Sampling hidden populations is particularly challenging by using standard sampling methods mainly because of the lack of a sampling frame. Respondent-driven sampling is an alternative methodology that exploits the social contacts between peers to reach and weight individuals in these hard-to-reach populations. It is a snowball sampling procedure where the weight of the respondents is adjusted for the likelihood of being sampled due to differences in the number of contacts. The structure of the social contacts thus regulates the process by constraining the sampling within subregions of the network. We study the bias induced by network communities, which are groups of individuals more connected between themselves than with individuals in other groups, in the respondent-driven sampling estimator. We simulate different structures and response rates to reproduce real settings. We find that the prevalence of the estimated variable is associated with the size of the network community to which the individual belongs and observe that low degree nodes may be undersampled if the sample and the network are of similar size. We also find that respondent-driven sampling estimators perform well if response rates are relatively large and the community structure is weak, whereas low response rates typically generate strong biases irrespectively of the community structure.