A likelihood-based approach supporting identification of contaminated food using sales data: The effects of shopping behavior on performance

Objective: Rapid identification of contaminated food is vital to minimize illness and loss in an outbreak. A likelihood-based approach, mapping geo-coded sales data against geocoded confirmed case reports, has the potential to reduce the time required for investigation.[1] The likelihood “score” defines a binary classifier defining a set of potentially “guilty” products which contains the contaminated source with high accuracy. In this work we report the effects of consumer travel on the performance of the method.

Study Design: We use 3 years of weekly sales data for 580 anonymous food products from a German retailer covering 3,513 of Germany's 8,235 postal areas. Replacing the simplest assumption where outbreaks occur where products are sold, [1] a more accurate model accounts for the fact that consumers do not only shop where they reside. A retail “gravity” model is used to test the effects of travel behavior [2].

Conclusions: ROC curves measure “sensitivity” based on the fraction of true positives vs. fraction of false positives. The area under the ROC curve (AUC) was measured for outbreaks with varying case number and with the gravity model “travel exponent”, k, varied over the expected range for grocery shopping. The performance of the method degrades as consumer travel distance increases. However, the degradation becomes less pronounced as more case reports are included in the analysis. For k = 3, the classifier performance having observed 50 cases versus only a single case dropped 6.2 %, while for k = 1 performance dropped 36.3 %.

[1] Kaufman J, Lessler J, Harry A, Edlund S, Hu K, Douglas J, Thoens C, Appel B, Kasbohrer A, PLoS Comput Biol 10(11): 2014

[2] Huff DL (1963) A probabilistic analysis of shopping center trade areas. Land Economics 39 (1) 81–90. doi: 10.2307/3144521