Reducing SNAP Payment Errors through Predictive Modeling: A Randomized Trial
How to cite
Lollo, Anthony; Wallace, Jacob. 2026. "Reducing SNAP Payment Errors through Predictive Modeling: A Randomized Trial." AEA RCT Registry. AEARCTR-0019325. https://doi.org/10.1257/rct.19325-1.0
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
This study is a two-arm randomized controlled trial that tests whether a predictive model—which incorporates a wide variety of case features—is better able to identify cases at risk for SNAP payment errors than a current rules-based process that primarily selects cases for review based on benefit amount. Over a 4-month period we will randomize approximately 2,500 cases in a 1:1 ratio to: (1) control (business-as-usual case selection process, N=1,250), or (2) treatment (predictive model case selection, N=1,250). After cases are selected by either process, they will go through a standardized Quality Assurance (QA) review in which all information related to income and expenses is reviewed. During this process, errors are corrected which may result in changes to the initial benefit determination. We think cases selected for review by the predictive model will be more likely to have underlying payment errors and will have larger errors. As a result, there will be a larger extent of benefit changes resulting from QA reviews for the cases selected by the predictive model.
Intervention
Arm 1: Control
Cases are selected for review using the state’s pre-existing manual “ABC” selection criteria. This selection process prioritizes reviews based on benefit amount, with priority given to cases with benefits > $400, then cases receiving between $200 and $399, and the lowest priority for review given to cases receiving < $200 in benefits. In addition to this benefit prioritization, reviewers are given discretion to select cases if any information pertaining to the top 6 error elements is “questionable or error prone”.
Arm 2: Treatment
Cases will be selected for review based on the model output from predictive models. The models are cross-validated XGboost classifiers that use historic federal SNAP payment error data and state case review data to identify the case characteristics most likely to predict whether a case has a payment error and whether the state’s current review process can fix those errors. Cases will be selected for review based on a composite scoring function which incorporates predictions of both models. Using this composite scoring function, the top 60 cases will be selected for review each week.
Experimental design
This study is a two-arm, randomized controlled trial designed to evaluate the effectiveness of different strategies for selecting cases for QA review within Connecticut’s SNAP program. Each week, cases are assigned to each arm in roughly equal proportions (1:1)
Design details
Cited by
No citation data available for this registration yet.