Docstoc

RDS Methodological Challenges

Document Sample
RDS Methodological Challenges Powered By Docstoc
					Lisa Johnston, Tulane University School of
Public Health and Tropical Medicine and
Center for Global Health Equity
Lsjohnston.global@gmail.com
   Target population must be socially networked;
   Starting with the seeds, participants allowed to
       recruit no more than a pre-specified number of
       recruits;
   Recruitment chains must be long so sample
       estimators stabilize, thereby reaching
       “equilibrium”;
   During data collection, must gather information
       about who recruits whom and the size of each
       participant’s social network;
   RDS data must be analyzed to account for biases
       inherent in chain referral sampling;
   Incentive for participation and recruiting peers.
   A recent meta analysis on RDS studies conducted in
    non U.S. countries found several studies suffering
    from operational, design or analytical challenges
    including
       inadequate measurement of social network sizes
       sampling from insufficiently networked population
       applying inadequate design effect (or no DE)
       combining samples from separate geographical areas
       improperly analyzing RDS data
       Inappropriate incentives
Johnston, L., et al. (2008). AIDS and Behavior; 12(suppl 1):131-141.
Malekinejad, M. et al. (2008). AIDS and Behavior; 12(suppl 1):105-130.
   Social network sizes set up the probabilty of
    selection into the sample
     Include ALL components of eligibility in the social
         network size question
       Break up the social network size question into several
         questions
       Develop systematic probing
       Ensure adequate training for staff*****
       Someone’s social network size cannot be zero
       Use short time frames in the question to reduce recall
         bias
   Dense social networks are necessary to sustain
      recruitment and ensure long recruitment
      chains
     Conduct sufficient formative research to
       understand social networks
     Make sure the populations interact on a variety
       of levels (multiplexity)
     Discover potential bottle necks
   The DE accounts for sampling variability
      including degree outliers and homophily and
      essentially makes your sample size larger
     DEs should be no less than 2, but more might be
       better
     Use larger DEs if you suspect bottle necks or weak
       social networks between some key groups of
       interest
Two distinct networks with few connections
(Red dots depict infection)

                                             Goel and Salganik, 2007
   An RDS sample must be comprised of one social
    network
     RDS is not suitable for producing one country level
       estimate if sampling from several geographically
       separate locations
     Analysis is based on strict tracking of who recruited
       whom—there are coupon tracking systems which
       allow you to track seed recruitment and maximum
       number of waves
     Even within the same city when using multiple sites,
       networks may not cross over
   Incentives are used for participation and for
    recruiting peers
     Some incentives are too high (results in coupon
       bartering/selling, non-target group members trying to
       participate, and repeated attempts to enroll)
     Keep incentives small
     Consider offering services such as clinical exams,
       testing with results and treatment, etc. (e.g.,
       asessment in Bangladesh, Dom Rep, India)
     Remember that peer pressure is an additional ‘incentive’
   Without proper analysis RDS is just a very good
    snow ball sample
     RDS data must be analyzed to account for homophily,
          differential social network sizes and recruitment
          patterns
       There is free software available to analyse RDS data
          (www.respondentdrivensampling.org)
       latest version of RDSAT is v. 6.0
       RDSAT is tedious to use
       RDSAT provides adjusted estimates and confidence
          intervals
       RDSAT is currently under revision and is an area that still
          needs development
   Without proper analysis RDS is just a very good snow
    ball sample
     RDS data must be analyzed to account for homophily,
          differential social network sizes and recruitment patterns
       There is free software available to analyse RDS data
          (www.respondentdrivensampling.org)
       latest version of RDSAT is v. 6.0
       RDSAT is tedious and difficult to use
       RDSAT provides adjusted estimates and confidence
          intervals
       RDSAT is currently under revision and is an area that still
          needs development
       No consensus on how to conduct regression with RDS data
   Include this question in the survey: “What is the
       main reason why you accepted a coupon and
       enrolled in this study?”
     Responses to include: to receive an HIV (or other
      infection) test and results, to receive the incentive, to
      receive a clinical exam, because the person who gave me
      this coupon asked me to enroll, I wanted to do
      something socially valuable, nothing better to do, etc.
 Include a short questionnaire at follow-up to
     assess who refused coupons and why
 Find target population members to serve as an
     advisory group during data collection (e.g.,
     India, Zanzibar)

				
DOCUMENT INFO