Title: Regression Modelling for Respondent-Driven Sampling
Authors: Michael Rotondi - York University (Canada) [presenting]
Lisa Avery - University Health Network (Canada)
Abstract: Respondent-driven sampling (RDS) is a relatively new technique used to recruit participants from hard to reach (hidden) populations. However, due to the statistical complexities of RDS, a number of methodological questions, including regression, remain unanswered. A simulation study was performed to evaluate the validity of various regression models that could control for the dependency between participant responses and unequal sampling probabilities in RDS. Networked populations with varying levels of homophily and prevalence, based on a known distribution of a continuous predictor were simulated and RDS samples were drawn from each population. Weighted and unweighted binomial and Poisson regression models, with and without various clustering controls were modelled for each sample to evaluate model validity. Our motivating example, examining factors associated with prevalent cardiovascular disease among the Indigenous community in Toronto is also discussed. Type-I error rates were unacceptably high for weighted regression models, dependency within the data was in general inconsequential. Even when the reported degree is accurate, as in this simulation, a low reported degree can unduly influence regression estimates. Based on the simulation results, unweighted regression should be used with RDS data and sample clustering can be ignored, at least under conditions of moderate homophily.