OBJECTIVES: To evaluate the effect of drotegogin alfa (DA) on bleeding events in patients with severe sepsis and multiple organ failures. METHODS: A pre-post design was conducted before and after DA's market introduction. An optimal propensity score matching method was undertaken to control for unbalanced characteristics. Several models were tested to explain the number of bleedings events. The more usual ones are the Poisson and the negative binomial (NB) models. Contrary to the NB model, including a dispersion parameter, the Poisson model supposes the mean equals the variance. An alternative consists on modelling separately the probability of experiencing at least one bleeding (with a binomial model) and the number of bleeding events (with a count model). Double-hurdle models suppose that, once a threshold has been exceeded, patients experience at least one bleeding event. In zero inflated models, both models can predict an event absence. Non-nested models were compared with the Vuong statistic. RESULTS: The matched sample includes 840 patients. Bleeding events were experienced by 17.6% of patients, 13.6% in the before and 21.7% in the after phase (p=0.0021). The mean number of bleedings was higher in DA treated patients (0.28 against 0.18, p=0.0208). The standard NB model fitted better than the double-hurdle NB model (p<0.0001) and was similar to the zero-inflated NB model (p=0.6815). We kept the NB model, the simpler one. Moreover, the dispersion parameter was significant (p=0.0013), favouring the NB to the Poisson model. In this multivariate model, patients in the after phase were still more at risk of experiencing bleeding events. Other risk factors included the presence of a central catheter infection and a high LODS score. CONCLUSIONS: DA use in addition to the conventional treatment leads to more bleeding events. In our study, over-parameterised models did not bring more information than simpler ones.