| |
|
Page 3
Respondent Driven Sampling
There may be some populations we want to sample that do not hang out at a specific place. In the case of intravenous drug users (IDU's) not only is there no suitable location for TLS, but the population is also not likely to self-identify. Either of these factors make TLS unsuitable for sampling. Respondent Driven Sampling (RDS) has the advantage that no location is needed and can also achieve wider coverage of the population. The method is based on the non-probability technique of Snowball Sampling, but is rendered a probability sample with mathematical sampling theory. We'll continue with the example of IDU's.
Back in City X we are trying to determine the proportion of IDU's that are HIV+. We have data from a needle exchange in the city, but we don't think it is reliable because we know many drug users do not participate. In order to get an unbiased sample with wide coverage of the population we decide to use RDS.
Sample Size
Software is currently being worked on for sample size. Like before, size is determined by the level of precision desired, but with RDS bias becomes a concern as well. Tests have shown that a sample of 500 is biased by .oo1, while a sample of 200 is biased by .002. The amount of bias for both samples is negligible, but different populations will produce different amounts of bias. For now, it is recommended that at least 200 people are sampled and for this example we will set the sample size at 300. 1/sample size
Selecting Seeds & Referral
In RDS the word "seed" refers to an individual belonging to the target population who is interviewed and then recruits other members of the population to be interviewed. Every study must begin by finding people in the community. The inevitable bias that results from this non-random selection can be controlled through recruitment. The study can be thought of in waves. The first wave is finding people in the community. The second wave are people they recruit. The third wave is recruited by the second, and so on. If enough waves are reached and there is a sufficient sample size, a probability sample is achieved and bias is not a problem. As a general rule, a minimum of six waves is needed and it is desirable to have as many waves as possible. In order to achieve the desired amount of waves we must limit the amount of people recruited by each individual in the study. This is done by giving each person unique coupons for recruitment, which also enable us to map who is being recruited by whom (this will be needed for analysis of the data).
In order to achieve our desired sample size of 300 we will have to balance the number of initial seeds with the number of referrals allowed. Assuming we start with 5 seeds and give each of them 2 coupons for recruitment, the minimum amount of waves before reaching our sample size will be:
mathematical formula
In order to analyze data from RDS three characteristics of each respondent must be known.
A computer program that takes these characteristics into account to organize a study is RDS Coupon Manager (download here). Click here for a brief tutorial.
Analysis
RDS software