How does hard-to-reach status affect antiretroviral therapy adherence in the HIV-infected population? Results from a meta-analysis of observational studies

We searched for studies containing ART adherence outcomes in HTR populations (drug users, homeless individuals, and sex workers). In accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) [28] guidelines, we developed a systematic literature search in electronic databases (PubMed, Cochrane Library and EMBASE) with no starting time limits, extending up to December 31, 2018, to identify studies that reported quantitative results regarding ART adherence by HIV-positive HTR individuals. The studies in the populations of interest had outcomes of effect sizes expressed as odds ratios (ORs). The search strategy was initially developed for the PubMed database and then adapted to the others. The following themes and keywords were then used to search for articles: (1) HIV infection: human immunodeficiency viruses OR HIV OR AIDS, (2) Antiretroviral therapy: ART OR HAART, (3) Adherence: adherence OR nonadherence OR suboptimal OR compliance OR noncompliance, and (4) HTR population: drug users OR sex workers OR homelessness. All four themes were then combined using the Boolean operator AND. References of eligible studies were hand-searched to identify additional relevant papers.

Articles that met the following criteria were included in this meta-analysis: (1) the study quantitatively examined the effect size of the association between HTR status and ART adherence; (2) the study reported the adherence level (e.g., ≥ 95, 100%) or provided information that could be used to estimate the adherence; (3) the study applied a clear definition of the measurements and timeframe (e.g., self-reported, prescription refill, 1 week, or 1 month); and (4) the study was a peer-reviewed full-text articles published in English. Papers were not excluded on the basis of study design, sample size, exposure or outcome measure method, or geographical region [29], but they were excluded if (1) they were literature reviews, meta-analyses, laboratory studies, descriptive studies, or case reports, (2) the outcome presented was not ART adherence, (3) they were general-population-based studies, or (4) there was no effect size reported. If data were duplicated, we included the study with the longest observation period. The study selection process is detailed in Fig. 1.

We developed a standardized data collection form in accordance with the search criteria, and two reviewers (DL and CZ) independently performed the data extraction and comparison. All data were checked by a third reviewer (ZH) to systematically establish a dataset. Any discrepancies in the process were resolved by discussion with another investigator (XZ) or through reference to the original articles. The extracted items included the title, first author, publication year, geographic region, study design, number of participants, cut-offs values for adherence, adherence measures, survey period, measures of correlation (i.e., adjusted ORs), and summary descriptions of the study population. Duplicate data in different publications were excluded, and the study with the most informative and complete data was selected.

The outcomes were expressed as ORs in the eligible studies. The ORs could be directly used if the original study presented the association between any HTR population (homeless individuals, sex workers or drug users) and optimal ART adherence (optimal adherence to ART is often defined by a certain threshold, e.g., ≥ 95%, and this threshold differs across studies). To ensure maximum comparability of the outcomes across studies, we inverted the reported ORs if studies presented the association between suboptimal/nonadherence of ART and any characteristics of the socially excluded populations.

Two investigators (DL and CZ) independently assessed the quality of each study using the 9-star Newcastle-Ottawa Scale (NOS) [30] for cohort studies. The overall score ranges between 0 and 9 stars, and we considered a study awarded 7 or more stars to be a high-quality study. The methodological quality of the cross-sectional studies included was assessed using the 11-item checklist recommended by the Agency for Healthcare Research and Quality (AHRQ) [31] from the National Institutes of Health, and we considered a study awarded 8 or more stars to be a high-quality study [32]. Any disagreements were resolved through consensus.

For the quantitative findings analysed in this study, outcomes expressed as ORs and their 95% confidential intervals (CIs) were extracted from each study. Crude (unadjusted) ORs or adjusted ORs were used directly in the pooled meta-analysis calculations. When ORs were generated from both univariable and multivariable models, the effect size from the multivariable model was used. All results were summarized in forest plots that showed the individual OR estimates.

Heterogeneity was tested using both the Q test and the I² statistic. Cochran’s Q test was applied to qualitatively assess the heterogeneity across studies [33]. The Q test assesses whether differences in study estimates are due to chance alone (typically a P value< 0.1 or < 0.05 indicates heterogeneity among study estimates). The I² statistic was used to quantify the extent of the heterogeneity and determine the proportion of the total variation in study estimates accounted for by it [34]. The results ranged from 0 to 100%, and a larger I² indicates that the total variation between studies is due to true heterogeneity rather than sampling error (chance). The interpretation of the amount of heterogeneity is as follows [35]: I² from 0 to 40%: might not be important; I² from 30 to 60%: may represent moderate heterogeneity; I² from 50 to 90%: may represent substantial heterogeneity; I² from 75 to 100%: considerable heterogeneity. A Q-statistic value of P < 0.05 also suggests the presence of heterogeneity [9]. As we anticipated high levels of heterogeneity, a DerSimonian & Laird random effects model [36] was employed to pool the ORs across studies regardless of the significance of the between-study heterogeneity. Additionally, potential sources of heterogeneity were explored by stratifying the analyses by subgroups. We examined the robustness of the pooled effect estimates in defined subgroups (according to study design, study quality, adherence measures, etc.). Similarly, within-study heterogeneity was explored using random effects meta-regression analyses. Factors associated with effect sizes (pooled ORs) were investigated, and the results are reported as ORs with 95% CIs.

We examined the possibility of publication bias by generating funnel plots and performing meta-bias analyses (Begg’s test and Egger’s test). The publication bias for each outcome was accessed by evaluating asymmetry in the funnel plot, and ln(OR) was plotted against its standard error. Due to the subjective nature of graphical evaluation, Begg’s rank correlation test was also used to examine the asymmetry of the funnel plot [37]. Egger’s regression asymmetry test [38] was used to examine the association between the effect estimate and its variance. If an asymmetric funnel plot was identified, a contour-enhanced funnel plot was then used to further explore the source of bias [39]. All analyses were 2 tailed, with a P value < 0.05 indicating statistical significance. To evaluate the stability of the conclusions and the influence of individual studies, a one-study removed approach [40] was applied by omitting one study at a time to explore whether the pooled estimates were strongly influenced by any single study. All analyses were conducted with Stata version 15 for Windows (Stata Corporation, College Station, TX, USA).

The electronic database search and manual search returned 2088 potentially relevant articles, and 1495 were excluded after title/abstract screening, resulting in 593 records for full text screening. Finally, 29 [41‐69] studies that fulfilled the study entry criteria were included in our meta-analysis. The overall search flow is presented in Fig. 1.

Overall, 16 cohort studies and 13 cross-sectional studies were included in this meta-analysis. These studies were performed between 1993 and 2017 and reported between 1999 and 2018; the studies were conducted in several countries or regions. Most of the studies were from the United States (n = 15, 52%) and Canada (n = 7, 24%), and most of them investigated the association between drug use and ART adherence (n = 26, 90%). The measurement period for ART adherence ranged from 2 days to 1 year, and most used 100% (n = 13, 45%) or ≥ 95% (n = 10, 34%) as the thresholds for optimal adherence. Most studies measured adherence using self-reported questionnaires (n = 21, 72%), 7 studies used pharmacy refills, and only one study used the self-reporting plus pill count method. The characteristics of the eligible studies are summarized in Table 1.

Of the 29 studies on ART adherence, 24 studies only investigated the association between drug use and ART adherence, 2 studies only studied the association between homelessness and ART adherence, and 1 study only reported the association between transaction sex and ART adherence. One study presented the relationships between ART adherence and all three HTR population characteristics, and one study presented the relationship between ART adherence and two HTR population characteristics (drug use and homelessness).

The quality assessments of the studies included in the analysis are listed in Additional file 1: Tables S1 and Table S2. The quality of the cohort studies ranged from 5 to 9 stars and was, on average, high, with a mean of 7 stars according to the NOS, while the quality of the cross-sectional studies ranged from 5 to 9 points and was, on average, moderate, with a mean AHRQ score of 7. The 16 cohort studies all described the selection of the non-exposed group, demonstrated that the outcome of interest was not present at the start of study, and followed the subjects long enough for the outcomes to occur. As the adherence measures for many cohorts were subjective (self-reported questionnaires), they had inaccurate outcome assessment and inadequate follow-up, leading to high levels of performance bias. However, performance bias due to confounding was low because of the reported adjustments. All 13 cross-sectional studies defined the source of information and indicated if the evaluators of the subjective components of study were blinded to other aspects of the status of the participants. Most of them, moreover, listed inclusion and exclusion criteria for exposed and unexposed subjects or referred to criteria used in previous publications, indicated whether the subjects were consecutive, and described the assessment and/or control of confounding variables. Similar to the cohort studies, many of the cross-sectionals did not describe undertaking any assessments for quality assurance purposes but clarified the expected follow-up and the percentage of patients for whom incomplete data or follow-up was obtained [30].

Twenty-nine studies examined the impacts of HTR status on optimal ART adherence. The pooled association measures (ORs) and 95% CIs and those from individual studies are presented in Fig. 2. Despite individual study results varying widely from OR = 0.17 (95% CI 0.04–0.57) to OR = 1.33 (95% CI 0.52–3.42), the overall pooled estimate indicates that compared with the general population, HTR populations had 45% lower odds of achieving optimal ART adherence (OR = 0.55, 95% CI 0.49–0.62, P = 0.000). The I² value was 49.3% (Q = 61.11, P = 0.001), indicating moderate statistical heterogeneity among the studies.

We examined publication bias by plotting the log-transformed association measures (ORs) against their standard errors. Asymmetry was observed according to the visual inspection of the funnel plot, indicating the presence of publication bias (Additional file 2: Figure S1). A contour-enhanced funnel plot was used to aid in interpretation (Additional file 2: Figure S2) and demonstrated that the majority of studies had very low statistical significance. Hence, publication bias was a more likely cause of the funnel plot asymmetry. Clear publication bias was also detected out by Egger’s linear regression test (P = 0.000) (Additional file 2: Figure S3) and Begg’s test (P = 0.009). The trim-and-fill method [70], used to correct for publication bias, demonstrated that no studies needed to be filled.

Leave-one study-out sensitivity analyses were conducted. The pooled estimate for optimal ART adherence ranged from OR = 0.57 (95% CI 0.53–0.62; when the study by Joseph [56] was excluded) to OR = 0.61 (95% CI 0.56–0.66; when the study by Sharpe [64] was excluded), suggesting that no study had undue influence on the pooled adherence estimate (Additional file 2: Figure S4).

We explored differences among studies and generated adjusted estimates by meta-regression and subgroup analysis as appropriate. The meta-regression covariates considered included study design, adherence threshold/cut-off point, adherence measure, region/country, observational period, and quality assessment results. Restricted maximum likelihood (REML) was applied to establish the regression model of OR-covariate. A P value of 0.03 was found for observational period, implying that the observational period was a source of between-study variance. When observational period was introduced into the regression model, the tau² reduced from 0.05 to 0.04 and could explain 22.09% of the between-study variance. The proportion of within-study variance explained by the observational period was 24.14%. From the value of I-squared_res in the output, 42.47% of the residual variation was due to heterogeneity, with the other 57.53% attributable to within-study sampling variability.

We conducted subgroup analyses to recalculate the pooled ORs according to study design, adherence threshold/cut-off point, adherence measure, region/country, observational period, and quality assessment results (Table 2). Significant inverse associations between HTR status and ART adherence were observed in all subgroups, and all I² test results indicated the presence of moderate heterogeneity within each subgroup [all P values (Q statistic) < 0.05]. In the interaction tests, the pooled ORs of the relationship between HTR and ART adherence in the subgroup with an observational period ≥6 months was 28% lower than that in the subgroup with an observational period < 6 months (OR value and its 95%CI for interaction: 0.72, 0.55–0.94, P value for interaction = 0.02), which meant that the HTR population is more likely to have suboptimal adherence during a longer observational period. However, there was no significant heterogeneity found in other subgroups.

There was extremely suboptimal adherence to ART in the HTR populations. We found a ratio of the pooled odds of optimal ART adherence in the HTR group to that in the general population of 0.55 (95% CI 0.49–0.62, P = 0.000), which means that compared with the general population, HTR populations have 45% lower odds of achieving optimal ART adherence. The relationships were markedly consistent across all subgroups (Table 2) and did not change in the sensitivity analyses (Additional file 2: Figure S4), adding further support for this conclusion.

We found a negative impact of social exclusion on adherence to HIV treatment, linking socioeconomic status to ART adherence. To the best of our knowledge, this is the first meta-analysis to comprehensively summarize the dire situation regarding suboptimal adherence to ART in socially excluded HTR populations. This meta-analysis reveals the gaps in ART adherence between HIV-infected people in the general population and those in HTR populations. In the context of the current focus on how combinations of social characteristics affect health [71, 72], we offer evidence of an association between health inequities and health outcomes. Health inequities are invisible social barriers created by ignorance and prejudice, and this discrimination hampers the equity of health services with regard to HIV treatment and transmission control. Therefore, policy makers and healthcare providers are the targets audience who need to be informed of the pooled data extracted from the existing studies involving health issues in socially excluded groups. These extreme inequities demand an intensive cross-sectoral policy and service response to prevent exclusion and improve health outcomes [73].

A meta-analysis of adherence to HAART by HIV+ patients in China revealed that the adherence over 1 month and ≥ 3 months was 80.9 and 68.3%, respectively, indicating an inverse association between adherence and treatment duration [74]. Likewise, we obtained similar results in the subgroup analysis, which indicated that the pooled effect size in the subgroup with an observational period ≥6 months was significantly lower than that in the subgroup with an observational period < 6 months. The gap (28%) between the ORs confirmed the suboptimal adherence to HIV treatment. However, data on ART adherence from HTR individuals with an observational period ≥6 months is scarce (9 studies with an observational period ≥6 months VS 20 studies with an observational period < than 6 months), although those studies reported the lowest pooled OR (OR = 0.44, 95%CI: 0.35–0.56, P = 0.000) across all subgroups in this meta-analysis. Due to the relative lack of data, the adherence to ART by HTR individuals for longer than 6 months should be addressed as a matter of priority in future research.

The reasons for lower adherence to ART in the HTR population are complicated. For drug users, it is unlikely that a single mechanism explains the adverse impact of active substance use on adherence. The results from a cohort study implied that the use of stimulants (i.e., cocaine or methamphetamine) proved to be particularly disruptive to adherence to therapy by HIV-infected adults, and adherence was most dramatically affected during periods of active stimulant use. There are also several potential mechanisms by which substance use may impact adherence behaviour, including neurocognitive deficits, psychosocial impairment, and exacerbation of psychiatric dysfunction [75]. Nevertheless, others have suggested that poor adherence among substance users in general may be due to their inconsistent, unpredictable, and chaotic lifestyles [76]. In addition, as demonstrated in other studies, in the settings with free ART access for sex workers, an array of individual, social and structural barriers to poor adherence were identified, such as stigma and discrimination in healthcare settings, a lack of support, the criminalisation of sex work for HIV-infected individuals, and geographic mobility [77, 78]. On-going drug use and poverty may also indirectly prevent sex workers from achieving optimal ART adherence. Furthermore, a study conducted by Royal SW et al. investigated adherence to HAART by homeless persons living with HIV/AIDS [19]. The results showed that coexisting problems of limited access to healthcare, an elevated risk of mental health problems, and worse attitudes toward treatment are associated with an increased likelihood of worse adherence. Others argued that barriers to care, including the lack of financial resources, lack of transportation, side effects, restrictions on when and how the medications should be taken and stored, and insufficient health insurance coverage, may be especially problematic for homeless individuals [79].

The interpretation of the findings should be considered in the context of the limitations of this study. We grouped different types of HTR populations (drug users, sex workers and homeless individuals) and believe that there is commonality with regard to their HIV treatment adherence owing to overlap in their experience of marked social exclusion. But we were unable to identify a sufficient number of studies performed with homeless and sex worker populations (only 2 studies included for sex workers and 4 included homeless individuals) and believe that further work is needed to specifically describe their HIV treatment experiences. And the representativeness of our analysis is limited for three reasons. First, publication bias was generated by our collection of papers published in English. As with most meta-analyses, the information reported in this source is restricted. Second, there was uniformity of the geographical distribution of the included studies. Fifteen studies were from the USA, and 7 studies were from Canada, resulting in an unequal weighting of the pooled estimates and making it impossible to generalize our findings to HTR populations in specific contexts or in a generally resource-poor settings. For example, in Eastern Europe, HIV infection is mainly driven by drug use, and access to HAART is uneven [80]. Last but not least, although we confirmed that the subjects investigated in Canada were from different research programs, most of the studies were based in the same city (Vancouver), which suggests the possibility of duplicated data and overly narrow CIs. Another limitation is the reporting bias that can be introduced by the methods used for measuring adherence. Twenty-one out of the 29 studies included in our analysis measured ART adherence by self-report questionnaires. Although self-reporting is the standard method of collecting behavioural information, it is not as accurate as objective measurements, such as electronic devices and concentrations of the medication in the blood [74], leading to overestimation of ART adherence. Furthermore, a considerable proportion of the observed heterogeneity may be explained by differences [81] in adherence thresholds, observational periods and study designs. The cut-off value for ART adherence ranged from 80 to 100%, and the recall time frame ranged from 2 days to 1 year. The absence of universal definitions of cut-off points and recall time frames for the measurement of ART adherence is likely to explain some of the variation. To use the most informative and complete data in cases of data duplication or overlap, we selected the study with the longest observation time. However, unsurprisingly, and consistent with findings from longitudinal studies of adherence [82], the adherence rates clearly decreased over time, and the participants dropped out, which introduced selection bias into our study. Meanwhile, all studies included in our meta-analyses were observational in design; randomized controlled trial data were not available. In addition, a random effects model was chosen based on the moderate level of heterogeneity [47] in the overall analysis, but substantial heterogeneity remained in the subgroup analysis, as is often the case in the meta-analysis of epidemiological studies [74]. Finally, although the included studies controlled for many important confounders, such as age, gender, race, education and employment status, it is still possible that there was residual confounding due to the presence of unknown confounders and/or imprecise adjustment strategies [29].