The Cross-Border Project was the first HIV prevention project for IDUs in which the same interventions were implemented on both sides of an international border, in this case the border between northern Vietnam and southern China. Beginning in the early to middle 1990s, this area experienced epidemics of heroin injection and subsequently HIV infection among IDUs. As shown by Beyrer et al. [20], an important heroin trans-shipment route ran from the Golden Triangle of Laos, Thailand, and Burma into Northern Vietnam and thence into Southern China and to Hong Kong and the larger world. As heroin was shipped through the border region, drug users shifted from traditional opium smoking to heroin smoking and then quite quickly, for reasons of cost-effectiveness, to heroin injection. A map of the Cross-Border Project sites is presented in Figure 1.

As we have shown in an earlier paper, there was a gradient of HIV prevalence in the Cross-Border sites from the highest in the localities in Vietnam farthest from the China border to the lowest in the Chinese sites farthest from the Vietnam border [21]. While other patterns of HIV transmission may also have occurred, this prevalence gradient suggests that the dual epidemics of heroin injection and HIV infection moved from south to north through the border area. We also showed that border crossing behavior was associated with being HIV-infected among IDUS from both Chinese and Vietnamese sites [22].

The Cross-Border interventions are based on a “peer outreach” model developed in the U.S. in the late 1980s [23]–[26]. This model was implemented from the beginning of the project in 2002 in Lang Son Province, Vietnam and Ning Ming County, China. In Ha Giang Town between 2003 and 2005, a “peer-driven intervention” (PDI) was implemented. This model, which essentially applies Respondent-Driven Sampling to delivery of HIV prevention messages and commodities, involves large numbers of IDUs as recruiters and deliverers of harm reduction services in return for small rewards [27], as opposed to the peer outreach model, which employs regular peer educators who receive monthly stipends for their work. In 2005, the PDI in Ha Giang Town was converted to a peer outreach model because we concluded that it had already reached all of the IDUs – indeed, the PDI reached 900 IDUs, far more than the 700 estimated by the health department to live in Ha Giang.

In the peer outreach model, initial groups of peer educators (PEs) were recruited by local health departments based on performance during initial training, recommendations from IDU “gatekeepers”, and expressed interest in joining the project. The original intention was that the peer educators be former IDUs. However, it quickly became apparent that very few IDUs in these sites had been able to stop injecting heroin. Therefore, it was necessary to recruit primarily current IDUs. Most of the PEs are young men but women PEs have been recruited to reach women IDUs as well as sexual partners of IDUs and sex workers in companion interventions for women at risk.

The initial groups of PEs received two weeks of training and the local health departments have provided periodic refresher training. In addition, the peer educators’ weekly meetings include training on special topics. The peer educators are supervised locally by health department staff. Because there were different funding sources and local cost norms, PEs’ stipends differed across sites: from $30 per month in Lang Son to $46 in Ha Giang to $117 in Ning Ming.

For a variety of reasons, mostly associated with the peer educators’ ongoing drug use, commitment to drug detention centers, arrest for petty crimes, or moving away, there has been substantial turnover among them. Turnover rates varied by year and by site but sometimes reached 50% per year. Local health department staff recruit and train replacements as needed.

The peer educators regularly contact other IDUs in the community and provide them with information on reducing drug use- and sex-related HIV risks, verbally and through distribution of brochures. They distribute new needles/syringes, ampoules of sterile water for injection solution, and condoms and vouchers redeemable for these items at participating pharmacies. The peer educators also perform a valuable public health service by collecting and disposing of used needles/syringes that might otherwise put community residents at risk.

The average annual cost of these interventions was $60,000–$75,000 per province (that is, all sites in the province), including the cross-sectional surveys. (More detailed information on the interventions may be found in a replication manual available from the corresponding author [28]).

The interventions aim to reduce HIV risk behaviors among IDUs, thereby stabilizing HIV prevalence and reducing HIV incidence on both sides of the border and preventing cross-border HIV transmission. We consider this a structural intervention because it seeks not only to influence the behavior of individual IDUs (e.g., reduction or elimination of the sharing of injection equipment) but also to diffuse HIV risk reduction messages and sterile injection equipment more widely through the IDU community [29]–[30].

The Cross-Border interventions have reached substantial proportions of the IDUs in the project sites through direct contact with PEs and “secondary exchange” in which IDUs receiving needles/syringes directly from the project pass on some of these to other IDUs. Through this and other means, the HIV risk reduction information and safer injection materials disseminated by the project are also reaching IDUs who do not personally interact with the project.

The cross-border project also seeks to diffuse knowledge of HIV and build support for the interventions in the general community. Project staff have regular meetings with police, local government bodies (e.g., People’s Committees), Communist Party officials, and mass organizations (e.g., Women’s Union, Youth Union). Staff and peer educators also regularly attend community meetings. These interactions build and maintain support for the interventions. Government leaders in Ning Ming, Lang Son, and Ha Giang gave initial approval to the interventions and directed all departments, including the police, as well as the general community, to give their support and cooperation. This support has been maintained but it requires constant nurturing through regular meetings and other interactions. Peer educators and project staff must repeatedly explain that the purpose of the project is to prevent disease and not to encourage drug use or give special treatment to drug users. Misinformation and misunderstanding are seemingly endemic, and the peer educators and project staff must constantly work to prevent misinformation and misunderstanding from undermining the interventions [28].

The evaluation of structural interventions poses complex methodological challenges. We believe that our interventions are appropriately evaluated at the community level rather than the individual level. Randomized controlled trials of these interventions pose serious ethical and cost issues.

Our design relies primarily on serial cross-sectional surveys of IDUs (including behavioral interviews and HIV testing) just before the start of the interventions and at follow up intervals thereafter. All survey waves reported on in this paper were carried out in the same sites in Ning Ming County, China (Ning Ming City, Aidian, Shilang, and Tongmian) and Lang Son (Lang Son City, Cao Loc, Loc Binh, Hop Thanh, Dong Dang, and Tan Thanh) and Ha Giang Provinces (Ha Giang Town), Vietnam. Data collection methods were largely parallel in Ning Ming County, Lang Son and Ha Giang, with some variation in the strategies used to recruit IDUs for the surveys. Routine monitoring data were employed to gauge levels of activities and commodities provided.

To be eligible, survey participants had to be at least 18 years of age and to have injected heroin in the past 6 months. In all sites, modified “snowball” techniques were used. These involved identification of initial respondents through project PEs, lists of registered IDUs, and/or IDUs present at mapped “shooting” places or other gathering places, with the remainder of sample quotas from each site recruited through snowball or chain referral methods. Chinese participants received 20 yuan (approximately $2.50) for the interview and blood sample, 5 yuan for each additional male respondent recruited, and 10 yuan for each additional woman respondent recruited. Vietnamese participants received 30,000 dong (about US$1.50) for the interview and blood sample.

An interviewer explained the survey procedures to each prospective participant. In Vietnam, a verbal informed consent was obtained from participants, with the interviewer certifying that consent by signing the form. This procedure was recommended by the Institutional Review Board of the National AIDS Standing Bureau in order to provide more assurance of confidentiality to prospective participants. In China, a standard signed informed consent was obtained from each participant.

Unique ID codes were constructed for each IDU participant composed of several letters and numbers representing, for example, the first letter of the mother’s family name and the subject’s date of birth. The objective was to have a unique identifier composed of items that participants could readily reconstruct in order to obtain their HIV test results. Test results were coded only by these identifiers.

A structured instrument was used for the interviews, based on version 2b of the questionnaire employed in the World Health Organization’s Drug Injection Study, Phase II [31]. The questionnaire includes a series of demographic and self-reported behavioral items. Most of the behavioral questions pertain to specific drug-related and sexual activities and measures of participation in the intervention during the 6-month period prior to the interview. (The full questionnaire is available upon request from the corresponding author.).

The interviews were conducted by trained staff of the local health departments. The baseline IDU survey in Lang Son Province, Vietnam was conducted in January, 2002 with 7 subsequent waves between March, 2003 and October, 2009. The 36-month and 60-month surveys had to be skipped because of temporary gaps in funding. In Ha Giang Town, Vietnam the baseline survey took place in January, 2003 with 9 additional waves between August, 2003 and February, 2011. The baseline survey in Ning Ming County, China occurred in July-August, 2002 with 7 subsequent waves between April, 2003 and June, 2008. Thus, in total we have data from 26 cross-sectional survey waves among IDUs across all sites.

The surveys included HIV antibody testing. Participants were given pre- and post-test counseling at local health centers. Blood was drawn by trained phlebotomists at the time of the interviews. Participants were told to return to the local health center on a certain date to receive their test results using their ID numbers.

In China, testing was by double ELISA (Vironostika HIV-uni-form, Organon [Holland]) with confirmation of initial HIV-positive results by Western blot methodology (Genelabs Diagnostics). All testing was conducted at the laboratory of the Guangxi Center for HIV/AIDS Prevention and Control in Nanning. In Vietnam, testing was performed at the laboratory of the Lang Son and Ha Giang Provincial Health Services using the Serodia SFD screening test (Biorad [France]) and double ELISA (Genescreen, Biorad [France]; Vironostika, Organon [Holland]). This is the official protocol of the Vietnam Ministry of Health.

To identify recent infections and inform HIV incidence estimates, we conducted BED capture-enzyme immunoassay testing [32]–[34] on HIV-positive serum/plasma samples using the Calypte HIV-1 BED EIA (Calypte Biomedical Corporation, Rockville, MD) for all Ning Ming samples, Lang Son 48-, 72-, and 84-month surveys, and Ha Giang 60-, 72- and 96-month surveys. The BED assay measures the proportion of IgG that is anti-HIV antibody and uses a multi-subtype synthetic peptide antigen for detection of HIV-1 subtypes B, E, and D. For Ning Ming, frozen HIV-positive serum/plasma samples were available from all survey waves and BED testing was carried out by the laboratory of the Guangxi CDC in Nanning, China. For the Vietnamese sites, we had frozen HIV-positive serum/plasma samples only from three survey waves (48, 72, and 84 months) in Lang Son and from three survey waves (60, 72, and 96 months in Ha Giang Town). This was because samples from earlier waves were not frozen. BED testing of Vietnamese samples, except for those from the Ha Giang 96-month survey, was done at the Thai Ministry of Public Health-U.S. Centers for Disease Control and Prevention HIV/STI Laboratory in Nonthaburi, Thailand. BED testing for Ha Giang’s 96-month survey was done at the National Institute of Hygiene and Epidemiology in Hanoi.

The study was reviewed and approved by the institutional review boards (IRBs) of the following institutions: Guangxi Center for HIV/AIDS Prevention and Control, the National AIDS Standing Bureau of Vietnam, Lang Son and Ha Giang Provincial Health Services, Abt Associates Inc., and Beth Israel Medical Center, New York, NY. The BED testing of Vietnamese samples in Thailand was done under a non-research determination from CDC.

Data sets were prepared in EpiInfo, version 6.04 by staff of the Guangxi Center for HIV/AIDS Prevention and Control (for Ning Ming) and, over time in Vietnam by staff of the National AIDS Standing Bureau and Abt Associates. The data were analyzed at Abt Associates Inc. using SAS, Version 9.22 (SAS, Inc., Cary, NC). For all analyses, we considered a measure to be statistically significant if it had an estimated p-value of less than or equal to 0.05.

To analyze trends in risk behaviors and HIV prevalence, we used regression analysis to estimate a time trend for each behavior and HIV prevalence. Separate regressions were run for each project site. Tests of significance for trends used a linear time trend, adding covariates for specific geographic subunits. We also took into account the potential correlation generated by multiple responses from the same respondents over time, using a robust covariance matrix with clustering on the survey respondent. The robust covariance matrix will not affect the estimated parameters from a model assuming independent and identically distributed errors, but will change the estimated parameter covariance matrix [35]. To estimate the model for these outcomes, we used the SAS SURVEYLOGISTIC procedure and the CLUSTER command to cluster the error matrix on survey respondents [36]. The model assumes clustering among responses from the same participant but independence across survey participants.

To explore further the possible associations between the interventions and trends in HIV prevalence, we compared the trends in our Vietnam project sites to those from sentinel surveillance in other northern Vietnam provinces without similar interventions (Dien Bien, Lao Cai, Bac Giang, Cao Bang, and Phu Tho) and compared the trends in Ning Ming to those in Yunnan Province, China, which also lacked similar interventions during the period under study. For these comparisons, we used a difference-in-difference estimator [37] to calculate the trend in HIV prevalence within project sites and “difference out” the underlying trend in HIV prevalence occurring in comparison provinces.

Ha Giang began its intervention in 2003, while Lang Son began in 2002. In Dien Bien, sentinel surveillance data were available for 2004 forward. In Lao Cai, sentinel surveillance reported HIV prevalence at 3 to 6 percent in 2000 through 2003, but not less than 14 percent in subsequent years (18 to 22 percent in 2004 through 2006). The rise in HIV prevalence seems suspicious between 2003 and 2004, so we elected to use data from 2004 to 2009 for the Vietnam comparisons. Changing the time period would not substantially change the results. To determine the trend in the Vietnamese comparison sites, we used a weighted least-squares regression, with the weights equaling 1 divided by the square root of the number of tests in each site for each year (for details, see Model Specifications S1). The dependent variable was the proportion testing positive for that site in that year. Independent variables included dummy variables for each site, and a time trend for each site (equaling 0 in 2004 and 1 in 2009). To compute the change in prevalence across the five provinces, we used a simple average of the percent decline in prevalence over the time period 2004 to 2009.

We used similar estimation techniques for China to compare HIV prevalence trends among IDUs in Ning Ming with those in Yunnan Province for the period 2002–2007, based on data in a paper by Jia [38]. We estimated a weighted least-squares regression, with the weights equaling 1 divided by the square root of the number of tests for each year. The dependent variable was the proportion testing positive in that year. The independent variable was a time trend (equaling 0 in 2002 and 1 in 2007). In the comparison with Yunnan Province, we also added a time-squared trend to the model because it greatly improved model fit (for details, see Model Specifications S1).

In both countries, we ran similar regressions for our project sites. For each project site, we ran separate regressions that included dummy variables for the subsite (in Lang Son and Ning Ming), demographic variables that included marital status, ethnic minority status, age (dummy variable for 21–30), indicator for male, and a time trend. For each site, we ran a linear probability model. In China, we used a generalized estimating equation with repeated measures for each person. For Vietnam, we used an OLS regression.

We employed two methods to estimate trends in HIV incidence: analysis of new injectors and BED testing. As described elsewhere [16], we estimated HIV incidence among new injectors, defined as individuals reporting injection histories of three years or less. To test the significance of a trend in estimated HIV incidence, we used a Generalized Estimating Equation [39] with a normal distribution and an identity link [35]. We estimated the GEE using the GENMOD procedure in SAS. In the GEE, the incidence was the dependent variable, the project site and survey wave were independent variables, and we allowed clustering when an individual appeared in more than one survey. The test of significance was based upon the z-statistic reported for the survey wave.

Our second method was based on BED testing, which can identify recent infections (approximately within the past six months) from samples that are HIV-positive on antibody tests [32]–[34]. BED results can be used to estimate HIV incidence from cross-sectional data [40].

Annual HIV incidence (I) was estimated based on the BED test results using the formula in the Calypte kit insert (I = ([365/w] N_inc)/(N_neg + [365/w] N_inc/2) where w, is the window or recency period, N_inc is the number of recent infections and N_neg is the number of HIV negatives in a survey wave. Using BED data from 190 seroconversion panels from Thailand [41], we determined the recency period (the mean period from initial seroconversion to an ODn of 0.8) as 152 days.

When specimens from all HIV-positives were not available for BED testing, the BED estimate was extrapolated by the factor: (# BED-recent/# of HIV-positive specimens available for testing). This assumes that the proportion of specimens that test positive for recent infection is the same as in those with available and missing specimens. Many of the calculated incidence rates were very close to or below zero. As a result, we computed confidence intervals based upon the Wilson score interval [42].

Adjustments to BED-based incidence estimates are advised because of the possibility of the test producing false-recent results. We selected the adjustment proposed by Welte [43]. This method adjusts well for variance in recency periods, among other factors proposed by McDougal [44] and Hargrove [45]. Such adjustments to very low incidence rates may generate negative values.

In addition to applying this adjustment, we examined survey responses for all participants with BED-recent results (linked through study ID numbers) on whether and when these individuals had been told that they had HIV infection and duration of any ARV treatment. Where linkage between names and study numbers was still available, we were able to check putative “recent” infections against registries of outpatient clinics providing ARV treatment. Any cases of long-term infection or ARV treatment initiated more than 6 months before the survey were excluded as recent infections for HIV incidence estimation.