Introduction

In 2010 it was estimated that 33.3 million people worldwide were living with human immunodeficiency virus (HIV); 2.6 million of these individuals were newly infected, of whom over 95% were living in developing countries [1].

Highly active antiretroviral therapy (HAART) has improved the clinical situation and the prognosis of most patients infected with HIV by decreasing morbidity and mortality [2–4]. However, high levels of adherence are necessary to achieve viral suppression [5, 6], prevent the development of resistant strains [7–9], and reduce disease progression [10] and death [11, 12]. The minimum cut-off for sufficient HAART adherence in order to achieve the highest treatment efficiency is not clearly established [13, 14] but usually ranges between ≥90% and ≥95%. [13, 15–17].

Despite the need for rigorous assessment of HAART adherence, a “gold standard” for measurement has not yet been found [18, 19]. A variety of measurement strategies are available, each presenting their own strengths and weaknesses. Therefore, the use of more than a single strategy to measure adherence is recommended [20, 21]. Self-report is most commonly used [22] because it is inexpensive, feasible in a wide variety of settings, and is clinically applicable. Yet it tends to overestimate adherence [23, 24] because of recall bias or a desire to please the treatment provider and prevent criticism [25, 26]. Aside from these problems associated with measurement, it is also not clear what individual- and structural-level factors are associated with greater levels of adherence.

Although some systematic reviews on adherence to antiretroviral therapy attempting to elucidate these questions have been published, [27–29] their results have not been unanimous. For example, while Puigventos [27] identifies a wide number of sociodemographic, health and clinical factors related to adherence from thirty studies worldwide, Mills [28] shows that the region where the study was conducted, adherence thresholds, and study quality can explain the variability in levels of adherence. On the other hand, Malta [29] finds that active substance use, depression and low social support are associated with poor adherence.

Despite the fact that meta-analysis is traditionally used to combine results from experimental studies, such as rand-omized control trials, the use of meta-analysis to combine results from observational studies is gaining popularity because of the many issues in public health that cannot be studied without the use of observational designs [30]. The current meta-analysis rigorously integrates statistical research findings from individual observational studies on HAART adherence with two main objectives: (a) to estimate the average rates of reported adherence in the available literature of people in the world that maintain an intake of ≥90% of prescribed HAART at baseline measurement, and (b) to identify the factors associated with such adherence.

Methods

Studies were selected from eight electronic databases published in English and Spanish (PsycInfo, ERIC, Medline, IME, Teseo, IBECS, ISOC and ISI Web of Knowledge), using a Boolean search: {[HAART OR highly active antiretroviral therapy] AND [adheren* OR compliance] AND [HIV OR AIDS OR (human immu* virus) OR (acquired immu* syndrome)]; searches were also conducted with the Spanish equivalent terms. We also reviewed the reference lists of the articles included in this meta-analysis and other known research about HAART adherence for additional articles. There were no exclusions by publication type, and all studies were included via our selection process (Fig. 1).

Fig. 1
figure 1

Selection process for inclusion in the meta-analysis

Full size image

The studies were included in this review if they fulfilled the following criteria: (a) measured adherence to HAART at least once and used one or more method of measuring adherence, (b) had a cross-sectional or cohorts design, (c) evaluated a sample of patients ≥18 years old with HIV who were in treatment with HAART, and (d) provided enough information to allow estimation of the proportion of patients with ≥90% adherence to prescribed HAART.

Using a predefined protocol (available on request) two trained raters independently coded each study across forty-three dimensions. These dimensions spanned biological (e.g., baseline clinical infection stage according to the Centers for Disease Control [CDC] guidelines [31], CD4 count, viral load), demographic (e.g., age, ethnicity, self-identified as MSM, IDU or heterosexual), and social (e.g., marital status, educational attainment, employment status) characteristics of the studies’ samples.

The methodological quality of the studies was assessed according to the twenty-two items recommended by the STROBE Statement [32]. Although no value is assigned to each item in the STROBE list, the research team assigned a value to each item by consensus. A value of 0 was allotted if the item did not fulfil the STROBE Statement, 0.5 was allotted if the Statement was partially fulfilled, and 1 if the Statement was fully fulfilled.

Additionally, we sought to characterize the environments that the individual studies were situated in by coding for country-level development indicators. Specifically, we recorded the human development index (HDI) value [33], Gini index value [34], and the HIV prevalence [34] of the country the study was conducted in. The HDI is a composite index that measures a country’s average achievements in three basic aspects of human development: life expectancy, education and standard of living. The United Nations categorizes values between 0.80 and 1 as indicative of high human development, between 0.50 and 0.79 as medium development, and below 0.50 as low human development. The Gini index is the most commonly used country-level measure of inequality, with values that range from 0 to 100, where a value of 0 expresses total equality and a value of 100 expresses maximal inequality.

The values corresponding to the HDI, Gini index and HIV prevalence were selected not only on the basis of the country the study was conducted in, but also the study’s year of data collection. If the value corresponding to year of data collection was not available in one of the above indices, we used the closest available value within 2 years from data collection. If there was no information on the year of data collection, the value closest to the year of publication was used. According to these criteria, index and prevalence values from 2 years prior to data collection were used in sixteen studies, and from the closest year to publication in six studies.

The proportion of people who reported ≥90% adherence to HAART was estimated in each study as the effect size index (ES). If adherence was measured with more than one strategy (i.e., refill-based adherence, pill-counts, the use of electronic devices, and/or plasma drug concentration) and an average value for adherence was not provided, an average was calculated. If the study evaluated adherence at multiple time points, the value from the first measurement time (most often occurring at baseline) was chosen to avoid dependence. Further, when enough information was possible to stratify data for multiple groups in a study (e.g., gender, use of complementary and alternative medicine [CAM] vs. none), separate adherence estimates were calculated for each group. In this way, the eighty-four studies included in our meta-analysis provided one hundred and two groups (k = 102), each with a single independent estimate (ES) of proportion of sample who were ≥90% adherent to HAART (see Fig. 1).

To ensure the normality of the effect size (ES) metric all the statistics were obtained using a logit transformation on the proportion adherent \( \left( {T = \ln \left( {{\frac{p}{1 - p}}} \right)} \right) \), where P was the proportion of people who reported an intake of ≥90% HAART for each comparison. The weighted effect size was then obtained following fixed- and random-effects assumptions. Both fixed- and random-effects results indicate the same pattern, but we present here the latter as the findings were more robust [35]. Finally, all of the results were transformed back into a proportion for a more comprehensive interpretation of the data using the formula \( p = {\frac{{e^{T} }}{{1 + e^{T} }}} \) [36]. This final outcome (p) is a proportion that ranges from 0 to 1, where 1 indicates that all patients have reported at least ≥90% adherence, while 0 means that no patient has reached this level of adherence.

The asymmetries of the effect sizes distribution due to publication or other types of bias of the effect size distribution were analyzed through three different strategies: Trim and Fill [37], Begg’s strategy [38], and Egger’s test [39]. The homogeneity was evaluated using the Q test and the index I 2 with its confidence interval [40]. The relation between study dimensions and ES variability was examined using independent modified least squares regression analyses for each independent variable with weights equivalent to the inverse of the random-effects variance for each effect size. Thus the moderator analysis was conducted under mixed-effects assumptions—a more conservative approach than a fixed-effects model. This more conservative approach was chosen due to the large heterogeneity of the sample and to avoid spurious significant differences [35, 41]. When the variables were categorical, a beta value was derived from a multiple R of contrast of comparisons [42, 43], and Bonferroni correction was considered in order to avoid significance by chance on a multiple comparison test. There were an insufficient number of studies per moderator to permit simultaneous entry of all moderator dimensions that exhibited significance on a bivariate basis.

A sensitivity analysis was performed in order to: (a) test the influence of possible outliers, (b) visualize trends in the results, and (c) evaluate whether adherence proportion varied significantly when comparing the three designated levels of adherence (≥90%, ≥95%, and 100%), the four first measurement occasions (at baseline, 4–16 weeks, 17–26 weeks and ≥27 weeks), the number of methods of measuring adherence (one, two or three), and the methods of measuring adherence employed (self-report, refill-based adherence, pill-count, electronic devices and/or plasma drug concentration).

Results

The search for this study ended on 27 January 2010 and 84 observational studies providing 102 independent estimations to HAART adherence (Fig. 1). These studies were performed between 1999 and 2009 (Table 1) and no significant pattern emerged in the bivariate analysis due to the year of data collection or publication. A pooled sample of 33,199 patients older than 18 years taking prescribed HAART was obtained from the eighty-four included studies. For characteristics of the sample, see Table 2.

Table 1 Description of the studies included in the meta-analysis
Full size table
Table 2 Description of the sample
Full size table

Seventy five (89.3%) studies were published and 9 (10.7%) unpublished; 40 (50.6%) utilized a cross-sectional design and 39 (49.4%) used cohorts. Methodological quality ranged from 0 to 17. Adherence was characterized in 21 studies (25%) as 100% adherence to HAART, in 36 studies (42.9%) as ≥95%, and as ≥90% in 27 studies (32.1%). Adherence was assessed using one method only in 63 studies (75%), two methods in 20 studies (23.8%) and three methods in 1 study (1.2%). Self-report was the most frequently used adherence measurement method and was employed in 77 studies (72.6%). Refill-based adherence was the second-most popular measurement, used in 19 studies (17.9%), followed by pill-counts in 4 studies (3.8%), the use of electronic devices in 3 studies (2.8%), and plasma drug concentration in 3 studies (2.8%). (See Table 1)

Intercoder reliability was 0.80. Cohen’s Kappa was used for the categorical factors (κ = 0.71) and the Spearman-Brown correlation coefficient was used for the continuous factors (r = 0.89). Disagreements were resolved through discussion.

The weighted mean effect size of people who reported ≥90% adherence HAART under random-effects assumptions was 0.62 (95% CI 0.59–0.66; I 2 = 97.75%). Chart 2 shows the forest plot of the ES of the 102 groups ordered according to their ES, as well as the overall weighted mean ES (at the bottom of the chart). Chart 3 shows the ES by region.

Chart 2
figure 2

Forest plot. Proportion people reported to intake ≥90 of prescribed HAART

Full size image
Chart 3
figure 3

Forest plot. Proportion of people reported to intake ≥90 of prescribed HAART by region

Full size image

The sensitivity analysis showed that: (a) after comparing all the possible outliers, none had biased the distribution of the ES sampled significantly, (b) after ordering and grouping the ES, no significant patterns emerged, and (c) statistical significance was not found when comparing the three cut-offs for HAART adherence, the assessment period post-baseline (i.e. <4 weeks, 4–16 weeks, 17–27 weeks, and >27 weeks), the number of measurement methods used, or the type of measurement methods employed. Further, the three strategies used to assess possible asymmetry of the ES distribution showed absence of any bias (Trim and Fill results indicate that no study is missing; Begg’s test [z = −0.26, P = 0.73]; Egger’s test [bias = −0.84, t = −0.87, P = 0.39]).

As can be seen in Table 3, significant factors in the bivariate analyses were (a) the human development index (HDI), (b) region, (c) ethnicity, (d) self-identified as MSM and/or IDU), (e) participating in methadone maintenance, and (f) clinical infection stage defined by the CDC (A and C). Specifically, large rates of adherence are more related to studies that include a greater proportion of participants who are white or IDU, and include a lower proportion of participants who are MSM or participate in methadone maintenance. Large rates of adherence are also associated with studies conducted in countries with lower HDI, and with samples including more participants who are diagnosed with clinical infection stage A than stage C.

Table 3 Bivariate analyses. Mixed effects model
Full size table

We employed an exploratory analysis to examine whether these six moderators retain significance and follow similar patterns when controlling for the adherence measure used (see Table 3). All of them are still significant and in the same direction regardless of the number of strategies that were used to evaluate adherence or if self-report or dispenser registration measures were employed.

Discussion

The findings from this meta-analysis suggest that the mean proportion of people reporting an intake of ≥90% prescribed HAART is 62% worldwide. This proportion does not vary significantly when comparing the three cut-off levels for adherence (≥90%, ≥95% and 100%), the four first measurement occasions, the number of measurement methods used, or the type of methods of measuring adherence employed. With only 62% of the studied sample reporting ≥90% adherence to HAART, a question arises–what is happening with the 38% of the sample reporting poor adherence? Perhaps HAART is still efficacious when patients are less than 90% adherent. Kitahata et al. [44] used pharmacy refill data among 923 HIV-positive patients and showed that there was no difference in the risk of disease progression between those with moderate (70–90%) and high (≥90%) levels of adherence compared to those with low (<70%) adherence [44]. Moreover, Lima et al. [45] reached the conclusion that although perfect adherence remains an important goal of therapy to prevent disease progression, individuals with long-term viral suppression may be able to miss more doses without experiencing viral rebound. The answer to this question remains beyond the scope of our analyses, and future research should aim to investigate the relationship between adherence and treatment efficacy more thoroughly.

The proportion of people who reported ≥90% adherence to HAART varied depending on the region where the study was conducted. Further, by taking into account country HDI values, we found that in countries with low HDI the average proportion of adherence is higher than in countries with high HDI. This finding is consistent with past research, which has found that samples in developing countries are just as, if not more, adherent to prescribed HAART than samples in developed countries [46–48]. In the meta-analysis by Mills et al. [28], which observed facilitators and barriers to HAART adherence in developed and developing countries, samples in Sub-Saharan Africa were 77% adherent to HAART, as compared to 55% adherent in North America. Further research must be undertaken to explain the variability of this trend. Group identity also related to levels of adherence such that samples with higher proportions of MSM and lower proportion of IDU were associated with higher proportions of participants reporting ≥90% adherence.

With the current meta-analysis we have not been able to firmly conclude what factors are correlated with the proportion of people who are ≥90% adherent to HAART. However, we have found that in addition to HDI and group identity, race, clinical infection stage, participation in methadone maintenance, and anxiety levels are significantly associated with an intake ≥90% of prescribed HAART. Still, with the exception of HDI, these factors registered in fewer than half of the studies. Therefore, although they were significant in the bivariate analyses, we were not able to conduct multivariate analysis with these factors because less than a third reported data across the factors, thus rendering it impractical to evaluate in a complex model.

This meta-analysis was limited by the high heterogeneity found in the proportion of people reporting ≥90% adherence, as well as the scarcity of available factors to explain the heterogeneity. Additionally, the ability of this meta-analysis to obtain more precise estimations of adherence levels was limited by the heterogeneity in measurement type across reports, which necessitated the combination of different types of outcome data. Another limitation is that all of the studies included in this meta-analysis utilized a cross-sectional or cohorts design, making it difficult to determine causal relationships between level of adherence and other factors.

HAART is an expensive, but necessary resource that is difficult to access in many poor countries or communities [46]. In order to make the most of the scarce resources available, it is essential that further observational studies continue researching the factors associated with HAART adherence. However, it is necessary that these observational studies collect and report data on a much wider range of biological and psychosocial factors to enable more comprehensive meta-analyses in the future.