published from 1996 through the end of 2007. Searches of electronic databases (MEDLINE, PubMed, PsycInfo) were conducted by crossing terms related to HIV (HIV; AIDS), antiretroviral therapy (HAART; antiretroviral therapy; antiretrovirals; ART; combination therapy; HIV treatment), medication adherence (adherence, nonadherence, compliance, noncompliance), and alcohol (alcohol, drinking, substance use, drug use, AOD [alcohol and other drug use]). All resulting abstracts were screened. Candidate studies were defined as those that reported results from a data-based study and indicated that adherence and alcohol or substance use were assessed.

Results of the study selection process are depicted in Figure 1. All candidate studies were retrieved for full review, and studies were retained for inclusion in the meta-analysis if they met the following criteria: (a) a measure of antiretroviral adherence was included in the study; (b) a measure of participants' alcohol use was included; (c) reported a quantitative test of the association of alcohol use and adherence; (d) provided sufficient statistical information for deriving an effect size; and (e) reported on a sample that did not overlap with that of another qualifying study. Studies that combined measures of alcohol with other substance use, or that combined measures of antiretroviral adherence with other medication adherence, were not considered for inclusion. In cases where 2 or more studies had overlapping samples, only 1 study was retained; we used sample size and specificity of drinking measures as the primary criteria for choosing among them. After arriving at an initial list of qualifying studies, we implemented 2 manual search strategies, which consisted of reviewing references cited in each of the qualifying studies (as defined above) and consulting review articles on correlates of antiretroviral adherence to identify additional candidate studies. Forty published studies were retained for the meta-analysis.^{9,14,16,17,24,27-33,35-61}

Figure 1

Data Abstraction and Coding

All studies were coded on attributes within 3 categories: descriptive characteristics, characteristics of adherence measurement, and characteristics of alcohol use measurement.

Descriptive characteristics included study location and year; gender composition; study aim (defined as whether alcohol/substance use was assessed as a primary or ancillary aim of the study); and sample size for the alcohol-adherence analysis. In cases where the number of participants for the alcohol-adherence analysis was smaller than that of the overall sample, sample size was coded based on the former. Studies from the United States were coded on racial composition of the sample using the 2 most commonly reported demographic groups (ie, percentage of African American and percentage of white participants). Non-US studies typically lacked data on race/ethnicity. Given the association of IDU with HAART adherence, we also estimated the proportion of each sample that reported IDU using available indicators; these varied across studies (eg, pastmonth IDU; lifetime IDU; IDU as the mode of HIV transmission).

For adherence measurement, studies were coded based on the assessment method (self-report, Medication Event Monitoring System [MEMS], etc.) and the length of time over which adherence was recalled or otherwise measured (eg, 3 days, 7 days, etc.). Studies also were coded based on whether adherence was analyzed as a continuous (eg, percentage of doses taken) or dichotomous (eg, adherent vs. nonadherent) outcome. For dichotomous outcomes, criterion cutoffs for adherence (eg, 95%) were recorded. With regard to alcohol use, studies were coded on the length of the alcohol use assessment period and whether the authors reported use of a previously validated measure to assess alcohol use or alcohol use disorders. The most frequently reported instruments/methods used to infer alcohol use disorders were the CAGE⁶² (4 studies), the Alcohol Use Disorders Identification Test (AUDIT)⁶³ (4 studies), and DSM-IV criteria (2 studies). Each effect size was coded based on the type of drinking examined in relation to adherence. Specifically, effect sizes were categorized as assessing (a) drinking frequency (eg, days per week); (b) drinking quantity (eg, drinks per drinking day); (c) drinking frequency and quantity (eg, 3 + drinks at least twice a week).

Our aim to examine a possible “dose response” effect of alcohol on adherence was complicated by the fact that the majority of studies used dichotomous drinking categories with highly variable cutoffs and/or definitions. Therefore, we created an objective index of drinking severity that was standardized across studies. Each effect size was coded on a variable reflecting drinking intensity, which included 3 levels-problem drinking, moderate drinking, and any or global drinking. A 3-level variable was chosen to allow evaluation of possible dose-response effects although retaining enough effect sizes per category for meaningful comparisons.

An effect size was coded as assessing problem drinking if it was derived from a dichotomous measure of alcohol use with a threshold that (a) met or exceeded the National Institute on Alcohol Abuse and Alcoholism (NIAAA) definition of at-risk drinking, using the threshold for men (>14 drinks per week or >4 drinks in a day),⁶⁴ or (b) met criteria for a probable alcohol use disorder based on diagnostic or screening criteria. For instance, a study that categorized participants based on whether they drank at least 5 drinks per drinking occasion, consumed at least 20 drinks per week, or had a CAGE score of 2+ or an AUDIT score of 8+ would contribute an effect size to the problem drinking category. Effect sizes were coded as assessing moderate drinking when based on a dichotomous outcome with a clearly defined threshold that reflected moderate drinking levels that did not exceed NIAAA criteria for at-risk drinking. For example, studies that categorized participants on whether they drank at least 2 times per week or at least 10 drinks per month were assigned to this category because these levels fall below NIAAA-defined at-risk drinking. Effect sizes were coded as any or global drinking if they (a) used categories that were exceedingly broad (eg, any alcohol use in the past month vs. none), (b) used a single, continuous drinking measure of drinking (eg, drinking days in the past 2 months), or (c) used indicators that were so vague as to preclude clear assignment to another category. A separate category for heavy episodic (binge) drinking (eg, 5 + drinks on 1 occasion) was not included because few studies reported this outcome and because our problem drinking category encompassed this definition.

Some studies had more than 1 effect size because they compared more than 2 drinking patterns/categories. In these instances, we selected the effect size that represented the most extreme comparison between drinking levels for the primary analysis to avoid violating the independence of effect sizes. However, we also conducted separate analyses that included (a) the least extreme comparison per study and (b) an average effect size for those studies with multiple effect sizes (described below).

Studies were also coded based on whether the reported alcohol-adherence association was determined using unadjusted or adjusted odds ratios because these relations could differ after controlling for other variables. Finally, studies were coded based on the degree of temporal overlap among the alcohol use and adherence measurements (ie, the degree to which these variables were assessed over the same time interval). This variable was included because temporal contiguity between the predictor and outcome variable has important implications for inferring causal relationships and could influence the pattern of effect sizes across studies. Temporal overlap was coded from 0 to 3. A score of 0 indicated that the alcohol and adherence measures had no overlap (or it was impossible to determine overlap based on the study description). A code of 1 indicated some but minimal overlap among assessment intervals. A code of 2 was used if there was evidence that the measures overlapped at least partially, and a code of 3 indicated that the assessment intervals were identical or almost identical.

Descriptive Information

The 40 studies included in the meta-analysis spanned a 10-year period (1998-2007) and totaled over 25,000 participants. Most studies (33) were conducted in the United States; other locations were France (3), Canada (1), Brazil (1), India (1), and Italy (1). Examining alcohol/substance use was identified as a primary study aim in 9 studies. Typically, alcohol use was one of several demographic or behavioral variables that were assessed and was not a major focus of the report. Twenty-one studies were derived from prospective cohort investigations, 16 from cross-sectional studies, and 3 from clinical trials. Of 24 studies derived from prospective designs (prospective cohort or clinical trial), 9 used prospective data to test the alcohol-adherence association. Table 1 presents descriptive information and results of the alcohol-adherence analysis for each study.

Table 1

Table 1

Table 1

Table 1

Table 1

Table 1

Table 1

Table 1

Table 1

Table 1

Table 1

Table 1

Table 1

Table 1

Effect-Size Abstraction

Effect-size data were entered and standardized before analysis. Odds ratios (ORs) were chosen for the effect size metric because the majority of studies compared 2 categories of alcohol use (eg, drinkers vs. nondrinkers; 72.5% of studies) on a dichotomous adherence indicator variable (eg, adherent vs. nonadherent, 77.5% of studies). Nine studies (20.5%) included multiple comparisons involving an alcohol-related indicator (eg, any use vs. none in addition to problematic vs. nonproblematic use or none). In order not to violate the assumption of independence of effect sizes, we did not include more than 1 comparison per study per analysis. Two alternative ways to extract effect sizes from these studies were explored: selecting the least extreme comparison (alternative 1; eg, any vs. no alcohol) and computing the average effect size (alternative 2). When studies included multiple adherence outcomes (eg, proportion of doses and timing of doses), 1 measure reflecting proportion of doses taken was selected because this was the primary outcome for most studies. If more than 1 adherence cutoff (eg, 90% and 100%) was evaluated in relation to drinking, the more stringent cutoff was used.

For each effect-size estimate, we computed a 95% confidence interval (CI), Z statistic and P value. Figure 2 presents the effect-size data and the forest plot. For each study, the forest plot indicates the point estimate for the study's effect size (OR) and its 95% CI. The size of the point represents the weight of the study in the context of the present meta-analysis. An OR of 1 indicates no effect of alcohol on adherence, ORs greater than 1 would indicate a benefit of alcohol on adherence, and ORs less than 1 indicate a detriment of alcohol on adherence. An initial test of homogeneity of variance indicated heterogeneity across samples, Q (39) = 82.235, P < 0.001; therefore, random effects models were used. The core analysis yielded an estimate of the overall effect. A sensitivity analysis was conducted whereby the overall effect was computed with each study removed in turn. A stratification analysis was conducted to examine the 3-level variable reflecting drinking intensity (any/global use, moderate use, problem drinking). Moderator variables were tested using linear meta-regression. Finally, for the overall effect, publication bias was evaluated via inspection of a funnel plot and the “Trim and Fill” method.⁶⁵

Figure 2

RESULTS

Core and Sensitivity Analyses

Under the random effects model, the point estimate and 95% CI for the combined studies was 0.548 (0.490 to 0.612), Z = −10.633, P < 0.001, indicating that those who used alcohol, or who drank relatively more, were 0.548 times as likely to be classified as adherent as compared with nonusers, or those who drank relatively less. Sensitivity analysis found no individual effect size to unduly influence the estimate of the overall effect; therefore, all were retained. The primary analysis used the most extreme comparison for those studies that included comparison of multiple drinking levels. As might be expected, alternative methods of effect-size extraction altered the estimate of the overall effect. Using alternative 1, the overall OR was estimated to be 0.628 (0.568 to 0.695), Z = −9.042, P < 0.001, indicating that when using the least extreme comparison per study, drinkers were 0.628 times as likely to be classified as adherent compared with nonusers. Using alternative 2, the overall OR was estimated to be 0.586 (0.531 to 0.647), Z = −10.647, P < 0.001, indicating that, when collapsing across multiple comparisons per study, alcohol users were 0.586 times as likely to be classified as adherent as nonusers.

Stratification Analysis of the Effect of Drinking Intensity

Participants classified as problem drinkers, defined in accordance with NIAAA guidelines for at-risk drinking or based on meeting diagnostic criteria for a probable alcohol use disorder, were 0.474 (0.408 to 0.550) times as likely as nonproblem drinkers or abstainers to be classified as adherent (14 effect sizes, Q (13) = 13.034, Z = −9.803, P < 0.001). Among studies examining drinking thresholds classified as moderate (ie, falling short of problem drinking criteria), drinkers were 0.480 (0.360 to 0.639) times as likely as abstainers, or those who consumed less, to be adherent (6 effect sizes, Q (5) = 2.180, Z = −5.021, P < 0.001). Among studies examining any or global alcohol use (eg, any use in the past year vs. none), the combined OR was 0.604 (0.531 to 0.687) (20 effect sizes, Q (19) = 17.312, Z = −7.704, P < 0.001). Overlap in the CIs for the effect sizes by level of drinking intensity indicates that, although the effect sizes are significantly different from zero, they are not significantly different from each other.

Moderator Analyses

Results of univariate moderator analyses are shown in Table 2. Findings indicate that the effect of alcohol on adherence was significantly moderated by a number of variables. Greater alcohol-related decrements in adherence were associated with higher proportions of men in study samples, lower proportions of participants reporting IDU, higher adherence criteria, nonuse of a self-report measure of adherence, nonuse of a dichotomous measure of adherence, assessing alcohol/substance use as a primary study aim, dichotomization of the alcohol variable, using an alcohol variable that took into account both quantity and frequency, use of a standardized alcohol measure, and use of the AUDIT in particular.

Table 2

Publication Bias

Inspection of a funnel plot revealed slight asymmetry, which is an indicator of publication bias. The Trim and Fill Method⁶⁵ indicated missing studies to the right of the mean. Eight studies were identified for trimming; the imputed point estimate was 0.638 (0.600 to 0.679). Because the imputed estimate is very close to observed estimate, 0.624 (0.585 to 0.664), and their CIs overlap considerably, publication bias seems to have been minimal.

DISCUSSION

This study provides the first meta-analytic evaluation of the association of alcohol use and antiretroviral adherence. Effect sizes for the combined studies suggested that those who used alcohol were 50%-60% as likely (OR = 0.548, 95% CI: 0.490 to 0.612) to be classified as adherent compared with those who abstained (or drank relatively less). Alcohol use that met or exceeded an objective threshold for problem drinking (defined as meeting NIAAA criteria for at-risk drinking or diagnostic criteria for an alcohol use disorder) was associated with the largest effect (OR = 0.474, 95% CI = 0.408 to 0.550), whereas the overall effect was smaller among studies examining any or global alcohol use (OR = 0.604, 95% CI = 0.531 to 0.687). Although these effect sizes were not significantly different from each other, they were significantly different from zero, and the point estimates can be viewed as broadly consistent with “dose-response” effects reported in previous studies.^9,14,25,27

Several variables moderated the alcohol-adherence association. This association was stronger in samples that included a higher proportion of men, a finding that is inconsistent with previous reports suggesting that alcohol's effects on adherence are more prominent among women.^28,29 The alcohol-adherence association was also stronger in samples with a lower reported prevalence of IDU. Given the established association of IDU with lower adherence, it is possible that any effects of alcohol on adherence are obscured in the context of IDU. The observation that effects were stronger in studies with larger samples presumably reflects greater statistical power. Aspects of alcohol use measurement also moderated the effects. Studies assessing both drinking quantity and frequency and those using the AUDIT (which assesses quantity and frequency) showed stronger effect sizes. A recent study found that when disaggregating the NIAAA at-risk drinking criteria into its 2 components (>4 drinks per day or >14 drinks per week), only the former predicted reduced adherence.²⁶ Taken together, the available evidence suggests that drinking quantity is a more robust and important predictor of adherence than drinking frequency, a finding that seems consistent with dose-related alcohol effects on adherence.^14,27 Dichotomous (compared with continuous) drinking outcomes were also associated with stronger effects, perhaps because studies using continuous measures tended to rely on global variables (eg, drinks per week) that did not index drinking quantity.

With respect to adherence assessment, moderator analyses indicated greater alcohol-related decrements in adherence in studies where adherence was defined using a higher criterion (eg, 100% vs. 90%). This result is consistent with event-level findings suggesting that alcohol's effects are more evident under more difficult adherence requirements⁶⁶ and suggests that alcohol use might be particularly detrimental to achieving perfect or near perfect adherence. The alcohol-adherence association was also stronger when using continuous adherence measures. Continuous measures presumably afford greater statistical power and have been shown to explain the most variance in viral load.⁶⁷ Incorporating continuous measures might allow more sensitive evaluation of alcohol-adherence associations in future studies. Finally, alcohol's effects on adherence were stronger when using assessment approaches other than self-report. Similarly, research on illicit drug use and adherence suggests that this association might be more reliable when using MEMS compared with self-report.⁶⁸ Use of MEMS specifically was not a significant moderator in this study, perhaps due to low power given that only 4 studies used MEMS. Objective measures might be more likely to detect significant associations due to fewer sources of measurement error, including social desirability influences.⁶⁷ Readers are referred elsewhere for comprehensive reviews of adherence assessment approaches.^67,69,70

In addition to establishing provisional effect-size estimates, this study offers a basis for discussing methodologic and conceptual issues in research on alcohol and HAART adherence. A primary concern is the substantial heterogeneity in the measurement and definition of alcohol use across studies, which makes it difficult to compare and aggregate findings. Researchers are encouraged to use standardized assessment approaches that include validated and multidimensional measures of alcohol use. The AUDIT⁶³ is a particularly useful measure given its brevity (10 items), established validity,⁷¹ and inclusion of items assessing drinking frequency, quantity, heavy episodes, and symptoms of alcohol dependence. Moreover, this measure is the recommended standard in primary care settings.⁶⁴ Timeline followback approaches, although relatively more time consuming, are extremely useful for providing nuanced assessments of the daily covariation among drinking and adherence.^25,26 Other event-level methods that permit fine-grained analyses⁷² warrant consideration in future studies. Relying solely on diagnostic criteria is probably less useful because traditional diagnostic schemes (and some brief screening methods) omit measures of drinking quantity, which is a significant limitation.⁷³ Consistent with this reasoning, a recent study showed that drinking quantity/frequency, but not alcohol-related problems, predicted reduced adherence.³⁵

Although the association of alcohol use and nonadherence is replicable and reliable, it remains difficult to speak to the causal nature of this association. The majority of studies included in this review were cross-sectional reports that evaluated global associations using retrospective measures of drinking and adherence. In a substantial proportion of studies, there was little or no overlap among the alcohol use and adherence assessment intervals. These limitations restrict the ability to infer causal effects and leave open the possibility that these associations could be attributable to other variables. If alcohol use is embedded in a broader context of problematic behaviors that also influence adherence, including IDU or other substance use, spurious associations could emerge (the association of tobacco use with nonadherence²⁴ likely reflects this phenomenon). The possibility that alcohol use is simply a marker for broader substance use involvement cannot be ruled out based on the current analyses; however, our finding that the alcohol-adherence association was significantly stronger in the context of lower IDU argues against this possibility and suggests a unique association of alcohol with adherence. Moreover, recent studies using sophisticated measurement approaches^25,26,66 provide compelling evidence that alcohol use is closely associated with decreased adherence. Continued use of these approaches would increase the ability to speak to causal associations. Researchers have also begun to examine specific intrapersonal and situational moderators of alcohol's effects on adherence.^26,66 We suggest that future research should continue to evaluate potential moderators to clarify the conditions under which alcohol use is likely to influence adherence. Because the association of alcohol and nonadherence seems significant and reliable across studies, further efforts to evaluate global associations may do little to extend knowledge in this area. That noted, there is a dearth of research on this issue in developing countries and establishing basic associations of alcohol and adherence in these settings would be useful.

A notable aspect of this literature is the omission of theoretical frameworks for understanding alcohol's association with adherence. Of the studies included in this review, the vast majority did not discuss possible mechanisms for these effects. One intuitive mechanism is cognitive impairment, such that acute intoxication might interfere with one's capacity to plan for or remember dosing requirements.²⁶ However, additional explanations are possible. Alcohol users might have decreased access to HAART⁷⁴ or may use alcohol to reduce or avoid HIV-related negative mood states,^75,76 a motive that could also lead one to neglect adherence requirements. It is also important to note that some patients intentionally skip medication doses when drinking due to misperceptions about possible toxic interactions.^15,18,77 These various explanations each have unique theoretical and clinical implications for research and intervention at the intersection of alcohol and adherence. An important direction for future research is to specify mechanisms that explain the link between drinking and nonadherence, which should aid in identifying intervention targets. Such mechanisms are likely to involve cognitive factors such as alcohol-related beliefs, expectancies, and motives, in addition to environmental and event-level factors.

The present study has several limitations. Given that the measurement and definitions of drinking and adherence varied considerably across studies, effect sizes should be considered provisional and interpreted as relative (rather than absolute) estimates of the likelihood of nonadherence in the context of alcohol use. Although we imposed a relatively objective measure of drinking intensity in the stratification analyses, there was still heterogeneity within categories due to measurement differences across studies, and these analyses relied on a modest number of effect sizes. Results concerning significant moderators should also be interpreted with caution. Another limitation is the omission of unpublished studies, although there was minimal evidence of publication bias.

The current findings support the need for interventions that address alcohol use in the context of HAART.^14,23 Given reported associations of alcohol use and immunologic function among those living with HIV/AIDS,^9,19-22 successful alcohol interventions could potentially show salutary effects on disease progression and, theoretically, life expectancy.⁷⁸ Few such interventions have been tested and more are needed.¹² In a recent study,²³ an alcohol/adherence intervention did not influence drinking but nonetheless led to improved adherence, decreased viral load, and increased CD4 cell counts, suggesting that adherence and biological outcomes can be improved even in the context of continued alcohol use. Similarly, meta-analytic research suggests that drug users often maintain adequate adherence, especially in the context of medical and psychosocial support.¹¹ Interventions might therefore aim not only to reduce alcohol use but also to promote strategies for maximizing adherence among those who are unlikely or unwilling to cease drinking. These efforts will benefit from an improved characterization of alcohol's relation to adherence and identification of factors that mediate or moderate this association.

ACKNOWLEDGMENT

The authors thank Jacqueline M. Otto for her assistance with literature reviews.

REFERENCES