Background. We examined adherence to highly active antiretroviral therapy (HAART) in the homeless population, a population thought to be at high risk for poor adherence to therapy and for development of drug-resistant strains of human immunodeficiency virus (HIV).
Methods. We performed a 12-month prospective study of 148 persons receiving HAART who were identified in a stratified screening of the homeless and marginally housed. We sampled in lunch lines, shelters, and hotels in 3 neighborhoods of San Francisco, California. We used pill counts at unannounced home visits as the primary measure of adherence.
Results. Of 148 individuals sampled, 46 (31%) discontinued HAART during the study. Average adherence in the group of those who discontinued HAART was 51%, and 9% of these subjects had undetectable virus loads (i.e., <400 copies/mL) at the last follow-up visit. Predictors of discontinuation of therapy were depressive symptoms, injection drug use, African American ethnicity, and early poor adherence. Of 148 subjects, 102 (69%) continued to receive HAART throughout the study period. Average adherence in the group of those who continued to receive HAART was 74%, and 55% of these subjects had undetectable virus loads at the last follow-up visit. Predictors of lower average adherence in this group were African American ethnicity and use of crack cocaine; men who had sex with men had higher adherence.
Conclusions. One-third of homeless and marginally housed persons receiving HAART discontinued therapy during the follow-up period and would benefit from adherence interventions directed at sustaining therapy; two-thirds continued to receive therapy at adherence levels comparable to those found with other clinical populations.
Antiretroviral therapy is less likely to be prescribed for individuals who are indigent [4, 5]. Transmission of multiple drug—resistant HIV has been demonstrated in the homeless population [6, 7], and multiple drug—resistant tuberculosis has been shown to develop in the homeless population as a result, in part, of nonadherence to therapy . Thus, health care providers have sometimes been reluctant to prescribe HAART in this population [9, 10]. However, a steadily increasing proportion of HIV cases in the United States are among poor individuals .
We set out to examine adherence to HAART in the San Francisco, California, homeless population, including both the “true homeless” population (i.e., the street- and shelter-dwelling population ) and the “marginally housed” population (i.e., hotel-dwelling). We chose to represent the population by residential strata because a replicable sample can be drawn by this method [13, 14]. An estimated 9% of homeless persons in San Francisco are HIV seropositive , one-fourth have been hospitalized for mental illness, and one-fourth have been in prison; in addition, rates of drug and alcohol use are extremely high [15, 16].
We report what is, to the best of our knowledge, the first community-based study of adherence to HIV therapy in the United States. We report, first, use of HAART in the cohort during a 6-year period and, second, adherence to therapy among those receiving HAART. We used pill count as the main adherence measure because electronic measures could not be obtained from subjects who used medication organizers. We have previously published cross-sectional data on some subjects [17–19]. Here, we present the first longitudinal study of adherence to HAART in the homeless and marginally housed, including predictors of both adherence level during therapy and discontinuation of therapy.
The Research in Access to Care in the Homeless (REACH) cohort. We screened homeless and marginally housed persons from April 1996 through December 1997 and from April 1999 through April 2000 in 3 San Francisco neighborhoods. We set out to represent the homeless population by service-use strata, following Burnham and Kogel . We constructed a multistage cluster sample that was stratified into shelters, free meal programs, and single-room occupancy hotels. The shelter sample was drawn from all overnight shelters housing ⩾50 people. The meal program sample included participants in 5 of 6 midday free-meal programs that served ⩾100 adults at least 3 days per week. Hotels were sampled with probability proportional to size from 292 low-cost residential hotels; hotels were eligible if they had ⩾20 “usually occupied” residential (nontourist) rooms licensed to rent for $400 (1996–1997) or $500 (1999–2000) per month or less. We observed all consenting individuals who were HIV seropositive [15, 16]. An estimated 8000 hotel rooms charged <$600 a month in the study area, with an 82% occupation rate and with multiple tenants in 37% of occupied rooms, for an estimated hotel-dwelling population of 9600. Recent estimates of the homeless population are in the range of 8600–9400 individuals . The study was approved by the Committee on Human Research of the University of California, San Francisco, and written consent was obtained.
Adherence measurement study. During the period March 1998 through April 2001, cohort members reporting receipt of HAART were recruited into a 12-month adherence study. Subjects received unannounced visits at their place of residence approximately monthly; blood samples were collected for virus load measurement, and adherence was measured using 3 methods [17, 20]. Briefly, in the first method, subjects reported all doses missed during the previous 3 days; doses received were reported as a proportion of doses prescribed ("self-report adherence"). In the second method, subjects received home visits at which pill counts were conducted; pills were counted, and “pill-count adherence” was defined as the difference between counts at successive visits as a proportion of pills prescribed. Pill count visits were completely unannounced (i.e., they began with an unscheduled knock on the subject's door, except in the small number of cases where access was controlled or subjects lived outside). The subjects' daily habits were tracked to identify appropriate times of day for making such a visit. When repeated visits were necessary, no messages were left. In the third method, subjects received electronic pill bottle caps that recorded bottle openings (Medication Event Measuring system [MEMS]; Aardex); cap information was downloaded at monthly visits, and the 3-day record of cap openings was adjusted with interview data regarding out-of-pocket doses collected at the visit ("adjusted MEMS adherence") . We report adherence to the principal active drug, usually a protease inhibitor (PI) or a nonnucleoside reverse-transcriptase inhibitor (NNRTI). MEMS adherence was not available for 63 (43%) of the 148 subjects who used medication organizers.
We defined early adherence as average adherence in the first 3 months of the follow-up period or average adherence until discontinuation of therapy. Subjects were defined as having discontinued therapy after 1 month in which no pills were taken.
Period of observation. Unannounced visits continued for 12 months or until discontinuation of therapy. Some subjects who began therapy in the second or third month of the follow-up period and who were censored after 12 months of follow-up had <12 months of pill counts. Subjects who were institutionalized or who moved away were censored. The median duration of follow-up for 102 subjects who did not discontinue therapy was 10.5 months. Virus load and CD4+ cell count measurements continued for 12 months for all subjects followed-up in the REACH study. Of 102 subjects who did not discontinue therapy, 16 dropped out of the study (4 died, 2 were institutionalized, 5 moved out of the San Francisco Bay area, and 1 was lost to follow-up).
Measurement. Physical health was measured by the SF-36 role-physical scale; general mental health was measured by the SF-36 mental health scale . Depressive symptomatology was measured by a score of ⩾15 on the Beck depression index .
Statistical methods. We compared subjects of the adherence study with those who were not eligible for the study and compared those with electronic pill count measurements to those without, in both cases with use of forward stepwise multivariate logistic regression (with P value for entry equal to .10).
We compared those who discontinued taking medication with those who did not with use of forward stepwise proportional hazards regression, with P value for entry equal to .10, modeling time to discontinuation. Model 1 excluded early adherence and virus load (as a continuous variable and as the dichotomous variable <400 copies/mL vs. ⩾400 copies/mL). Model 2 included all variables. All variables significant (P < .01) in either model (except the exclusions) were included in the final models.
We used repeated measures analysis of months receiving therapy to estimate the difference in pill count between those who discontinued therapy and those who did not, to estimate the trend in pill counts over time, and to determine whether the trend differed according to whether therapy was discontinued. Fixed effects were study month (range, 0–12 months), group (discontinued therapy vs. did not discontinue therapy), and an interaction term (study month by group). Within-subject variability was modeled using autoregressive covariance, and between-subject covariance was unstructured.
Because the monthly pill counts did not change significantly over time among those who did not discontinue therapy (neither when unadjusted or when adjusted for calendar time and regimen change), we examined predictors of average adherence in this group. Forward stepwise linear regression (with P value for entry equal to .10) of average adherence produced the same predictors as trichotomous logistic regression. Linear regression results are presented for ease of interpretation.
Use of HAART in the REACH cohort 1996–2002. Of 4055 persons screened, 48% were street- or shelter-dwellers, and 52% were hotel-dwellers. Of 411 HIV-infected persons identified, 330 (80%) consented to follow-up in the REACH cohort.
The proportion of these individuals receiving any antiretroviral therapy rose to ∼50% by late 1997 and has remained approximately constant (figure 1), as has the proportion of those with a CD4+ cell count <350 × 106 cells/L who are receiving any antiretroviral therapy (56% in mid-2002). Among all cohort members, the proportion with suppressed virus (i.e., HIV RNA load <400 copies/mL) increased to 30% (figure 1). Of those receiving HAART, 52% had suppressed virus in mid-2002.
Recruitment to adherence study. Of 330 persons in the REACH cohort, 183 (56%) received HAART at some point during the adherence study recruitment period. Of these 183 subjects, 148 (84%) consented to and completed ⩾1 monthly visit. Of these, 85% were male, 38% were African American, and 43% were white; most of these subjects were resident in hotels, and the levels of drug and alcohol use were extremely high (table 1). Median CD4+ cell count at baseline was 289 × 106 cells/L, median log virus load (log10 HIV RNA copies/mL) was 2.7, and 49% of subjects had virus loads <400 copies/mL. Of these subjects, 62% were receiving a PI regimen at baseline, 34% were receiving an NNRTI regimen, and 4% were receiving other regimens.
When we compared adherence study subjects with all other subjects in the REACH cohort (i.e, with those who had not received antiretroviral therapy during the recruitment period), they were more likely to have a CD4+ cell count <200 × 106 cells/L (OR, 1.9; P = .02) and to have a regular health care provider (OR, 3.0; P < .01). They were less likely to use crack cocaine (OR, .5;P = .03) and to have a Beck depression index score of >15 (OR, .5; P < .01).
Discontinuation of therapy. Forty-six (31%) of 148 subjects discontinued therapy during the follow-up period after a median of 7 months. Those who discontinued therapy had a median of 4.5 months of prior HAART at baseline, compared with 10 months for those who continued therapy (P < .01, by Wilcoxon test).
Early adherence was a strong univariate predictor of discontinuation; those with early adherence ⩽50% had a hazard ratio (HR) for discontinuation of 6.2 (95% CI, 2.8–14.1) when compared with those with 91%–100% adherence (figure 2). Independent predictors of discontinuation in a model that did not include early adherence and virus load at baseline were depressive symptoms, injection drug use, and African American ethnicity; age was marginally significant (table 2). When we included detectable virus load at baseline and early adherence in the model, poor early adherence (⩽50%) remained significant (adjusted HR, 6.1; 95% CI, 2.4–15.6) (table 2), as did depression and injection drug use.
Adherence while receiving therapy. Mean and median rates of pill count adherence were 67% and 71%, respectively, across all visits. Mean and median rates of self-reported adherence were 81% and 89%, respectively. Mean and median adjusted rates of MEMS adherence among those not using medication organizers were 55% and 57%, respectively; mean and median pill count adherence rates in this group were 65% and 69%, respectively. Overall crude correlations were 0.69 between pill count adherence and adjusted MEMS adherence (n = 85 persons), 0.65 between self-reported adherence and pill count adherence (n = 148), and 0.61 between self-reported adherence and adjusted MEMS adherence (n = 85).
The mean pill count adherence rate during months in which therapy was received was 51% among those who discontinued therapy and 74% among those who did not. Among the 46 subjects who discontinued therapy, monthly adherence decreased by an estimated 3.6% per month in a mixed model; among the 102 subjects who did not discontinue therapy, adherence decreased 0.3% per month (figure 3).
Predictors of adherence level. For the 102 persons who continued to receive therapy, independent multivariate predictors of lower adherence were African American ethnicity and use of crack cocaine; men who had sex with men had a higher adherence rate than did other risk groups (tables 3 and 4).
Adherence and virus load. The mean HIV RNA level at baseline was 2.7 log10 copies/mL in all subjects, 3.4 log10 copies/mL in those who discontinued therapy, and 2.6 log10 copies/mL in those who did not discontinue therapy. Among those who discontinued therapy, the mean log virus load increased by .88 log10 copies/mL over 12 months, and the percentage of subjects with viral suppression decreased from 33% at baseline to 9% (figure 4) (virus loads were obtained for subjects who discontinued HAART but continued in the parent REACH study). Among those who did not discontinue therapy, the mean log virus load remained at ∼2.6 log10 copies/mL, and the percentage of subjects with viral suppression decreased from 57% at baseline to 54% after 12 months. Mean CD4+ cell count increase, from baseline to last follow-up visit, was 14 × 106 cells/L; it was –57 × 106 cells/L in those who discontinued therapy and 47 × 106 cells/L in those who did not.
Among those who did not discontinue therapy, we examined viral suppression at 12 months by average adherence level in 6 categories: adherence rates of ⩽50%, 51%–80%, 81%–85%, 86%–90%, 91%–95%, and 96%–100%. The percentages of subjects with suppressed virus were 19%, 33%, 43%, 64%, 100%, and 70%, respectively (P < .01, by Cochran-Armitage test for trend). We made the same comparison in the 59 persons with measured MEMS adherence who did not discontinue therapy. The percentages with viral suppression were 27%, 57%, 57%, 86%, 100%, and 100%, respectively (P < .01, by Cochran-Armitage test for trend). Comparable results were obtained with use of pill counts.
Pill count adherence compared with adjusted MEMS adherence. Compared with 63 persons without MEMS data, those with MEMS data were more likely to have a CD4+ cell count of 200–500 × 106 cells/L (OR, 3.7; P < .001) or >500 × 106 cells/L (OR 3.4, P = .02), compared with <200 × 106 cells/L, and they were more likely to be true homeless persons at baseline (OR, 2.4; P = 0.1). They were less likely to have viral suppression at baseline (OR, 0.5; P = .05), to be injection drug users (OR, 0.4;P = .03), or to drink heavily (OR, 0.3; P = .03).
We compared adjusted MEMS adherence and pill count adherence as predictors of virus load at 12 months among 59 persons with both measures who did not discontinue therapy. Pill count adherence was a slightly better predictor (χ2 = 15.9 vs. 10.7).
The rate of antiretroviral use in our homeless cohort has remained stable since the late 1990s, and in 2002, it was at 56% for those individuals with a CD4+ cell count <350 × 106 cells/L (almost all of whom were receiving HAART). This can be compared with San Francisco surveillance data, which shows that, of those with a diagnosis of AIDS, 72% were receiving HAART in 2001 and a further 16% were receiving other antiretroviral regimens (W. McFarland [San Francisco Department of Public Health], personal communication).
Subjects receiving HAART did not adhere to therapy in 2 ways. Approximately one-third of those receiving HAART discontinued therapy (resulting in the low overall treatment rate), and other subjects had low adherence but managed to continue receiving therapy. Different variables predicted these 2 phenomena.
In our study, 31% of subjects discontinued therapy, at a median of 7 months from study baseline. This group had average adherence of 51%, decreasing rapidly during follow-up. Discontinuation of therapy was predicted by depressive symptoms, injection drug use, and African American ethnicity. Early poor adherence was also a strong predictor of discontinuation of therapy.
Although emphasis in the literature has generally been placed on maintaining the level of adherence during receipt of therapy, even perfect adherence will not delay progression to AIDS and death unless treatment is sustained [23, 24]. Our data suggest that physicians treating indigent persons may need to attend to depression or drug-use issues before or during HIV treatment . Depression has been repeatedly identified as a predictor of nonadherence in studies of AIDS  and other diseases , and it is a predictor of mortality in patients with AIDS . Our data suggest that the effect on mortality is explained in part by failure to persist in receiving treatment.
We found that injection drug use was associated with discontinuation of therapy. Chen et al.  also found that injection drug users were at high risk for discontinuation, and injection drug use has been identified as a risk factor for poor adherence [29–32]. This suggests an approach combining drug treatment and HAART. HAART has been applied with success in methadone and buprenorphine maintenance programs [33, 34]; however, the effect may not persist after maintenance therapy is discontinued .
Two-thirds of our subjects continued to receive therapy, with adherence rates averaging 74% and not decreasing significantly over time. More than one-half of subjects had undetectable virus loads at the end of follow-up, and, as in other studies [17, 23, 36, 37], viral suppression was predicted by better adherence. Predictors of lower adherence were African American ethnicity and use of crack cocaine (but not injection drug use); men who had sex with men had higher adherence rates than other risk groups.
African American ethnicity predicted both discontinuation of therapy and low adherence in those who continued to receive therapy. In the continuing group, adherence in African American subjects was 60%, compared with 81% in all other subjects. Lower adherence in African Americans has been reported in several studies [30, 31, 38] and is a major issue in improving adherence in the homeless population.
How does adherence in the homeless population compare with that in other populations? In most people, near-perfect adherence is required for sustained viral suppression. However, adherence is often lower in practice . Adherence levels of 53%–83% have been reported in 7 North American studies using objective adherence measures [30, 31, 36, 40–43]; higher rates have been reported using self-reported adherence measures [29, 38, 44]. In our study, average adherence rates in those who continued to receive therapy were not strikingly low.
In cross-sectional data, the proportion of those receiving therapy in the REACH cohort with viral suppression was 52% in 2002. In San Francisco surveillance data, 61% of patients receiving therapy in mid-2002 had virus loads <400 copies/mL (W. McFarland [San Francisco Department of Public Health], personal communication). In a population-based cross sectional study in England, 53% of patients had complete viral suppression (i.e., <50 copies/mL) in 2002 . Although the frequency of discontinuation of therapy in the homeless population means that effective treatment levels are low, the level of viral suppression in those who continue to receive therapy is not dissimilar to that in other populations.
Our data suggest that interventions should be directed at sustaining treatment, as well as at improving adherence to treatment. Adherence interventions have traditionally focused on the latter, either by adherence case management [38, 47] or directly observed therapy [46, 47].
One limitation of the current study is the use of pill-count adherence measures even though electronic measures (such as MEMS adherence measures) have been thought to be more accurate in comparative studies [42, 43, 23]. However, these studies compare MEMS adherence to clinic-based pill counts. Our pill counts were obtained, unscheduled, at the subject's place of residence. In our study, among those with both measures, average adjusted MEMS adherence was 55%, compared with 65% by the pill count measure—which is much closer than in studies using clinic pill counts [30, 42, 43]. Pill count adherence was a slightly better predictor of viral suppression at 12 months than was adjusted MEMS adherence.
A second limitation of the current study is selection bias. The sampling method used underrepresents both the least functional subjects (such as those with severe psychosis) and the most functional subjects (such as those who have jobs). It may overestimate or underestimate the proportion of subjects receiving HAART and may exclude both high-adherence and low-adherence groups. However, early validation studies suggest that a high proportion of street-identified homeless persons are accessible by this method [13, 14].
AIDS is increasingly common in economically and socially marginal individuals. Given the improvements in treatment since 1996, failure to provide therapy amounts to withdrawal of life. As HIV infection increasingly becomes a disease associated with poverty—in the United States and in the world as a whole—development of functional methods of delivering HAART to disadvantaged populations will be an increasingly important part of the combined treatment and prevention approach to HIV infection.
We would like to acknowledge the work of the REACH study field staff.
Financial support. National Institute of Mental Health (grant MH 54907) and the Universitywide AIDS Research Program of the State of California. D.R.B. received additional funding from The Doris Duke Charitable Foundation.
Conflict of interest. All authors: No conflict.