Respondent Driven Sampling
2 likes1,748 views
Respondent Driven Sampling (RDS) is a technique for sampling hard-to-reach populations. It works by having initial participants (seeds) recruit a small number of people from their social networks, who are then eligible to recruit others from their networks. This process continues in successive waves. RDS relies on assumptions about network structure and recruitment behaviors. Analysis adjusts for network size and recruitment patterns. The technique was used to sample migrants in Morocco to estimate HIV, syphilis, and tuberculosis prevalence and understand their demographics, risks, and access to services. Results provided insights to guide health programs for this population.
Respondent Driven Sampling
- 1. Respondent Driven Sampling Lisa G Johnston, MA, Mph, PhD Affiliations: University of California, San Francisco; Tulane University of Public Health and Tropical Medicine www.lisagjohnston.com; www.respondentdrivensampling.com lsjohnston@gmail.com
- 2. Topics •Respondent Driven Sampling Background •Hard to reach populations •How it works •Assumptions •Migrants in Morocco •Innovations •Population Size Estimation (SS-PSE)
- 3. Respondent- Driven Sampling Background First described in 1997* Adopted by US CDC in early 2000s for use in HIV biological behavioral surveillance surveys† Endorsed by Global Fund, WHO and UNAIDS^ for sampling hard to reach populations Started to be noticed in late 2000s by other disciplines It is a sampling and analysis technique *Heckathorn DD. Respondent-driven sampling: a new approach to the study of hidden populations. Soc Probl. 1997;44(2):174–99. †Johnston LG. Conducting respondent driven sampling (RDS) studies in diverse settings: a training manual for planning RDS studies. Centers for Disease Control and Prevention, Atlanta. 2007. ^UNAIDS. Guidelines on surveillance among populations most at risk for HIV. 2012; WHO. Bio-behavioural survey guidelines for populations at risk for HIV. 2017.; Johnston LG. WHO Module 4: Introduction to respondent driven sampling. Introduction to HIV/AIDS and sexually transmitted infection surveillance. 2013.
- 4. Hidden and hard to reach populations •No sampling frame •Often want to remain hidden due to practices and behaviors that are illegal or stigmatized
- 5. Is the population of interest socially networked (i.e., they know each other as part of that population)? Can you find a small number of persons who are part of the population? Questions to ask yourself
- 6. What is RDS? Seed Wave 1 Wave 2 Wave 3 Wave 4 Wave 5 Wave 6 Wave 7 Wave 8 Wave 9 Wave 10 Wave 11 Wave 12 Wave 13
- 7. Assumptions in RDS • Respondents know each other as members of the target population (and these are reciprocal) • Respondents are linked by a network composed of a single component • Sampling occurs with replacement (population from which a sample is gathered is infinitely large) • Respondents can accurately report their personal network size (and this needs to be measured) • Peer recruitment is a random selection from the recruiter’s network • Each respondent recruits a single peer Heckathorn DD. Extensions of respondent-driven sampling: analyzing continuous variables and controlling for differential recruitment. Sociol Methodol. 2007;37(1):151–207; Johnston LG. Introduction to Respondent Driven Sampling. Introduction to HIV/AIDS and sexually transmitted infection surveillance. WHO, Geneva Switzerland; 2013. Available from: http://applications.emro.who.int/dsaf/EMRPUB_2014_EN_1686.pdf
- 8. Anatomy of a Survey Conducted using RDS Objectives were to measure: prevalence of HIV, Syphilis and TB • Socio-demographics • Sexual risk • use of and access to health and social welfare programs • HIV testing • The size of the population Develop recommendations to guide programs and expand services
- 9. Pre-survey research • Social network properties • Seed Selection • Acceptability of RDS and research in general • Survey Logistics (willing to undergo testing, venous blood draw, anal/vaginal swabs, etc.) Johnston LG, et al. Formative research to optimize Respondent Driven Sampling surveys among hard to reach populations in HIV behavioral and biological surveillance: Lessons learned from four case studies. AIDS Care. 2010. 22(6):784-92.
- 10. Survey components • Primary compensation: 30 Moroccan Dirham (~8 Euro) to complete both biological and behavioral parts of survey. • Secondary compensation: 30 Moroccan Dirham for each (a maximum of three) eligible recruit who completes survey. • Two interview locations
- 11. Eligibility • Eligibility for both surveys • 18 years or older • Living and/or working in Rabat • Residing three months or more in Morocco. • in an irregular administrative situation in Morocco (undocumented, asylum seeker, refugee) • Anglophone: • Originating from anglophone Sub-Saharan countries: Nigeria, Gambia, Ghana, Liberia, or Sierra Leone, etc. • Francophone: • Originating from francophone Sub-Saharan countries: Senegal, Cameroun, Mali, Cote d’Ivoire, DRC, Guinea, etc.
- 12. Measurement of network size • How many migrants do you know, and they know you who are native to Ghana, Liberia, Nigeria, Sierra Leone or Uganda who speak English, who have lived and/or worked in Rabat for 3 months or more? • How many of these persons above are in an irregular administrative situation? • How many of these persons are 18 years and above? • How many have you seen in the past 30 days?
- 13. Typical steps in a standard Bio-behavioral RDS survey (visit 1) Enrollment (valid coupon, screening, consent) Interview Testing (Pre-test counseling) Coupon explanation, give coupons, primary incentive Test results Lsjohnston.global@gmail.com, www.lisagjohnston.com Typical steps in an RDS survey (visit 2) Return to Interview site (2nd visit coupon, test voucher) Follow-up interview Optional: Biological test results Secondary incentive for each eligible recruit who completed survey
- 14. Survey Results • Survey lasted two months (March and April 2013) • 410 francophone (6 seeds) and 277 anglophone (5 seeds) Social network size Age Nationality Sex Max. number of recruits Max. number of waves Percent of sample Seed 1 23 32 Ivory Coast Female 84 7 20.7 Seed 2 10 30 Ivory Coast Female 39 6 9.8 Seed 3 500 43 Ivory Coast Male 88 6 21.7 Seed 4 12 26 DRC Female 75 6 18.5 Seed 5 10 38 Congo Brazzaville Female 59 5 14.7 Seed 6 3 25 Guinee Conakry Male 59 5 14.6 Seed 1 20 30 Gambia Male 5 3 2.2 Seed 2 17 35 Liberia Female 158 14 57.4 Seed 3 3 43 Nigeria Male 7 2 2.9 Seed 4 3 31 Cameroun Male 61 6 22.4 Seed 5 10 27 Nigeria Male 41 7 15.2 AnglophoneFrancophone
- 15. Francophone Anglophone Recruitment Graphs of Males and Females
- 16. RDS Analysis and diagnosis • RDS data must be adjusted by the inverse of each respondents’ personal social network size • Must collect data on who recruited whom • Software programs: RDS Analyst (www.hpmrg.org), RDSAT (www.respondentdrivensampling.org) • There are multiple estimators • Assess data for bottlenecks and convergence
- 17. Convergence-Sex Definition: measures progression of enrolling subjects to determine when the proportion for a characteristic approaches and remains stable in relation to the adjusted estimate. What it tells you: That your final estimate is independent from the bias of the non- randomly selected seeds. ANGLOPHONEFRANCOPHONE
- 18. Convergence-Nationality for Francophone 1=Cote Ivoire, 2=Congo Brazzaville, 3=DRC, 4= Guinee Conakry, 5=Cameroun, 6=Mali, 7=other
- 19. Bottlenecks-Sex Estimate precision is based on underlying network structure and traits within the structure. Assume that the structure makes up one complete social network component. Some structures have sub-populations not connected within the network structure. Especially problematic when traits differ between unconnected sub-populations in a sample. FRANCOPHONE ANGLOPHONE
- 20. Bottleneck-Nationality for Francophone 1= Cote I’voire, 2=Congo Brazzaville, 3= DRC, 4=Guinee Conakry, 5=Other, 6=Cameroun, 7=Mali
- 21. Sociodemographic Findings Francophone N = 410 Anglophone N = 277 n %, (95% CIs) n %, (95% CIs) Sex Male 276 63.9 (55.0, 72.8) 198 70.9 (63.8, 77.9) Female 134 36.1 (27.2, 45.0) 78 29.1 (22.1, 36.2) Religion Muslim 99 24.6 (17.4, 31.9) 24 9.2 (4.3, 14.2) Christian 281 67.9 (60.7, 75.1) 245 88.7 (83.4, 94.1) Catholic 17 4.2 (2.1, 6.4) 1 0.5 (-0.1, 1.0) Other 10 3.2 (1.1, 5.3) 4 1.5 (0.4, 2.7) Age group <25 101 25.9 (20.5, 31.3) 96 33.0 (27.2, 38.9) ≥25 309 74.0 (68.6, 79.5) 180 67.0 (61.1, 72.8) Education level Primary 124 32.9 (27.9, 38.0) 176 74.5 (68.2, 80.8) High School 136 34.0 (28.3, 39.8) 25 10.4 (6.3, 14.5) University 120 33.0 (26.8, 39.2) 32 15.1 (9.8, 20.4) Current marital Status Single 325 78.7 (73.5, 83.9) 196 71.3 (64.2, 78.5) Married 62 15.6 (10.7, 20.5) 76 27.3 (20.7, 33.8) Divorced, Widowed 23 5.7 (3.3, 8.1) 4 1.4 (0.2, 2.5)
- 22. Nationalities of Franco and Anglophone migrants 15.4 7.7 25.2 8.7 5.3 33.5 4.3 Cameroun Congo Brazzaville Ivory Coast Guinee Conakry Mali Democratic Republic of Congo Other 3.7 4.4 2.6 66.9 17.2 5.1 Ghana West Africa other Liberia Nigeria Cameroun East Africa
- 23. Differences Between Male and Female Francophone and Anglophone migrants Francophone Anglophone Females N = 134 Males N = 276 Females N = 78 Males N = 198 n %, (95% CIs) n %, (95% CIs) n %, (95% CIs) n %, (95% CIs) Main source of revenue Salaried 24 16.0 (7.9, 24.0) 128 46.4 (40.1, 52.8) 3 3.0 (2.7, 3.3) 79 39.3 (32.0, 46.5) Hair/ Massage 30 22.2 (11.5, 32.9) 7 2.6 (0.2, 5.0) 10 13.0 (1.8, 24.2) 4 4.0 (0.9, 7.1) Trades 12 5.6 (1.2, 10.1) 26 9.7 (5.5, 13.8) 1 1.6 (-0.2, 3.5) 18 9.1 (2.8, 15.5) NGO/ service 13 8.3 (3.5, 13.0) 18 5.7 (1.8, 9.6) - - 6 3.3 (0.5, 6.2) Begging 7 4.4 (0.9, 8.0) 11 2.9 (1.3, 4.5) 43 60.3 (45.0, 75.5) 49 24.7 (17.4, 32.0) Help from family/ friend/others 48 43.4 (31.9, 54.9) 85 32.7 (26.9, 38.5) 21 22.1 (12.8, 31.3) 42 19.6 (14.0, 25.1) Had sexual intercourse in exchange for money or gifts in past 12 months Yes 37 22.7 (16.1, 29.4) 20 19.4 (8.5, 30.3) 39 28.7 (19.2, 38.1) 6 12.4 (3.7, 21.1) Knows where to go to have HIV test Yes 95 70.8 (61.9, 79.7) 167 56.9 (49.3, 64.5) 25 32.1 (20.5, 43.8) 63 30.3 (24.1, 36.6) No 37 29.2 (20.3, 38.1) 109 43.1 (35.5, 50.7) 51 67.9 (56.2, 79.5) 135 69.7 (63.4, 75.9) Has had an HIV test in past 12 months (among all participants) Yes 64 52.1 (41.2, 63.1) 63 21.4 (16.4, 26.5) 34 49.7 (34.5, 64.9) 35 17.2 (11.6, 22.8) No 69 47.8 (36.9, 58.8) 213 78.6 (73.5, 83.6) 44 50.4 (35.2, 65.6) 163 82.8 (77.2, 88.4)
- 24. Infections among Francophone and Anglophone Migrants Francophone N = 410 Anglophone N = 277 n %, (95% CIs) n %, (95% CIs) HIV (ELISA, Discordant-Immunoblot; Tie breaker-PCR) Yes 17 3.4 (1.8, 5.0) 6 3.2 (0.0, 6.4) No 381 96.0 (95.0, 98.2) 262 96.8 (93.6, 100.0) Syphilis (VDRL and TPHA ) Yes 11 2.8 (0.2, 5.4) 1 0.3 (0.0, 0.6) No 387 97.2 (95.6, 99.8) 267 99.7 (99.4, 100.0) Tuberculosis (confirmed w/sputum) Yes 1 0.5 (0.2, 1.2) 1 0.2 (0.0, 0.4) No 397 99.5 (98.8, 100.2) 267 99.8 (99.6, 100.0)
- 25. RDS strengths •Assumed representative of the network sampled (understand the underlying network structure) •Can diagnose bias •Efficient for rare and hard to reach populations •Manuals and tools available to help you •Reaches less visible segments of population •Free computer software available
- 26. Recent Innovations •Estimators •Web-based RDS •Respondent Driven Intervention •Imputed Visibility† •Population Size Estimation (SS- PSE)* • Multipliers (unique object, service) † McLaughlin KR, Johnston LG, Japuki X, Gexha-Bunjaku D, Deva E HM. Inference for the visibility distribution for respondent-driven sampling. Manuscr Prep. * Handcock MS, Gile KJ, Mar CM. Estimating hidden population size using Respondent-Driven Sampling data. Electron J Stat. 2014;8(1):1491–521. Available from: http://projecteuclid.org/euclid.ejs/1409619420 Johnston LG, McLaughlin KR, Rhilani HE, Latifi A, Toufik A, Bennani A, et al. Estimating the size of hidden populations using respondent-driven sampling data: Case examples from Morocco. Epidemiology. 2015;26(6); Johnston LG, McLaughlin KR, Rouhani SA, Bartels SA. Measuring a hidden population: A novel technique to estimate the population size of women with sexual violence-related pregnancies in South Kivu Province, Democratic Republic of Congo. J Epidemiol Glob Health. 2017;7(1).
- 27. Population Size Estimation-successive sampling assumptions Option to use Visibility Imputation
- 28. Imputation of network sizes-FSW Yerevan Population Size Estimation: 2000 or 0.6% of the adult female population in Yerevan McLaughlin, KR, et al. Use and interpretation of population size estimations among hidden populations using successive sampling in respondent-driven sampling surveys: Case studies from Armenia. J. Med. Internet Research (in Press)
- 29. Resources and courses Resources • Johnston, L. G. Introduction to Respondent Driven Sampling. 2013. Geneva, Switzerland. http://applications.emro.who.int/dsaf/EMRPUB_2013_EN_1539.pdf • www.hpmrg.org (RDS Analyst Software) • www.respondentdrivensampling.org (RDSAT, references, RDS materials) • www.lisagjohnston.com; www.respondentdrivensampling.com (resource materials, journal articles, etc.) Upcoming courses • Tulane University of Public Health and Tropical Medicine, New Orleans, Louisiana, USA. January 7 to 11, 2019. For more information contact: lsjohnston.global@gmail.com • ECPR Methods Winter School, University of Bamberg, Germany. 25 February– 1 March 2019. Contact: https://ecpr.eu/Events/EventTypeDetails.aspx?Event TypeID=5
- 30. Thank You!