Applying Statistical Methods to Reduce Product Recalls

Recent product recalls of meat and other foods have received wide media coverage, placing the efficacy of government inspection and food industry quality assurance systems under intense public scrutiny. The U.S. Dept. of Agriculture, Food and Drug Administration, and industry have been working hard to eliminate pathogens from the food chain, and the relative rarity of food poisoning outbreaks originating from processed foods testifies to their success. But even under the best manufacturing practices, total elimination of pathogens in many foods, fresh meats, and poultry, in particular, is impossible because the animals and the workers that handle the foods will continue to harbor them. Also, cross contamination can occur almost anywhere—from the field to the supermarket, restaurant, or the consumer’s home.

Heat destroys harmful bacteria like Salmonella and E. coli O157:H7. But there is always a possibility that improperly cooked food (with a viable pathogen present) will be consumed somewhere, resulting in acute poisoning. Therefore, reducing the initial microbial load to the lowest possible level is a sensible and effective strategy.

Despite efforts to eliminate foodborne pathogens and reduce the number of microorganisms of all kinds, there is always a finite probability that sometime in the future they will be encountered at a dangerous level without any warning signs. The reason is that such an occurrence would be the result of a random coincidence of unnoticed or undocumented events, each insufficient to cause a problem by itself. Under certain circumstances, the probability of such random coincidences when conditions remain unchanged can be estimated directly from the microbial counts’ fluctuation pattern (Peleg and Horowitz, 2000; Peleg, 2006). (Known malfunctions of equipment, refrigeration system, or similar accidents need no modeling.)

The estimation method is based on identifying the counts’ distribution, which in most cases is expected to be lognormal, finding its parameters by the Method of Moments (or Maximum Likelihood Estimation), and using them to calculate the probability of future counts exceeding any level deemed unacceptable. The method’s predictive ability has been demonstrated with actual industrial records (Peleg et al., 2000; Corradini et al., 2001) and made available to industry as a free downloadable Microsoft Excel® program at http://www.unix.oit.umass.edu/~aew2000/Mic-CountProb/microbecounts.html.

All the user has to do is paste the counts record and choose up to five high count levels. The program will automatically test the counts’ independence and their distribution’s normality or log-normality, plot their histogram and autocorrelation function, calculate their distribution parameters, and display the probabilities of encountering a count exceeding the user’s chosen levels. This and similar programs, based on other parametric distribution functions, can estimate the probabilities of outbursts having a magnitude not previously recorded. They can assess the efficacy of improved sanitation or prophylactic measures not only in terms of lowering the mean count, but also by how much they have reduced the probability of potential future problems—the two are not the same! In addition, such programs can be used to simulate microbial fluctuation patterns and examine their potential food safety implications.

Counts records having segments of zero entries or showing aperiodic outbursts having variable magnitude and duration require more elaborate mathematical models and different programs (Hadas et al. 2004; Engel et al., 2001; Corradini et al., 2009). But these too can be used to estimate the frequencies of outbursts if conditions are not improved, at least in principle—see http://www.unix.oit.umass.edu/~aew2000/MicPopExpl/MicPopExplModel.html, for example.

To date, the food and meat industries have shown little interest in the concept’s development or implementation. Consequently, most microbial records, collected at great expense, are subjected only to rudimentary analysis. They are then kept on file for a while before being discarded, without any attempt to translate the information that they contain into the probability of future problems.

Like all events governed by probabilities, especially very low ones as in our case, no one can be certain whether a past microbial outburst could have been avoided, and if or when the next one will occur. But in light of the cost of a recall and an actual food poisoning outbreak, the food industry should think twice about expanding the application of statistical methods to interpret their microbial records and using them to identify and quantify the risk of a recall.