A 2006 respondent driven sampling (RDS) survey of injecting drug users (IDUs) in Bristol, UK, estimated 40 per 100 person years HCV incidence but in 2009 another RDS survey estimated only 10 per 100 person years incidence amongst the same population. Estimated increases in intervention exposure do not fully explain the decrease in risk. We investigate whether the underlying contact network structure and differences in the structure of the RDS trees could have contributed to the apparent change in incidence.

Method
We analyse the samples for evidence that individuals recruit participants who are like themselves (assortative recruiting). Using an assortativity measure, we develop a Monte Carlo approach to determine whether the RDS data exhibit significantly more assortativity than is expected for that sample. Motivated by these findings, a network model is used to investigate how much assortativity and the structure of the RDS tree impacts sample estimates of prevalence and incidence.

Results
The samples suggest there is some assortativity on injecting habits or markers of injecting risk. The 2009 sample has lower assortativity than 2006. Simulations of RDS confirm that assortativity influences the estimated incidence in a population and the structure of RDS samples can result in bias. Our simulations suggest that RDS incidence estimates have considerable variance, making them difficult to use for monitoring trends.

Conclusions
We suggest there was likely to have been a decline in risk between 2006 and 2009 due to increased intervention coverage, but the bias and variance in the estimates prevents accurate estimation of the incidence.