Another Nail in the Coffin? Moving Away from Population-Based Case-Control Studies during Outbreak Investigations

BACKGROUND:   Population-based case-control studies (CCS) have been a mainstay of foodborne outbreak investigations. Because CCS are resource intensive and slow, other methods have been developed for testing associations between suspected foods and illness. We compared CCS (odds ratios) to Bernoulli trials (binomial probabilities) using existing food exposure data in two outbreak investigations.

METHODS:   We conducted population-based CCS during a 2013 Shiga toxin-producing E. coli (STEC) outbreak associated with cucumbers from restaurant chain A and a 2014 Salmonella outbreak associated with ground beef. We then calculated binomial probabilities for the suspected food item from each outbreak and compared resources used and outcomes. 

RESULTS:   In the STEC outbreak, all 8 cases interviewed with a detailed food exposure questionnaire recalled eating at outlets of a single restaurant chain (Chain A). To test the association between illness and eating at chain A, controls were enrolled using sequential digit dialing. Eight interviewers made 574 calls to enroll 12 controls (mean 47.8 calls per control enrolled). The odds ratio for eating at chain A was undefined with a p-value of 0.0007. In a second CCS using online orders from chain A to determine specific food items, 112 controls were enrolled in 175 phone calls (mean 1.6 calls made for each control enrolled). This second CCS implicated cucumbers (odds ratio undefined, p-value of 0.002). Using a background exposure rate of 2%, from Minnesota Department of Health’s sporadic STEC cases, the binomial probability of 8/8 cases eating at chain A is p<0.0001. Using the 2006/2007 FoodNet Population Survey exposure rate of cucumbers of 46.7%, the binomial probability of 8/8 cases eating cucumbers is p=0.002. In the Salmonella outbreak, controls were enrolled using geographically based phone lists created using whitepages.com. Ten interviewers made 560 calls, enrolling 16 controls (mean 35 calls per control enrolled). Additionally, staff spent 15 hours creating the phone lists. Consuming any ground beef (OR=11, p<0.04) and purchasing Chain B ground beef (OR=10, p=0.01) were associated with illness.  Relevant exposure rates for ground beef included: 32% (Oregon Salmonella cases) and 40% (Colorado FoodNet Population Survey, 2007) with binomial probabilities that 11/12 cases consumed ground beef of p<0.00001 and p=0.0003, respectively. 

CONCLUSIONS:   These two outbreak investigations demonstrate that using binomial probabilities with available background exposure rates may yield the same results as population based CCS while expending less time and effort and fewer resources. Identifying representative background exposure data can be challenging, as food consumptions varies regionally.