The Challenge of Evaluating Universal School Meals: A Cautionary Tale on Recall-Window Mismatch and Limited Pre-Treatment Data

Abstract

Between 2022 and 2023, nine U.S.\ states adopted universal free school meals policies, providing breakfast and lunch to all public school students regardless of family income. This paper investigates whether these policies reduced household food insecurity beyond direct child beneficiaries through a "resource reallocation" mechanism. Using the Current Population Survey Food Security Supplement (2022–2024), I estimate difference-in-differences models comparing households with school-age children in treatment versus control states. The naive TWFE point estimate of 4.7 percentage points (SE = 2.0 pp, 95% CI: [0.9 pp, 8.5 pp]) on a restricted sample of 2023 adopters versus never-treated states is statistically significant but meaningless as a causal estimate—it does not identify a treatment effect because the 12-month recall window does not align with survey-year treatment coding. More informatively, a triple-difference specification with state$\times$year fixed effects comparing households with versus without school-age children yields a precisely estimated null effect ($-0.8$ pp, SE = 1.3 pp, 95% CI: [$-3.4$ pp, 1.8 pp]). With only 2–3 years of data, no true pre-treatment periods for 2022 adopters, and policy adoption coinciding with major post-pandemic economic shifts, credible causal inference is not possible with this data structure. This paper contributes a concrete illustration of how recall-window mismatch invalidates standard DiD designs, with implications for researchers using survey data with rolling reference periods.