School Nutrition and Student Discipline: Effects of Schoolwide Free Meals

National school-level panel from the Civil Rights Data Collection (biennial, 2010-2016) linked to CEP participation data, using state-by-state timing of Community Eligibility Provision pilot rollout (2012-2015). Matched sample about 49,074 elementary and 15,150 middle schools; the preferred elementary specification uses 34,437 observations across 11,430 schools and the preferred middle specification 7,138 observations across 2,625 schools.

US public elementary, middle and high school students in high-poverty schools adopting schoolwide free meals. CEP-eligible schools are a disadvantaged subset of all US schools, so heterogeneity contrasts within this sample are compressed relative to the national distribution.

Difference-in-differences with school and year fixed effects on within-school changes in suspension rates as states phased in the CEP pilot; treatment is defined as being in a state ELIGIBLE to participate in year t, which leverages only immutable state-level timing rather than school-level selection into the programme. Standard errors clustered by district. Credible identification, and the authors explicitly reject regression discontinuity (reported categorical-eligibility shares bunch at the 40% and 62.5% thresholds) and instrumental variables (estimates are unstable across instrument choices because the suspension distribution is highly skewed). THE HEADLINE IS SPECIFICATION-DEPENDENT and our original extraction did not say so. In the full sample with school and year fixed effects the effect on multiple out-of-school suspensions is NULL for elementary schools (-0.0905, se 0.0642) and marginal for middle (-0.263, se 0.157); restricted to CEP-eligible schools it is null for both (elementary -0.0633; middle +0.0673). Significance appears only after two further restrictions: limiting the sample to schools that ever participated in CEP, and dropping 17 large districts (Chicago, Los Angeles, Philadelphia, Miami- Dade and others) that changed formal discipline policy during the study window. The main outcome is also narrower than "suspensions": it is the rate of students with MULTIPLE out-of-school suspensions, a severe and rare event. On any out-of-school suspension the middle-school effect disappears entirely. The authors are explicit that they cannot observe dietary intake, so any nutritional interpretation is inference, not measurement; stigma reduction and household income relief are equally consistent mechanisms, and they say they cannot distinguish them.

In the preferred (twice-restricted) specification, modest reductions in the rate of students with multiple out-of-school suspensions for elementary students (-0.32 per 100, about a 15% reduction from a 2.17 baseline, p < 0.01) and middle-school students (-0.48 per 100, about 6% from a 7.65 baseline, p < 0.1), with no effect at high school - and, on the broader "any out-of-school suspension" measure, a marginally significant INCREASE for high school students (+0.72 per 100 on a 14.78 baseline, about +5%, p < 0.1) that our original extraction missed and that contradicts Kho 2018 finding of the largest reductions in high schools. For elementary students the reductions concentrate in counties with above- median predicted child food insecurity (-0.54 v -0.16 per 100; interaction -0.379, p < 0.01); the equivalent middle-school contrast is similar in magnitude but imprecise (F-test p = 0.087). Race subgroups show no significant effects on their own. The authors frame this as evidence that school nutrition services matter beyond test scores.

Low for selection into treatment (state pilot timing, and treatment defined by state eligibility rather than school choice). But as with all programme-level designs the treatment bundles food, household income relief and stigma reduction, and the paper cannot separate them.

Two prior single-state studies of the same policy on the same outcome disagree with each other, and this paper reproduces BOTH depending on the sample. Kho (2018), comparative interrupted time series in Tennessee, found a 1.2 percentage-point (about 10%) reduction in students ever suspended or expelled, LARGEST IN HIGH SCHOOLS. Comperatore and Fuller (2018), difference-in-differences in North Carolina, found NO effect on the share of students suspended at any level, in aggregate or in high-poverty or urban subsets. Gordon and Ruffini find modest elementary and middle-school reductions in their preferred restricted sample, a marginally significant INCREASE for high schools (contradicting Kho), and the North Carolina null when they broaden to all eligible schools. So: mixed, not unreplicated.