Waimar Tun is a program associate with the Horizons Program at the Population Council who manages the implementation of operations research activities in Africa, South America, and Asia. Her research focuses on HIV and sexually transmitted infections among at-risk populations, including men who have sex with men, sex workers, injection drug users, and mobile populations. To meet these aims, she utilizes both qualitative and quantitative methodologies, integrating behavioral research with epidemiology. She holds a Ph.D. and M.H.S. in epidemiology from Johns Hopkins University.

In this interview Waimar Tun, program associate at Horizons, shares details of the program’s experience with Respondent-driven Sampling, a technique for sampling hard-to-reach members of a population.

Q: What is Respondent-driven Sampling?

A: Respondent-driven sampling (RDS), developed by Douglas Heckathorn (Heckathorn 1997), is a way of sampling hard-to-reach and sometimes hidden members of a population through a technique similar to snowball sampling. The RDS method relies on individuals recruiting a limited number of their peers, who in turn recruit a limited number of their peers, and so on, continuing along this recruitment “chain” until the sampling goal is reached. A special feature of RDS is that it takes into account that study participants were not recruited randomly and uses statistical weights based on the participants’ network size (i.e., the number of people that the participants knows who would be eligible for the study) and recruitment patterns to yield estimates of characteristics of the target population and the confidence interval around the estimate.

Q: In what way is RDS different from snowball sampling?

A: It is different statistically and operationally. Statistically, because snowball sampling is a convenience sample—as the participants are selected by connection to another member already in the study and not randomly from a sampling frame—the sample may not be representative of the larger population and could lead to biased estimates. Convenience samples only allow you to make subjective guesses about how precise these sample proportions are when using them to infer to the larger target population. In other words, with any kind of convenience sampling such as the snowball method, inferences of the target population may not be valid. Unlike snowball sampling, RDS can potentially produce unbiased population estimates.

It is also different operationally in a couple ways. First, RDS limits the number of peers that a participant can recruit into the study (typically 3–4 recruits per participant). This is done so that recruitment will not be biased by relying on only a few individuals who are extremely successful at recruiting; it essentially equalizes the opportunity for those with large networks as well as those with small networks to recruit their peers. The limitation on recruits thereby allows for greater penetration into the target population and yields a sample that is less influenced by the characteristics of the set of initial participants who were non-randomly selected to start the recruitment process.

Second, unlike snowball sampling, RDS uses an incentive system: the first incentive is given to the participant for completing the study, and the second incentive is given to the participant for each peer the participant successfully recruits into the study. These incentives, while not imperative, can improve the chances of successful recruitment. Typically, monetary incentives are used, but non-monetary incentives such as toiletries, night club entrance tickets or non-perishable food items, have been used successfully.

Q: In what situations is RDS the preferred method?

A: This method is particularly useful when recruiting hard-to-reach populations or hidden populations, such as drug users, men who have sex with men, or commercial sex workers. It is extremely challenging to develop a non-biased representative sample for hard-to-reach populations who are typically highly stigmatized and therefore “hidden”. Because these populations are hard to identify, developing a sampling frame (i.e., lists of homes in the community, schools in the district, or hospitals in the province) from which to draw a random sample is not generally feasible. As such, we are not able to obtain a representative sample of these groups by using traditional random sampling techniques such as household surveys or facility-based surveys.

Other random sampling procedures do exist, such as time-location sampling but since some members of these populations do not tend to congregate at public venues, it makes it difficult to do time-location sampling. Even if time-location sampling was feasible for some of these populations that do congregate in public venues (i.e., street-based sex workers, men who have sex with men at nightclubs), the sample obtained might be biased because the sample would exclude those who avoid public places.

Because it is difficult to construct a sampling frame for these hard-to-reach populations, RDS is now the preferred method of sampling if the goal is to obtain an unbiased sample of these hard-to-reach populations.

Q: What are the advantages of the method?

Recruitment chains from a RDS study of men who have sex with men in Ciudad del Este, Paraguay. Arrows indicate the direction of recruitment. Photo credit: Waimar Tun

A: The key advantage of RDS’s recruitment method is that it allows researchers to access, in a systematic way, members of traditionally hard to reach target populations who may not otherwise be reached (i.e., those who don’t hang out in public venues, access facility-based services or who are not in contact with outreach workers). RDS is based on the belief that people are best recruited by their own peers due to greater trust, familiarity, or peer influence.

Because RDS is a probability sample, it can yield population estimates as mentioned earlier. It does this by taking into account that people sampled using RDS have different network sizes (i.e., circle of friends or sexual partners) from which to recruit, that there is a strong tendency to recruit others like themselves, and that some people recruit more than others. The analysis adjusts for these factors to produce unbiased population estimates. For example, in the Horizons study of men who have sex with men in the triple-border city of Ciudad del Este in Paraguay, the sample proportion of men who have sex with men who identified themselves as homosexual was 24 percent. However, when the analysis was weighted, the estimated population proportion of homosexually-identified men who have sex with men was only 12 percent (95% CI: 7–18). The reason the population estimate was half that of the sample proportion was because homosexually-identified men who have sex with men had on average larger network sizes than non-homosexually identified men who have sex with men, and they were also more likely to recruit other men who have sex with men who also identified themselves as homosexuals. As a result, the sample over-represented homosexually identified men who have sex with men, but the RDS analysis statistically adjusted for this over sampling of homosexually-identified men who have sex with men and down weighted their contribution to the sample to obtain a population based sample. RDS also has the potential for rapid recruitment because every participant becomes a recruiter. So for each subsequent participant, there is potential for exponential growth in recruitment. This is particularly true if you have large networks and strong ties among members of the networks. RDS can be quite successful at rapid recruitment in dense urban environments.

Q: What are the disadvantages of this method?

A: While there is potential for rapid recruitment, it is also possible that recruitment can be very slow if your participants are simply not recruiting their peers. This can happen for a variety of reasons, including small network sizes, lack of ties among members of the target population, lack of enthusiasm to participate, inadequate incentives, and high levels of stigma. So essentially, the real disadvantage here is that recruitment rate is unpredictable. For example, we did not anticipate that it would take over 12 months to recruit roughly 600 men who have sex with men in a behavioral study we conducted in Campinas, Brazil (with free syphilis and optional HIV tests) despite indications from a formative assessment that it would take much less time.

Another challenge right now is that the analysis of data collected using RDS can be difficult since it must take into account weighting for network size and recruitment patterns. Currently, there is a software called RDSAT (RDS Analytical Tool), which is designed to analyze data collected using RDS. The software, available at www.respondentdrivensampling.org, provides only basic statistical estimates such as estimates of proportions and cannot handle more complicated statistics at this time. It also lacks user-friendly ways to easily import and export data. Experts in this methodology are continually making refinements to the analysis software and holding discussions to address some of the analytical issues, including sample size required, design effects, checking if assumptions are met, conducting regression analyses, and assessing statistical significance between groups.

Q: Why do you think the recruitment of men who have sex with men was slower than you expected in Campinas?

A: We conducted a qualitative assessment to identify and understand the reasons for the slow recruitment process. This assessment indicated that getting a free HIV rapid test was an attractive feature for some; however, it may also have deterred others from participating due to fear of taking the HIV test and potentially having to face a positive result. Additionally, the fear of disclosing their sexual orientation remained a strong barrier to participation. In fact, when men who have sex with men who successfully recruited other participants into the study were compared to those who did not recruit anyone into the study, the non-recruiters were less likely to belong to gay NGOs, suggesting that they may have been less comfortable identifying and being open about their sexual orientation. Essentially, slow recruitment was likely due to people being scared of testing for HIV and/or being uncomfortable disclosing their sexual orientation to peers and study staff.

Q: What did you do to boost recruitment?

A: First, we contacted owners of nightclubs and bartenders at gay bars and clubs to help enlist new seed recruiters. Because these people know their clientele well, they are well positioned to help identify men who have sex with men who are considered to be ‘sociometric stars’ or ‘good seeds’. In other words, bartenders and nightclub owners were able to identify men who have sex with men who were popular and influential, and had a large network of friends in order to spread the word about the study and use their social standing to influence peers to participate in the study. Another strategy the investigators used was to advertise in strategic locations such as nightclubs and strategic events, such as the Campinas Gay Pride Parade. Posters and staff informed potentially eligible persons to participate in the study if they had ever received a recruitment coupon for the study even if the coupon had expired already, extending the recruitment period.

Q: What contributions has Horizon’s research made in the use of RDS?

A: The implementation of the study in Brazil (PDF of Report) contributed substantially to the development of the national guidelines for the implementation of the national HIV surveillance in high-risk populations. Further, the assessment of reasons for participation and non-participation will contribute to a greater understanding of some of the operational challenges of using RDS. Additionally, both the studies in Brazil and Paraguay (PDF of report) using RDS have contributed to a better understanding of risk behaviors and HIV seroprevalence among men who have sex with men as they were among the first studies to provide population estimates.

 

Heckathorn, D. 1997. "Respondent-driven sampling: A new approach to the study of hidden populations," Social Problems 44(2): 174–99.

Back to text