OBJECTIVE: Discrete choice models are widely used in pharamacoeconomics. If correctly applied, bootstrapping is a useful tool for these models because small sample distributions of the dependent variables are not known. In this paper, we will show how to apply bootstrapping to have consistent and efficient estimators under discrete choice models. METHOD: Four common bootstrapping techniques were analyzed: paired, non-parametric, parametric, and wild bootstrapping. The extension of parametric bootstrapping for linear regression to parametric discrete choice models is presented: Let U be the probability that the binary dependent variable y=1. Then for each application we choose y*, which is the new independent variable for each bootstrap, from Bernoulli distribution with probability of success given by U. RESULTS: The Market Scan® private insurance database was used in this study. The analytic sample comprised 36,341 individuals with asthma whose healthcare was provided under a variety of fee-for-service (FFS), fully capitated, and partially capitated health plans. We estimated hospitalization for FFS and non-FFS asthma patients. Logit models were selected depending on the distribution of the dependent variable. The Pearson chi-square goodness of fit test (p=0.3742) and the Hosmer and Lemeshow test (p=0.2904) suggested that the model fit well. Treatment patterns had no significant effect on hospitalization after controlling for demographic and clinical factors. The illness severity of the patient (proxied by the number of three-digit ICD-9 codes), however, had a positive and significant effect on hospitalization. We would not have seen this significant effect if we had chosen paired, non-parametric, or wild bootstrapping as a way to bootstrap standard errors. CONCLUSION: Despite the obvious benefit of bootstrapping in discrete choice models, the method should not be used blindly. Once the model is estimated under parametric assumptions, as in logit or probit models, deviations of the assumptions for bootstrapping will yield inefficient estimators.