This study focuses on evaluating statistical models for count data that exhibit both over-dispersion and zero inflation, emphasizing the suitability of zero-inflated approaches. Among the models considered, including Poisson, negative binomial, zero-inflated Poisson (ZIP), zero-inflated negative binomial (ZINB), and hurdle models, the ZINB model emerged as the most appropriate due to its superior performance and its capacity to distinguish structural zeros from random zeros. Using data collected from 433 culinary students in Thailand, the models were applied to a real-world case involving kitchen accidents in culinary education settings. While the hurdle model yielded similar values for AIC, BIC, and -2LL, it lacked the ability to distinguish between structural zeros and zero-counts from at-risk individuals, highlighting a clear advantage of the ZINB approach. The ZINB model revealed that male students exhibited lower accident rates than female students. Students in Western kitchens had fewer accidents compared to those in Asian kitchens, which served as the baseline, whereas those in bakery kitchens were more likely to fall into the structural zero group. Moreover, the academic year was associated with reduced accident rates, possibly due to increased experience or improved safety adherence. Although the relationship between GPAX and accident frequency was negative, it was not statistically significant, whereas stress showed a significant positive effect on accident occurrence. These findings also emphasize the value of applying advanced zero-inflated models to analyze count data with excess zeros and over-dispersion and present useful information for the development of specific safety measures and the improvement of training programs in culinary institutions.
Rungjindarat et al. (Wed,) studied this question.