Studies of injection drug users (IDUs) in the mid-20th century relied on institutional settings such as hospitals, jails, and drug treatment centers as sampling frames. These sampling frames had inherent selection biases because of differences in prevalence of disease or risk behaviors among IDUs in or entering institutional settings compared to IDUs in community-based settings.^7,8 As the HIV/AIDS epidemic emerged in the United States in the mid-1980s, IDUs were identified as a high-risk group that needed to be studied outside of institutional settings. Two sampling methods were developed and utilized: targeted sampling (TS) and snowball sampling, which later was modified to develop respondent-driven sampling (RDS).

In 1986, Watters and Biernacki⁹ developed TS methods in San Francisco to recruit IDUs directly from communities. This method involves using secondary analysis of existing data (from drug treatment programs, hospitals, jails, etc.) and primary collection of ethnographic data to first establish characteristics of the target population. Then targeted enrollment plans (quotas) for each geographic area and demographic characteristic are established, and recruitment is conducted through the use of community health outreach workers.⁹ TS is an iterative process designed to assess the characteristics of the sample at several points so that sampling can be adjusted in service of obtaining a final sample similar to that of the hypothesized target population. It quickly became the most common recruitment method to study IDUs in the United States, including its use in the 23-city, National Institute on Drug Abuse (NIDA)-funded Cooperative Agreement study in the 1990s.^10,11 While TS has been a successful sampling method of IDUs, there is no formal way to assess whether the samples are representative. As such, it is not feasible to calculate valid prevalence estimates. In more recent years, researchers have enhanced TS by estimating the density of drug users in the target areas and using proportional sampling quotas.^12,13 These enhancements, however, have been shown to have significant selection biases.^4,8

Snowball sampling (also called chain referral) was developed in the 1960s by Goodman¹⁴ and is a logistically simple way of recruiting a convenience sample when the target population is hidden, and there is reason to believe that there are large social networks within the target population who would be willing to recruit members into a study. It involves starting with a few members of the target population and then giving them an incentive to bring in their acquaintances who qualify for the study. Those new recruits are also given incentives to bring in their acquaintances, and so on. This method of sampling was used to recruit IDUs and heroin users in the 1980s,¹⁵ including in the NIDA-funded 29-city National AIDS Demonstration Research project from 1987 to 1992.^16,17 An important tenet of snowball sampling is that initial recruits (also called “seeds”) be selected at random from the target population.¹⁴ However, because selecting the initial recruits at random is not possible with IDUs, the studies using snowball sampling to recruit IDUs in the 1980s and early 1990s were essentially considered convenience samples. However, in 1997, Heckathorn¹⁸ used mathematical modeling and data simulation to demonstrate that it is possible to generate population-based estimates by tracking the social networks of participants who refer each other to the study, given certain assumptions. He called this refined snowball sampling method RDS. By keeping track of who is referring whom, it is possible to identify a sequence of recruiter–recruit chains called “recruitment waves.” If these chains are sufficiently large, the composition of the sample (i.e., with respect to its demographic and behavioral characteristics) gradually stabilizes. At this point, the sample is said to have reached “equilibrium,” meaning that subsequent recruitment does not greatly alter the makeup of the sample and that sufficient data are collected to estimate and adjust for recruitment biases. Throughout the survey, the recruiter–recruit chains need to be carefully monitored and the size of the social network of each participant needs to be recorded. Using these network-size data and the characteristics of recruits and recruiters, statistical adjustments are made to the survey results to account for different probabilities of inclusion and the cluster effects introduced by the networked nature of the sample.¹⁹ In the 2000s, RDS become a very common method of recruiting IDUs in the United States, including in the Centers for Disease Control and Prevention (CDC)-funded 25-city National HIV Behavioral Surveillance (NHBS) study, and internationally.^20–28

We compared demographic characteristics, risk factors, and utilization of treatment and prevention programs in the TS and RDS studies (Table 1). In both samples, the majority were over 45 years old and male. The samples were similar demographically, with the exception of race/ethnicity. TS estimated a higher proportion of African Americans than RDS (49.3% vs. 35.7%), and the 95% CIs did not overlap. More TS participants than RDS participants reported having tested positive for hepatitis C virus (HCV) (79.7% vs. 55.9%). RDS study participants were less likely to report using methadone maintenance and methadone detoxification services (6% and 5%, respectively), as compared with participants in the TS study (19.2% and 14.1%, respectively). Finally, less RDS participants than TS participants reported obtaining syringes from SEPs (81% vs. 95%), while a larger proportion of RDS participants reported purchasing syringes at pharmacies (17.0% vs. 9.2%) in the past 12 months.

We found that the TS and RDS studies reached similar samples of IDU in terms of demographic characteristics, with the exception of African Americans. African Americans represent a small minority of the population in San Francisco (6.9% in 2007 per the US Census Bureau³⁹) and are largely concentrated in one neighborhood in the southeastern part of the city (Bayview/Hunter's Point, zip code 94124), which in 2005 was isolated from the rest of the city geographically by highway systems and poor access via public transportation (two bus lines, no streetcar, no subway). Through the secondary analysis and ethnography components of TS methodology, this neighborhood was identified as having a large population of IDUs, and a field site was placed in its midst. While the RDS study attempted to include IDUs from this neighborhood (three seeds), we hypothesize that the long travel time between the neighborhood and the RDS data collection site may have limited study participation. This finding suggests that when utilizing RDS, it may be wise to implement some of the steps of TS during the planning stages, which can be used to decide how many and where to establish data collection sites. This could include collecting secondary indicator data and conducting a brief ethnography to figure out where IDUs are located and what cultural factors may be important to consider in designing the procedures of the study. The finding also underscores the importance of including geographic markers (zip codes or census tracts) on surveys of IDUs to assess geographic reach.

The two samples differed substantially with respect to the proportion of participants who had utilized current prevention and care programs in San Francisco. Fewer RDS participants reported use of drug treatment and SEP, although SEP use was very high overall. This finding implies that RDS may be more effective than TS at reaching IDUs not receiving services. If the use of RDS is coupled with a proactive system of referral to services, the study can potentially bring IDU with less access to care into prevention and treatment programs.

The TS sample had a higher proportion of IDU who had reported “ever testing positive” for HCV, compared to the RDS sample. This finding might be because the TS sample included more IDU who had ever been in drug treatment. Those who have access to care are more likely to be tested for HCV. Another explanation is that the TS study provided HCV testing for participants from 1998 to 2001, and many in the 2005 TS sample had participated in the earlier cross-sections. HCV prevalence among IDUs in the TS study during those years was 91%.³⁴

There are several limitations to this study that need to be considered when interpreting its results. Although TS and RDS were both very effective in generating diverse samples, there is no way of knowing whether these samples are representative of the target population as a whole. TS has several limitations that should be noted. It requires that IDUs are part of a street culture that is easily accessed by an outreach worker, relying on the talents of the outreach workers who are involved in the ethnography and recruitment. For example, younger IDUs may not be interested in talking with or may not trust an older outreach worker. In RDS, it is possible to assess whether homophily exists and then corrects for it using weighting in RDSAT. Another limitation of the study is that the TS sample consisted of the 37th cross-section of a long-standing study, which may have generated a different sample than if it had been a first-time TS sample. The reputation of the study in the community may have biased the attributes of those who were willing to participate. RDS also has several limitations. It cannot access those who are not socially networked or isolated. For example, in an RDS survey of IDU in Cairo, Egypt, chains of referrals did not reach women.³⁵ In a survey of IDU in Tehran, Iran, the final sample lacked women and Afghan IDUs, despite empirical evidence supporting the existence of such groups.³⁶ Because RDS does not generally involve an intensive formative research phase, it is not easy to understand how the study procedures might bias who decides to participate.

There are also limitations common to both of these studies. Response bias is a limitation of all surveys, regardless of sampling methods, particularly when studying populations most at risk for HIV. Measuring the response rate is more challenging in surveys of IDUs, given that researchers usually are not present when study subjects are recruiting their peers to find out how many potential subjects are approached by recruiters. There were some differences in eligibility criteria in the two studies, which may account for some of the observed differences. Specifically, the RDS study was limited to English speakers while the TS study also included questionnaires in Spanish. However, in reality, no study participants in the TS study chose to be interviewed in Spanish. The time frame for injection criteria was 12 months in the RDS study and 30 days in the TS study. This may mean that some of the participants in the RDS study were less likely to be active IDUs. This could have biased the drug treatment estimates for example, even though drug treatment was less prevalent in the RDS sample. Finally, the RDS study required San Francisco residency. From our 20 years of experience conducting research with IDUs in San Francisco, we feel it is highly unlikely that more than 2% of the TS sample consisted of IDUs who reside outside of San Francisco. The majority of variables is self-reported and subject to recall bias and social desirability responses. However, we expect that these sources of biases would affect both samples similarly given that populations and type of questions were relatively comparable. Moreover, previous studies of IDUs have found good reliability with respect to the measures we used to assess the main outcomes in this study.^37,38 And finally, because of the different statistical methods needed for analyzing RDS and TS data, we were not able to combine the datasets and conduct multivariate comparisons to assess whether the data are significantly different. Instead, we relied on assessing whether the 95% CIs overlapped in the estimates.