The studies were selected from (a) seven electronic databases (PsycInfo, Medline, IME, EMI, Teseo, IBECS, ISOC and ISI Web of Knowledge) using a boolean search on the title of any type of publication: {[HAART OR highly active antiretroviral therapy] AND adheren* AND [HIV OR AIDS OR (human immu* virus) OR (acquired immu* syndrome)] AND [Spanish OR Spain]}; (b) Web browsers; (c) the summary of the latest conference proceedings; (d) funded projects on HIV / AIDS in Spain and (e) tracing of references cited in others studies. The studies had to be written either in English or in Spanish.

Studies were included in the review if they (a) aimed to evaluate adherence to HAART in a Spanish sample, (b) had a cross-sectional or cohort design, (c) evaluated an HIV-positive sample over 18 years old under treatment with HAART, (d) measured adherence at least once and using one strategy, (e) established an intake of >90% or 95% of the medication prescribed as the cut-off point for adherence, and (f) provided sufficient information to obtain the proportion of adherence to HAART.

The search ended on September 13, 2009 and studies from 1998 were included. Twenty-three independent studies that met the six selection criteria were included in this meta-analysis (table 1). These 23 primary studies19–41 provided 34 estimates of adherence to HAART. In three studies20,32,34 the total independence among their samples could not be checked.

The authors of 21 of the 23 primary studies included in the meta-analysis were emailed to request information on the databases they had used in their studies or at least a set of data that would allow two groups to be constructed: men and women. Nineteen authors responded to the message. The databases of eight studies and additional data from three studies were received.

Two independent trained raters coded each study following the coding manual (available from the main author on request). The final coding form registered 42 variables, which were grouped into three sets: “extrinsic”, of the “design” and the “sample”.

Although no scale to assess the methodological quality was used, a small group of variables related to the methodological design were coded to compare adherence according to the studies’ characteristics.

The proportion of adherence to HAART was estimated in each study as the effect size index. When, due to gender or location, more than one group could be drawn from a study, separate estimates of adherence were calculated for each group. If adherence was measured with more than one strategy, the average adherence was calculated. If the study evaluated adherence at different time points, the first assessment of adherence was chosen to avoid dependence, although a sensitivity analysis was performed for each set of measurements including all the possible comparisons. Because the length of follow-up varied widely across studies, we divided outcomes into four measurement intervals as a strategy to examine all study assessments: between 0-3 weeks (k=21), between 4-16 weeks (k=3), between 17-26 weeks (k=6), more than 27 weeks (k=4).

To ensure the normality of the effect size index, all the statistics were obtained using a logit transformation of the proportion of adherence (T=lnp1−p), where p was the proportion of adherence for each comparison. Then, the meta-analysis using a random effect model weighted by the inverse variance was performed. Finally, all the results were transformed back to a proportion for a more comprehensive interpretation of the data using the formula42p=eT1+eT. This outcome ranges from 0 to 1, where 1 indicates that all the patients have achieved high adherence (at least >90%), while 0 indicates that no patient has exceeded this cut-off point.

Homogeneity was evaluated using the Q test and the I2 index with its confidence interval43. The relationship among study dimensions and the proportion of adherence variability was examined by using modified least squares regression analyses with weights equivalent to the inverse of the variance for each effect size. When feasible factors related significantly to the proportion of adherence, they were entered into a series of models controlling for intercorrelations among the maintained study dimensions. These combined models allowed determination of the extent to which variation might be exclusively attributed to surviving study dimensions or not. The continuous variables that were significant in the univariate analyses were zero-centered to reduce multicollinearity44; if they were categorical, dummy variables were created44. Models with simultaneous independent variables were created if these factors were registered in more than five studies (k >5) and under mixed-effects assumptions, which are considered to have more conservative statistical power45.

The publication bias was analyzed through three different strategies: Trim and Fill46, Begg's strategy47, and the Egger test48.

The characteristics of the 23 primary studies were as follows: 21 (91.3%) studies were published and two (8.7%) were unpublished (doctoral theses), 10 (43.5%) were written in English and 13 (56.5%) in Spanish, 12 (52.2%) had a cross-sectional design and 11 (47.8%) a longitudinal design. All of the studies were performed between 1998 and 2008.

Patients with an intake of >90% of prescribed HAART were considered adherent. Adherence was assessed using just one strategy in 11 studies (47.8%), two strategies in eight (34.8%) and three strategies in four (17.4%). Self-report was the most frequently used strategy to quantify adherence in 19 studies (48.7%), followed by registration of dispensing medicines in 11 (28.2%), plasma drug concentration in four (10.3%), counting of surplus medication in two (5.1%), viral load in two (5.1%) and electronic devices in one (2.6%).

From the 23 studies included in table 1, a sample (table 2) composed of 9,331 HIV-positive patients older than 18 years under HAART was obtained, with 6,740 (72.2%) men and 2,591 (27.8%) women and a mean age of 37.9 years (SD: 2.83, range: 33-44).

The average adherence to HAART under the random effects model for the Spanish sample was 0.54 (95%CI: 0.49-0.59), showing wide heterogeneity (I2=91.20; 95%CI: 88.75-93.13) under the fixed-effects model. Therefore, only the results under the random-effects model and mixed-effects model to explain the heterogeneity not explained by the model (I2=51%; 95%CI: 11.45-72.88) are presented. Chart 1 shows the forest plot of the proportion of adherence of the 34 groups as well as the overall mean proportion of adherence (at the bottom of the chart). Chart 2 shows the proportion of adherence by sex.

Chart 1.

Forest plot.

(0.74MB).

Chart 2.

Forest plot.

(0.43MB).

A sensitivity analysis was performed to (i) test the influence of possible outliers, (ii) test the patterns of the set of studies divided by the week of measurement, and (iii) determine whether the results were affected by the exclusion of some studies in which the total independence among their samples could not be checked. Comparison of all the possible outliers revealed that none worked as a real outlier. The time interval showed the same patterns when the data were analyzed by interval as a separate meta-analysis rather than as a moderator without inclusion of more than one measurement per study. Finally, we decided to maintain all studies with a likely dependence because their exclusion did not affect the final results but would involve a substantial loss of information.

Intercoder reliability was higher than 0.90. Cohen's kappa was used for categorical variables (κ=0.98) and the Spearman-Brown correlation coefficient for continuous variables (r=0.955). Disagreements were solved through discussion.

The three tests used to assess possible publication bias aimed to show absence of bias. These tests were the Trim and Fill (the results indicate that there are no missing studies), Begg's test (z=–0.10, p=0.922) and Egger's test (bias=-1.58, t=–1.98, p=0.054).

A higher percentage of participants in stage A and with baseline viral load <200 copies/ml showed greater adherence. However, lower adherence was found when a higher percentage of patients were in stage B, with a baseline viral load >200 copies/ml, and with a high level of education, that is, education from secondary school onwards (table 3; this table depicts the variables registered in more than five groups, k >5).

Some of the significant moderators showed high collinearity (rs > 0.90, e.g., stage A and stage B, the two variables for baseline viral load) and therefore separate combined models were tested by different groups of moderators, avoiding dependence among them and excluding high educational level due to the small number of studies reporting this data (k=6). After all the possible combinations of stage and baseline viral load, the only variable that remained significant when baseline viral load (> 200 copies/ml because this value was measured in the largest number of studies) was controlled for was the percentage of patients in stage A. The model explaining 52% of the variance, more than expected by chance, was stage A (β=0.69, p=0.0005) controlling by baseline viral load > 200 copies /ml (β=–0.21, p=0.30), yielding the following regression equation (expressed in the unstandardized coefficients, the B values): –1.637+0.060 x (stage A) –0.0078 (baseline viral load > 200 copies/ml).