Attribution of Foodborne Illnesses, Hospitalizations, and Deaths to Food Commodities by using Outbreak Data, United States, 1998–2008

Each year, >9 million foodborne illnesses are estimated to be caused by major pathogens acquired in the United States. Preventing these illnesses is challenging because resources are limited and linking individual illnesses to a particular food is rarely possible except during an outbreak. We developed a method of attributing illnesses to food commodities that uses data from outbreaks associated with both simple and complex foods. Using data from outbreak-associated illnesses for 1998–2008, we estimated annual US foodborne illnesses, hospitalizations, and deaths attributable to each of 17 food commodities. We attributed 46% of illnesses to produce and found that more deaths were attributed to poultry than to any other commodity. To the extent that these estimates reflect the commodities causing all foodborne illness, they indicate that efforts are particularly needed to prevent contamination of produce and poultry. Methods to incorporate data from other sources are needed to improve attribution estimates for some commodities and agents.

, , ⋯ , , ⋯ , where indexes the etiologic agents, are the public health burdens (e.g., numbers of illnesses, hospitalizations, or deaths), and are column vectors with elements equal to the numbers of reported cases in simple and complex outbreaks respectively, and are 0/1 matrices corresponding to simple and complex outbreaks of etiology i with rows indexing outbreaks and columns corresponding to whether or not a given commodity was represented by ingredients of the contaminated food or foods from the outbreak. In some instances, one or both of and will be null matrices because no outbreaks of a given etiology were reported in that category. In addition, when is null, is defined to be null and the term in the summation is defined to be 0. When a column of S i is null then the corresponding column of is null. This follows from the general rule that no illnesses from complex food outbreaks due to a given etiologic agent will be allocated to commodities that are not represented among the simple food outbreaks due to that agent. The vector is a column vector of 1's with dimension 17.
Estimated proportions ( of all illnesses attributed to the 17 defined commodity groups are given by related equations: , , ⋯ , , ⋯ , Note that the minimum estimated totals sum to less than the total estimated number of illnesses, ∑ , and the maximum estimated totals sum to more than ∑ , when there are attributable illnesses due to complex foods. The same is true of the minimum estimated proportions and maximum estimated proportions; they sum to less than and more than 1,  for simple foods, the same concepts apply to source attribution for complex foods. Differences between outbreak-associated illness-based and outbreak-based attribution proportions reflect differences in outbreak sizes among commodities. The degree to which one or the other measure, outbreak-associated illnesses or outbreaks, reflect illnesses as they are caused by commodities in overall domestic foodborne illness is a subject for future research.

Expanded Methods
In this appendix, as a complement to the technical appendix, we provide a narrative description and examples of key elements of our method. To estimate illnesses attributable to specific commodities from reports of outbreaks of foodborne illnesses, we 1) attribute illnesses to specific commodities for each etiologic agent, and 2) sum the etiology-specific estimates, weighted by estimates of the number of illnesses (i.e., illnesses, hospitalizations, or deaths) for each etiology.

Attributing illnesses to specific commodities for each etiologic agent
To determine etiologic agent-specific attribution, we wanted to sum the number of illnesses from a complex food outbreak to each commodity associated with the outbreak, as long as that commodity was also implicated in a simple food outbreak caused by that agent (which establishes the commodity as a possible causal exposure). This last allocation counts illnesses multiple times, but that is consistent with the maximum estimate as providing an upper bound for the number of illnesses attributed to individual commodities.
In Table A, we illustrate the attribution of illnesses in a dataset of four hypothetical outbreaks of illnesses caused by Etiologic Agent X. Illnesses in simple food outbreaks in which ground beef, lettuce, and apple juice were implicated were attributed to the commodities beef, leafy vegetables, and fruits-nuts, respectively, for all three estimatesminimum, MP, and maximum. Outbreak D was due to a complex food so no illnesses from this outbreak were included in the minimum estimate. For the maximum estimate, all six illnesses were attributed to the beef commodity (because the vehicle contained ground beef) and all six were also attributed to the leafy vegetables commodity (because the vehicle contained lettuce), but no illnesses were attributed to the vine-stalk vegetables commodity (although the vehicle contained tomato) or the grain-beans commodity (the vehicle contained bread). This is because the dataset contained at least one simple Etiologic Agent X outbreak attributed to the commodity beef, and at least one simple Etiologic Agent X outbreak attributed to the commodity leafy vegetables, but no simple Etiologic Agent X outbreak was attributed to either vine-stalk vegetables or grains-beans.
In this example, to partition the illnesses in Outbreak D into the most probable number of illnesses for each commodity, we determined the proportion of illnesses in simple food outbreaks caused by that agent that were attributed to any commodity included in the hamburger sandwich: of the illnesses in the simple food outbreaks due to these commodities, 69% were attributed to beef and 31% to leafy vegetables. We applied these proportions to the six illnesses in Outbreak D, which yielded 4 illnesses attributed to beef and 2 attributed to leafy vegetables. The crude percentage of Etiologic Agent X illnesses attributed to each commodity was calculated by summing the number of attributed illnesses and dividing by the total number of actual illnesses in all Etiologic Agent X outbreaks. Note that although the actual number of illnesses was 46, only 40 illnesses were attributed to commodities for the minimum estimate and 52 were attributed for the maximum estimate; only the MP estimate counted each illness once and only once.

Summing the etiology-specific estimates, weighted by estimated number of domestically-acquired foodborne illnesses for each etiology
To calculate the total number of illnesses attributed to each commodity, we summed the etiologic agent-specific estimates obtained by applying the proportion of illnesses for each commodity to the estimated number of domestically acquired foodborne illnesses. In Tables B1-B3, we illustrate the calculations for the number of illnesses and deaths in a dataset of two hypothetical etiologies (X and Y). In Table B1, the minimum, MP, and maximum estimates of illnesses attributed to each commodity were calculated as above, and shown as a percentage of the total for each etiology. In Table B2  †Illnesses in outbreaks in which the implicated food was simple were included in the minimum, maximum, and most probable estimates; illnesses in outbreaks in which the implicated foods were complex were included only in the most probable and maximum estimates. ‡Hamburger sandwich ingredients: ground beef, lettuce, tomato, bread. ** For Outbreak D, the MP estimate of illnesses due to each commodity relies on information from the simple food outbreaks due to those commodities in the dataset. The total number of outbreak-associated illnesses caused by Etiology X and due to simple foods contained in the hamburger sandwich was 32, with 22 (69%) due to beef and 10 (31%) due to leafy vegetables. Rounding to the nearest integer, the MP estimate of illnesses in outbreak D attributed to beef was 4 (69% of 6) and the MP estimate of illnesses attributed to leafy vegetables was 2 (31% of 6).