Targeting SNAP Quality Control with a Predictive Model: A Randomized Pilot in Connecticut

• In Event: Leveraging Data and Partnerships to Help States Navigate SNAP Policy Changes

Abstract

The One Big Beautiful Bill Act of 2025 introduces substantial cost-sharing requirements forstates with SNAP payment error rates (PERs) above 6%, exposing state agencies to significantfinancial liability beginning in FY 2028. With Connecticut’s PER currently above this threshold,their Department of Social Services (DSS) faces urgent pressure to identify and correct errors.In collaboration with DSS, we developed, and are now piloting, a predictive model designed totarget the state’s local quality control (QC) reviews toward the highest-risk cases in an effort todrive down the overall PER.

This work has two main contributions: (1) the construction of a case-level error-risk predictionmodel using four years of federal QC determinations linked to state eligibility data, and (2) thedesign of a randomized pilot to rigorously evaluate the model’s effect on SNAP payment errors.Using federal QC reviews between 2022 and 2025 merged with longitudinal eligibility records,we trained a classifier to predict whether newly certified or recertified cases contain a SNAPpayment error. The model prioritized operationally meaningful signals: volatility in earned,unearned, and SSI income; changes in shelter and dependent-care deductions; addresschanges; and signals of case complexity. Within evaluation data, the highest-risk tenth of caseshad a 33% error rate and accounted for over 20% of total payment error, while the lowest-risktenth contained virtually no errors. To assess model fairness prior to deployment, we evaluatedits calibration and false positive rates by race and ethnicity.

To evaluate whether this strong statistical performance translates into real-world impact, we areimplementing the model through a three-month randomized pilot. Each week, we will randomlyassign cases to either: (1) be assessed by staff based on the state’s existing rules-basedprocess, or (2) to be assessed by the model for errors. For cases flagged by each process wewill measure the share of cases with a confirmed benefit correction above the federal thresholdand the average dollar change per case reviewed.

The pilot offers a template for how states facing cost-sharing penalties can pair administrativedata and experimental methods to evaluate operational interventions quickly and credibly. Wewill present the modeling framework, pilot design, and early implementation lessons, withparticular attention to the tradeoffs between rapid deployment and rigorous evaluation thatstates will need to navigate in light of these new cost-sharing requirements.

Author

• Anthony Lollo, Yale University
  Presenting Author