Purpose: We proposed an analytic approach to estimate attributable fractions of grouped exposures by comparing exposure profiles of patients with SE infection to those infected with other Salmonella serotypes.
Methods: Patients living in FoodNet sites during January 2014 – October 2015 were asked about 48 food and environmental exposures. Missing exposures were imputed based on multiple imputation scheme. Related exposures were grouped into causal pathways for analysis by random forest model. Counterfactual simulation was conducted in which exposed statuses were changed into non-exposed ones and the influence of such changes on reduction of predicted illness was estimated.
Results: We examined data on 1,604 patients with SE infection and 5,667 infected with other serotype. The median proportion of missing information for the 48 exposures was 4% (range, 1% to 18%). The exposure frequencies of the imputed data were similar to those of the observed data. Analyses in which each of the three chicken exposures (consuming any chicken, chicken outside home, and consuming ground chicken) were removed from the model resulted in reductions of SE illnesses by 4%, 2% and 0%, respectively. However, removal of all three chicken exposures reduced SE illnesses by 14%, indicating interactions between nested exposures.
Significance: Our results suggest that case-case comparisons in a random forest model might provide estimates of the proportion of illness attributable to related exposures. Grouping all exposures related to a particular food category can provide important information for evaluating the impact of food safety policies and interventions that target that food.