(P29-011-23) Structural Missingness and Algorithmic Bias in Supply Chain Stage and Food Product Records of Foodborne Disease Outbreaks: The National Outbreak Reporting System, 2009-2019

Objectives: As food regulatory agencies aim to leverage technology to create more reliable data repositories, a proactive approach is needed to detect and minimize algorithmic bias due to structural, or non-random, missingness and ineffective reporting. Using 9407 foodborne disease outbreak (FBDO) records from Centers for Disease Control and Prevention’s National Outbreak Reporting System in 2009-2019, we identified patterns of missingness in supply chain stage and food product variables and assessed their contribution to algorithmic bias on estimated duration and caseload.

Methods: We explored categories of missingness in supply chain stage and food product variables and identified eight mutually exclusive plausible patterns of missingness (P1-P8) among date-time variables, needed for illness, exposure, and global duration estimation. P2-P8 were missing at least one date-time variable: P5-P8 the illness end date, and P2-P4 either the exposure start or end date, or both. We examined the extent of algorithmic bias across categories and patterns of missingness by comparing durations and total caseload using pairwise t-tests with a Bonferroni correction and logistic regression models.

Results: Out of 9407 FBDOs, 13%(1222), 32%(3030), 42%(3907), and 13%(1248) occurred before, in preparation, were missing, and unknown, respectively. For 7111(76%) FBDOs, data on food products were categorized as missing (5456,58%), undetermined (353,4%), unclassifiable (51, < 1%), other (50, < 1%) or invalid (5, < 1%). 1 in 4 FBDOs were associated with land (958, 10%) or aquatic animal (670, 7%) or plant food products (668, 7%); 1196(13%) FBDOs reported multiple food products. Among 7037(75%) FBDOs with complete date-time data, 5803(82%) FBDOs were also missing records on supply chain stage or poorly defined food products and on average, had illness, exposure and global durations ~2 times shorter and ~1.5 times lower in caseloads (p < 0.001) than FBDOs with complete records.

Conclusions: While missingness in supply chain stage introduces potential algorithmic bias, the use of two broadly aggregated categories limits the understanding of food storage and distribution factors on outbreaks’ duration and magnitude. In contrast, seemingly more granular yet ineffective categorization, as seen with food products, may bias statistical inferences.

Funding Sources: DoD, NSF (#1855886)