Interventions and survey interviews were conducted in funded sites between April 2004 and December 2006. Sites were selected based on a competitive application process and were located in Chapel Hill, NC; Boston, MA; Baltimore, MD; New York City, NY; Seattle, WA; Sacramento, CA; San Diego, CA; Birmingham, AL; Philadelphia, PA; Decatur, GA; Miami, FL; Chicago, IL; and Tucson, AZ. Some sites had more than one clinic, and all participating clinics received funding from the federal Ryan White Program with 300 patients or more. Each site required local human research subjects approval in addition to the approval obtained from the University of California, San Francisco, which served as the cross-site evaluation center.
Although 15 sites were funded through this initiative, the analysis presented here is restricted to data from 13 of the sites. The Office of AIDS Policy and Planning, Los Angeles site was omitted because study investigators used different interviewing procedures and the resulting data were systematically different from the other 14 sites. Data from the Whitman Walker Clinic site in Washington, D.C. was omitted because there was an interruption in the study due to fiscal hardship within the organization; this interruption resulted in information not being available on patient receipt of the intervention and assignment to the intervention or assessment-only conditions.
A total of 3,556 HIV-infected patients completed audio-computer assisted self interviews (ACASI) at baseline. These interviews included standardized cross-site questions and additional questions that a site’s study team added and used in local analysis. The ACASI survey was programmed using Questionnaire Development System version 2.0 by Nova Research Company (Bethesda, MD). Inclusion criteria were HIV-infected status, receipt of primary care at the clinic, age of 18 years or older, and ability to provide informed consent. For the cross-site evaluation, we translated the interview instrument into Spanish; however, this version was used at only one site. All other respondents were English-speaking. Sites had the option of adding additional inclusion requirements, which are summarized in Table 1. The cross-site evaluation survey took approximately 30 minutes to complete.
Recruitment and screening of potential respondents was undertaken exclusively in medical clinics serving HIV-infected clients. Sites used a variety of recruitment materials including brochures, posters, and project descriptions, as well as direct contact by study staff in clinics. Interested patients were briefly screened by project personnel to determine their self-reported HIV status as well as basic demographic and contact information. Then, eligible participants were scheduled for a baseline interview. Screening took place in a private setting, usually in a room or quiet place in the clinic. Most sites used incentives worth approximately $25 such as cash, a grocery voucher, or a gift certificate, to encourage participation in the evaluation portion of the project. Participants were not given incentives to attend intervention visits.
Patients were individually randomized to an intervention group or a comparison condition on a site-by-site basis. Patients in the intervention group were assigned to receive interventions from their medical care providers during routine HIV care visits alone or in combination with services delivered by health educators, case managers or social workers or by HIV-infected peers trained to conduct HIV prevention counseling. Providers delivered interventions as stand alone sessions or in conjunction with routine clinical visits. Social workers, health educators and HIV-infected peers hired by these projects served the same role. In sites that included either medical provider-delivered or prevention specialist-delivered interventions, patients in the comparison group received the standard of care, routine HIV care visits by their medical care provider. Some sites’ interventions used a combination of approaches so that patients received prevention counseling from both medical providers and prevention specialists. In these multi-provider sites, patients in the comparison group received medical provider-delivered prevention counseling only. For this analysis, data were combined from patients in sites where the comparison group received the standard of care with those in multi-provider sites where the intervention condition was medical provider prevention only. Intervention and comparison types are summarized in Table 1.
Demographic Characteristics and Health Status Indicators
Detailed background and demographic items included participant’s age, race/ethnicity, gender, self-identified sexual orientation, relationship status, educational level, employment status, and income. In addition, self-reported health status indicators were assessed including: most recent CD4 count, HIV viral load, and current use of antiretroviral medication.
Participants were asked to report sexual behavior over a 6 month recall period. Separate but equivalent versions of questions were developed for men and women, each with language tailored to be consistent with the participant’s gender and sexual orientation. Participants were asked to provide the number of times they had engaged in insertive or receptive vaginal and/or anal sex with HIV-infected partners, HIV-uninfected partners and partners of unknown HIV status. Participants were also asked about the number of times they had used condoms (male or female) from the beginning to the end of penetration and the number of times sex was unprotected. Unprotected sex was limited in the questioning to any act of insertive or receptive anal or vaginal intercourse in which a participant did not use a condom, a definition that excludes risk acts produced by accidental condom slippage or breakage.
Sexual Transmission Risk
Transmission risk acts were defined as reports of unprotected anal or vaginal intercourse without the use of a condom with any HIV-uninfected or unknown status partners. These acts were dichotomized into, “Yes” or “No.” Our outcome indicator does not include sex with HIV-infected partners because superinfection has not been observed in individuals such as those participating in this project, primarily characterized as having been infected for more than 3 years, engaged in care and on antiretroviral drugs.
Use of legal and illegal substances was assessed over a 3 month recall period. Items included alcohol, cocaine/crack, sedatives, tranquilizers, stimulants (such as crystal methamphetamine), analgesics, inhalants, marijuana, hallucinogens, and heroin. Use of injected drugs was assessed over the past 30 days. Items included frequency of injection and whether a participant had lent a needle to someone else after using it.
Patients repeated the assessments at 6-month intervals for 12 months after their initial baseline assessment and randomization.
We compare participant characteristics, sexual behavior and substance use at study entry by intervention type using chi-square tests of homogeneity. Patients assigned to peer or health educator/social worker-led interventions responded similarly to the interventions with respect to their sexual transmission risk behavior and were combined in the analysis into a category entitled “prevention specialists.” We present the prevalence of sexual transmission risk behavior among participants in each intervention type over the 12-month assessment period graphically (Fig. 1). We compare the odds of sexual transmission risk behavior among participants in each intervention type to the odds of sexual transmission risk behavior among participants in the standard of care group at each time point using generalized estimating equations (GEE). In this model, we estimate the odds of sexual transmission risk behavior among participants in each intervention type at study entry, 6 months and 12 months using indicator variables for each intervention type at each time point. In order to account for clustering of participant behavior by intervention site, the GEE models adjust for sexual transmission risk behavior at study entry, as well as the correlation among participants at each site and within intervention groups.
Because randomization occurred at the level of individual sites, rather than across the study as a whole, we employed inverse probability of treatment weighting (IPTW) to account for the effect of differences in the patient populations across intervention types on the observed effect of each intervention type [12, 13]. This method has the effect of creating samples in each intervention type that have been weighted to balance out the distribution of demographic and other patient characteristics associated with sexual transmission risk behavior. This creates a weighed population where characteristics of individuals are similar across all intervention types. Estimated parameters thus reflect the effect of each intervention type that would have been seen if it were possible to randomize the treatments across settings.
To estimate these weights, we used categorical logistic regression to estimate the probability of being assigned to each intervention type, given an individual’s demographic characteristics (age, gender, sexual orientation, race/ethnicity) and reported substance use (alcohol, speed and/or injecting drug use) at baseline. Factors in the model were chosen based on their association with the primary outcome at baseline, which are described in our paper on predictors of risk in this population . The inverses of these probabilities were then applied as weights to the GEE model described above. However, application of these weights impacted the observed results by less than 1%. Therefore, they were not retained in the final model.
Because loss to follow-up was different across study sites and treatment arms, we used additional weighting methods to assess the impact of loss to follow-up on observed study results. These methods have the effect of weighting the data at each follow-up to resemble the sample at baseline with respect to factors associated with sexual transmission risk. Estimated parameters thus reflect the expected effect of each intervention type had no loss-to-follow-up occurred. To estimate these weights, we used logistic regression to estimate the probability of remaining in the study at each time point, given an individual’s demographic and substance use at baseline as defined above. The inverse of these probabilities were then applied as inverse probabilities of censoring weights (IPCW) to the final GEE model described above. Application of these results did impact the observed results. Therefore these weights were retained in the final model. The combination of the GEE model and the weighting procedures allows us to generate estimates for the effect of treatment at a population level accounting for differences across sites and the correlation of individuals results within sites. All analyses were implemented in SAS (version 9.1, SAS Institute Inc, Cary, NC).