Respondent-driven sampling (RDS), a link-tracing sampling and inference method for studying hard-to-reach populations, has been shown to produce asymptotically unbiased population estimates when its assumptions are satisfied. However, some of the assumptions are prohibitively difficult to reach in the field, and the violation of a crucial assumption can produce biased estimates. We compare two different inference approaches: design-based inference, which relies on the known probability of selection in sampling, and model-based inference, which is based on models of human recruitment behavior and the social context within which sampling is conducted. The advantage of the latter approach is that when the violation of an assumption has been shown to produce biased population estimates, the model can be adjusted to more accurately reflect actual recruitment behavior, and thereby control for the source of bias. To illustrate this process, we focus on three sources of bias, differential effectiveness of recruitment, a form of nonresponse bias, and bias resulting from status differentials that produce asymmetries in recruitment behavior. We first present diagnostics for identifying types of bias and then present new forms of a model-based RDS estimator that controls for each type of bias. In this way, we show the unique advantages of a model-based estimator.