DescriptionIn quantitative microbiological risk assessment (QMRA), risk estimates are a function of the variability in concentrations of the bacterial pathogen in the food product concerned. The choice of the probability distribution describing this variability may therefore be important. The lognormal distribution is a common choice, which is however complicated by the observation of relatively high proportions of zeros and low numbers in microbial counts. Discrete distributions have been indicated as alternatives to model counts with a considerable amount of low numbers, and their zero-inflations counterparts are considered appropriate to model counts with many zeros. In this study, we analyze the effect of the use of different distributions fitted to (artificial) data on the risk estimates in QMRA, using an existing QMRA model. A distribution of Campylobacter jejuni counts in chicken at retail was simulated and 9 datasets of 500 microbial counts were constructed, with 9 different proportions of zeros (10 to 90%). The outcome of fitting a Negative Binomial (NB) versus a zero-inflated NB (ziNB) distribution to each of those samples was assessed, as well as the resulting estimates for the probability of illness (Pill). For each data set under evaluation, the fitted NB and ziNB distributions showed equivalent mean concentrations. However, the ziNB distributions showed higher variances, with lower probabilities for low counts and extended right tails. This difference was sharper for datasets with higher proportions of zeros. There was a growing departure between the computed Pill depending on the selected frequency distribution as the amount of zeros increased. These results show that the choice of the distributions fitted through count data may have an important impact on the risk estimate. Guidelines for the most appropriate choices to be made in this context are discussed.
|Period||13 Sep 2011|
|Event title||7th International Conference on Predictive Modelling of Food Quality and Safety: null|