"They got their program, and I got mine": a cautionary tale concerning the ethical implications of using respondent-driven sampling to study injection drug users.

BACKGROUND This article examines the ethical implications of using respondent-driven sampling (RDS) to conduct HIV behaviour surveillance among injection drug users (IDUs) in Chicago. Ethnographic inquiry illustrates how the design and implementation of RDS invites if not promotes manifold violations of federal guidelines governing human research subject protections. METHODS Post hoc structured interviews with approximately 13% (n=70) of the behaviour surveillance sample (N=529) focused on how RDS's "dual incentive" structure affected participants' social, economic, and cultural milieu. Triangulated methods include interviews with owners of 20 "shooting galleries", unofficial and illegal locales where IDUs congregate and 400 h of traditional ethnographic observation of individual IDUs and IDU networks. "Consensus analysis" allows identification of key cultural domains that define the RDS coupon market. RESULTS The study reveals the power of RDS to foment a stratified market of research participation that reinforces pre-existing economic and social inequalities among IDUs. Participants co-opted RDS to develop various "underground" revenue-generating modalities that produced differential risks and benefits among participants. Deleterious outcomes include false advertising regarding the study's risks and benefits, exploitation of relative economic deprivation, generation of sero-discordant social networks, and interpersonal and organised conflict, coercion, and violence. CONCLUSION Although RDS may involve serious ethical violations it remains the best available means for accruing a representative sample of hidden populations. It is critical, however, to supplement RDS with research into (1) the subjects' cultural, social, economic, and political contexts, (2) the potential human subjects violations that participants experience, and (3) how these two issues might affect data integrity and interpretation.