Taxi drivers are a group with high driving exposure and are involved in a significant number of urban traffic casualties. Using two modelling approaches, this study examines whether the intention to speed, as measured by the Theory of Planned Behaviour (TPB), can better fit a crash frequency model than errors or lapses as measured by the Driving Behaviour Questionnaire (DBQ). Data from 1000 drivers in Tehran was collected through questionnaires. The crash prediction model included a cross-sectional model using negative binomial (NB) regression methods and a tree regression model from a previous study. In the last three years, the drivers had been involved in 544 road crashes, and of those, 42 resulted in serious injuries. Due to the rare and random nature of crashes, the empirical Bayesian (EB) method was used for model testing. Comparing AIC and BIC showed that zero-inflated NB (ZINB) models performed better. The final selected model was the intention-based ZINB model without the age variable. The coefficients for intention, exposure, and driver experience were 0.205, 0.103, and −0.443, respectively. The high EB coefficients indicated strong reliance on predicted crash values. The conclusion is that road crashes are closely related to taxi drivers’ intention to speed rather than errors and lapses. This indicates that it can be described as a traffic violation, rather than a mistake. Therefore, significant efforts are required to increase compliance with speed limits and reduce road crashes. Further education and high-quality campaigns are essential elements to achieve this goal.