Universal free school meals and child body mass index: impact among children in community-based health care organizations

Abstract

Introduction

The Community Eligibility Provision (CEP) is a federal policy that allows high-poverty schools to provide universally free breakfast and lunch to all children. Providing universal free meals has potential to decrease childhood obesity, but so far, studies are limited and findings mixed.

Methods

We used electronic health record data from a large network of community-based health care organizations and linked school-level data paired with extended 2-way fixed-effects models for staggered policy adoption to compare child body mass index z-scores (BMIz) from schools that adopted CEP to eligible, nonadopting schools.

Results

The sample consisted of 149 052 distinct lower-income children who attended a balanced panel of 1085 schools in 12 states. Mean age was 10.8 years, with 84% being publicly insured, and plurality race/ethnicity being Hispanic (43.1%). Children in CEP schools gained less in BMIz compared with children in eligible, nonadopting schools (difference-in-differences: −0.02; 95% CI: −0.04, −0.004), with estimates becoming more negative over time. However, we also found some evidence of heterogeneity by the year of adoption with increases in BMIz in some specifications.

Conclusion

This study builds on evidence suggesting that, for some low-income populations, universal free school meals are associated with relative decreases in BMI.

Introduction

Universal free school meals (UFM) policies allow eligible schools and districts to provide free breakfast and lunch to all students, with the goal of improving child food security and health.¹ The Community Eligibility Provision (CEP) is a federal UFM policy that became available to higher-poverty schools nationwide in 2014 as part of the Healthy Hunger Free Kids Act (HHFKA). The CEP has high reach, providing access to free meals for 23.6 million children attending participating schools in 2023–2024.² Evidence strongly suggests that such policies increase participation in school meals and reduce food insecurity.^3-5 Evidence also suggests that the improved nutrition standards, also enacted by the HHFKA, resulted in improved dietary quality for children participating in the National School Lunch Program, and that this effect was stronger for lower-income compared with higher-income children.⁶ However, studies that examine whether UFM policies such as CEP improve health outcomes are just beginning to emerge.

A key health indicator of interest in association with UFM is childhood body mass index (BMI)/obesity status, since this was a health outcome targeted in the HHFKA and because childhood obesity remains a serious public health issue.^7,8 Universal free school meals may affect BMI directly by prompting individuals to substitute healthier foods for unhealthier foods by eating school meals.^6,9-13 Additionally, by providing 2 meals a day at no cost, the CEP may indirectly affect BMI by decreasing food insecurity, which has been linked to increased obesity risk in some youth populations,¹⁴ or by increasing household disposable income, thus potentially reducing household stressors or enabling purchase of other health-promoting goods or services.^15-25 Thus, CEP has the potential to be a sustainable and equitable obesity-prevention policy since it targets high-poverty schools and addresses a social determinant (household budgets and time allocation) by providing potentially health-promoting meals (breakfast and lunch) free of charge. At the same time, some scholars have argued that UFM has the potential to widen health disparities because the lowest-income children already have access to free or reduced-price (FRP) school meals via individual applications and UFM may offer a new benefit to middle- and higher-income populations.

A few studies have previously examined child BMI or obesity status in association with UFM programs. The findings are mixed, but most studies found a decrease in BMI or obesity for some, but not all, population groups. Furthermore, groups experiencing beneficial associations have not been consistent between studies.^26-30 For instance, on the question of whether only higher-income populations (who were not previously eligible for FRP meals) benefit from UFM while lower-income populations do not, a study of UFM in New York City schools found that only higher-income populations (>185% federal poverty level [FPL]) experienced a reduction in obesity,²⁶ while a nationwide study of early implementation of CEP found that only lower-income students experienced a decrease in overweight.²⁸ Other studies have examined different heterogeneities, and found heterogeneity in the CEP–BMI associations by school level. A school-level analysis of CEP-eligible schools in New York State found a decrease in obesity among middle-school students, with larger effects among students in rural schools.²⁷ A study in an Atlanta metro-area school district found increases in BMI in association with CEP that were also strongest among middle-school students.²⁹ A school-level analysis among public and charter schools in California and found decreases in obesity prevalence in association with CEP and did not examine heterogeneities in effects.³⁰

A key limitation of these studies is that nearly all have been limited to 1 city or state. We build on this evidence by pooling data derived from electronic health records (EHRs) from community-based health care organizations from multiple municipalities across 12 US states. Our sample consists primarily of children with lower family incomes. While we cannot directly answer whether lower-income children benefit more than higher-income children with this sample, we can investigate whether low-income children experience any impacts on BMI. Representation of lower-income children is vital for understanding the impacts of this policy on children who are often omitted from research due to language, time, or resource barriers to participation in traditional study designs, or due to a lack of health insurance in most EHR-based studies. We additionally leverage recent improvements in difference-in-differences (DID) methods to better account for staggered adoption and potential time-heterogeneous effects.

Methods

Additional policy details

Without a UFM policy, lower-income children at schools participating in the National School Meals Programs (95% of public and charter schools) are individually eligible to receive FRP meals. However, many income-eligible children do not participate.³¹ Community Eligibility Provision eligibility is based on a school’s or district's “identified student percentage” (ISP), which is the percentage of children who are identified in state or federal systems as having low family incomes through participation in other means-tested programs or through being in foster care or homeless. Until 2024, schools or districts with an ISP over 40% were eligible for CEP. Once schools/districts participate, they can remain participating for 4 years without recertification. Meals served in CEP-participating schools are reimbursed by the federal government at the established rate for free meals based on 1.6 times a school's ISP. This means that only at an ISP 62.5% or greater would a school be reimbursed for 100% of meals served. We use school (rather than district) ISP greater than 40% to define eligibility for our comparison group.²

Key data sources

OCHIN data warehouse

The OCHIN data warehouse contains EHRs from the OCHIN (not an acronym) network of 1239 Community-based Healthcare Organizations across 21 states. We used outpatient data including height, weight, payer type, date of birth, race and ethnicity, preferred spoken language, and residential address collected at each clinical encounter.

School Attendance Boundary Assessment Survey

We used the School Attendance Boundary Assessment Survey, which contains information on the catchment area and open-enrollment participation (schools that accept students via mechanisms other than residential assignment) for each school³² to estimate the residentially assigned school for each child. We geocoded children's addresses and assigned them to the non–open-enrollment (residentially assigned) school in whose catchment they lived. When more than 1 set of boundaries encompassed a child's address, we assigned them to the nearest non–open-enrollment school. If only open-enrollment schools were within the catchment area, we assigned the child to the closest school.

Additional data sources

We summarized school characteristics of school size, school level, and percentage of children eligible for FRP meals, which were obtained from the longitudinal yearly National Center for Education Statistics (NCES) Common Core of Data (CCD) files. Social deprivation index was obtained from the Robert Graham Center. Rural–urban continuum codes, which categorize counties by the size of their urban population and/or adjacency to a metropolitan area, were obtained from the US Department of Agriculture Economic Research Service.

Exposure

Our exposure was CEP. We considered schools who adopted the policy as treated and schools that were eligible to participate but did not do so as the comparison. School CEP eligibility and participation came from multiple sources: state Departments of Education, NCES CCD, Food Research and Action Center, as well as from a replication dataset made publicly available from another researcher, prioritized in this order.⁵

Outcome

We used child-level BMI z-score (BMIz) as a proxy measure for adiposity, using the 2022 updated Centers for Disease Control and Prevention extended BMI growth charts.³³ We extracted clinically measured height and weight data from the OCHIN data warehouse and excluded implausible values identified by growthcleanR.³⁴ We interpolated missing height or weight values when 1 measurement was present but not the other for 2% of height observations and 0.2% of weight observations (Materials S1). For each child and school year (SY) in which they had data, we used the BMI measure closest to June 30 if multiple measures were available.

Covariates

Child-level characteristics potentially associated with BMI included as covariates were as follows: child race/ethnicity (Hispanic, non-Hispanic [NH] Black, NH-White, NH-Asian, NH-American Indian or Alaska Native, NH-Native Hawaiian or other Pacific Islander, multiple racial identities, unknown racial identity), preferred spoken language (English, Spanish, other), health insurance type (public, private, uninsured), and child age at BMI measurement (calculated from birth date; partial months).

Exclusion/inclusion criteria

We assembled a balanced panel of schools followed throughout the study period from SYs 2013–2014 through 2018–2019, reducing potential bias due to compositional change that could occur if we allowed schools to enter and leave the sample each year. We balanced the sample at the school level rather than the student level since the treatment happens at the school level.

Among 11 534 schools matched to 416 033 participants with height and weight observations measured at an OCHIN clinic, we excluded 6774 schools that were never eligible for CEP (Figure S1). We required that schools have linked patient data during the school year (August 15 through June 30; 94 schools excluded), with non-missing, valid height and weight values (66 schools excluded). We also required that schools have linked patient data for the entire study period (3335 schools excluded), be unexposed to a UFM program prior to 2014 (3 schools excluded), and have non-missing ISP after imputation (2 schools excluded; Materials S1). We excluded all 175 schools that adopted CEP in school year 2018–2019 since we could not achieve parallel or near-parallel pretreatment trends (Materials S1, Figure S2).³⁵

Statistical analysis

We estimated means and frequency distributions of key characteristics for the sample overall and by cohort (year in which schools opt into CEP). We plotted mean BMIz by cohort.

To assess whether school-level CEP participation was associated with children's BMIz, we used a staggered DID model called extended 2-way fixed effects³⁶ implemented with the Stata package “wooldid” (Materials S1), with the event study specification.³⁷ Data are organized such that the baseline/referent year for both pre- and post-period estimates is the year before treatment starts (t = −1). We assessed parallel pre-trends between comparison and intervention schools by testing whether DID estimates at each of the pre-policy years differed from zero. Our preferred treatment effect estimate was the pooled estimate for all post-treatment years. We also present separate treatment effects for all post-treatment years.

All models included the covariates listed above, plus age-squared, sex, an age-by-sex interaction, and an age-squared-by-sex interaction to control for compositional change in age within school (Materials S1). Within child, age and insurance status were time-varying. Including school fixed effects enabled within-school estimates of associations and controlled for all baseline, time-invariant school-level factors that might otherwise confound the relationship between treatment and outcomes. Year fixed effects controlled for any common secular trends in the outcome. Standard errors accounted for clustering at the school level. In secondary analyses, we examined treatment effects by ISP (<62.5% vs ≥62.5%), grade level, and cohort.

In robustness checks we (1) checked that results were not driven by selection bias in attending OCHIN clinics or imputation of some CEP scores; (2) assessed pre-trends in a subsample of the primary balanced panel of schools that also had measurements 2011–2012 and 2012–2013, which was to enable assessment of pretreatment trends for the 2014–2015 cohort since this was the largest cohort but had only 1 pretreatment period observed; (3) limited our sample to only children in households with incomes less than 185% FPL to check whether results were consistent for this group who would have been individually eligible for FRP meals; (4) controlled for whether and when states adopted direct certification by Medicaid; and (5) allowed for multi-way clustering of standard errors.

Results

Overall, the sample consisted of 149 052 distinct children who attended 1085 schools in 12 states. A total of 47.8% of children in the sample had repeated BMI measurements. Table 1 presents means and frequencies of sample characteristics overall and by cohort, including eligible schools that never opted into treatment. The sample was 49% male with a mean age of 10.8 years, 43.1% Hispanic, 26.8% White, and 13.6% Black. The preferred spoken language was English (61%), followed by Spanish (31%). The majority (84%) had public insurance.

Table 1.

Patient and school characteristics overall and by year of Community Eligibility Provision adoption.

	Overall	Year schools adopted CEP
		Never	2014–2015	2015–2016	2016–2017	2017–2018
Patient characteristics
Total no. of patients	149 052	65 120	62 814	8805	14 598	8341
Male, %	49.1%	47.4%	50.2%	48.9%	50.4%	51.1%
Age, mean (SD), y	10.8 (3.7)	12.3 (3.8)	9.8 (3.3)	11.2 (3.6)	9.8 (3.1)	9.3 (2.6)
Race or ethnicity, %
Hispanic	43.1%	32.7%	53.1%	28.7%	25.2%	80.4%
Non-Hispanic White	26.8%	40.2%	19.8%	34.6%	14.8%	4.5%
Non-Hispanic Black	13.6%	8.3%	10.4%	20.8%	48.6%	7.6%
Asian	4.4%	5.8%	4.4%	3.4%	1.1%	2.0%
American Indian or   Alaskan Native	0.5%	0.6%	0.4%	0.7%	0.1%	0.2%
Native Hawaiian or   Other Pacific Islander	0.7%	0.9%	0.7%	0.7%	0.2%	0.3%
Multiple races	1.6%	2.0%	1.4%	1.8%	1.5%	0.4%
Unknown	9.4%	9.6%	9.9%	9.5%	8.6%	4.7%
Primary language
English	61.3%	70.6%	51.5%	60.8%	76.1%	47.5%
Spanish	31.1%	22.0%	40.6%	20.4%	19.6%	49.7%
Other	7.6%	7.4%	7.9%	18.8%	4.3%	2.8%
Household percentage of federal poverty level
0%–50%	27.4%	29.0%	37.2%	60.1%	16.9%	31.1%
50%–130%	29.4%	35.2%	29.2%	24.2%	39.2%	31.9%
130%–185%	8.4%	7.9%	6.1%	2.7%	8.4%	7.5%
185%–200%	1.6%	1.2%	1.0%	0.3%	0.9%	1.2%
>200%	8.7%	5.5%	5.0%	1.3%	2.1%	6.1%
Missing	24.5%	21.2%	21.5%	11.4%	32.6%	22.2%
Health insurance type, %
Public (eg, CHIP)	84.1%	78.1%	88.0%	79.6%	85.4%	95.6%
Private	8.5%	13.3%	5.6%	10.2%	4.5%	2.7%
Uninsured	7.4%	8.5%	6.3%	10.1%	10.3%	1.7%
BMI z-score, mean (SD)	0.79 (1.1)	0.73 (1.1)	0.82 (1.2)	0.67 (1.2)	0.83 (1.1)	0.97 (1.1)
School characteristics
Total no. of schools	1085	444	460	71	78	32
Patients per school per year, mean (SD)	44.2 (78.0)	40.9 (85.6)	43.5 (63.8)	32.2 (43.6)	58.8 (102.9)	89.3 (114.1)
School type, %
Elementary	68.9%	59.7%	77.6%	63.4%	76.9%	65.6%
Middle	15.8%	21.2%	11.5%	16.9%	6.4%	21.9%
High	13.7%	17.8%	10.0%	18.3%	12.8%	3.1%
Other	1.6%	1.4%	0.9%	1.4%	3.9%	9.4%
Total no. of students, mean (SD)	661.8 (501.0)	727.6 (626.8)	624.0 (389.4)	647.3 (498.9)	538.3 (207.8)	623.5 (326.6)
Baseline percentage of students eligible for free or reduced-price meals, mean (SD)	76.9% (15.5%)	65.7% (12.8%)	81.2% (12.5%)	79.2% (14.2%)	95.6% (7.7%)	85.8% (7.3%)
No. of schools in district, mean (SD)	111.7 (236.1)	46.9 (130.0)	197.2 (312.0)	36.6 (70.7)	46.5 (45.1)	36.8 (72.4)
Social Deprivation Index score,a mean (SD)	71.9 (21.6)	63.3 (21.8)	79.4 (18.3)	74.2 (20.4)	74.6 (19.6)	70.0 (25.7)
Rurality,b mean (SD)	1.7 (1.5)	1.9 (1.7)	1.6 (1.3)	1.9 (1.5)	1.5 (1.1)	1.6 (1.0)

Source: Authors’ analysis of data from OCHIN, US state Departments of Education, National Center for Education Statistics, Food Research and Action Center, Robert Graham Center, and US Department of Agriculture, 2013–2019. The table shows total number, percentage, or mean and SD of patient and school characteristics overall, and by cohort (year that schools adopted the CEP). The sample includes 1085 schools followed from school year 2013–2014 through 2017–2018, matched to 149 052 distinct patients.

^aSocial Deprivation Index is a composite measure of area-level deprivation, measured at the census tract level; scale is from 1 to 100 with 100 indicating the highest level of socioeconomic deprivation.³⁸

^bRurality is measured at the county level on a scale from 1 to 9, with 9 indicating most rural.³⁹

Abbreviations: BMI, body mass index; CEP, Community Eligibility Provision; CHIP, Children’s Health Insurance Program.

The mean (SD) number of OCHIN patients per school per year was 44.2 (78.0). Schools were majority elementary (69%) and had a mean (SD) total number of students per year of 661.8 (501.0), with 76.9% of students eligible for FRP meals. The mean social deprivation was high and the sample was primarily urban. The geographic distribution was driven by the location of OCHIN-participating clinics: 32% of all schools in the balanced panel were located in California, 27% in Oregon, 11% in Texas, and the remaining 30% in 9 other states (Table S1).

Table 1 also shows the sample characteristics, averaged over the study years, by cohort. Across treatment status and cohort, some demographic characteristics are similar while others have more variability. Age, racial composition of children in the sample, primary language spoken, OCHIN patients per school, and district size showed some differences between groups. Unadjusted BMIz was generally increasing or stable for most periods in all cohorts (Figure 1).

Figure 1.

Unadjusted trends in mean BMI z-score by year of Community Eligibility Provision adoption. Source: Authors’ analysis of data from OCHIN, US state Departments of Education, National Center for Education Statistics, Food Research and Action Center, Robert Graham Center, and the US Department of Agriculture, 2013–2019. The figure displays trends in yearly mean BMI z-score among 149 052 patients matched to 1085 schools followed from school year 2013–2014 through 2018–2019. BMI z-score was calculated using the 2022 updated CDC extended BMI growth charts.³³ Vertical dashed lines indicate when the Community Eligibility Provision was first adopted by each cohort of schools. Abbreviations: BMI, body mass index; CDC, Centers for Disease Control and Prevention.

Figure 2 shows the DID estimates for each year in event study time. Pre-policy trends support the parallel trends and no anticipation assumptions—that is, in the pre-period, the DID estimates are each close to zero, do not indicate an upward or downward slope, and CIs indicate estimates are not significantly different from zero. The parallel trends assumption also held in the subsample with a longer period of pretreatment data (Figure S3).

Figure 2.

Treatment effect estimates of participation in the Community Eligibility Provision on BMI z-score aggregated by years since policy adoption. Source: Authors’ analysis of data from OCHIN, US state Departments of Education, National Center for Education Statistics, Food Research and Action Center, Robert Graham Center, and US Department of Agriculture, 2013–2019. Sample includes 149 052 distinct patients matched to 1085 schools followed from school year 2013–2014 through 2018–2019. BMI z-score was calculated using the 2022 updated CDC extended BMI growth charts.³³ The figure shows pre- and post-policy difference-in-differences estimates, aggregated by years since policy adoption (first year of policy adoption = 0, years prior to policy adoption are negative, and years since policy adoption are positive). Point estimates are shown as dark blue dots, and 95% CIs are shown by the light blue shaded area. Abbreviations: BMI, body mass index; CDC, Centers for Disease Control and Prevention.

Children attending schools that adopted CEP gained less in BMIz compared with children who attended schools that did not adopt the policy (Table 2, top rows; DID: −0.02; 95% CI: −0.04, −0.004). Figure 2 and the bottom rows of Table 2 display the estimates for time since policy adoption, which suggest that the impact of the policy on children's BMIz becomes increasingly negative (comparatively slower BMIz gains) over the first 5 years of the policy: for example, 0.01 (−0.02, 0.01) in the first year of participation and −0.05 (−0.08, −0.03) in the fifth year of the policy.

Table 2.

Treatment effect estimates of participation in the Community Eligibility Provision on BMI z-score aggregated by pre- or post-policy and years of policy participation.

	Estimate	95% CI
Aggregated pre- and post-policy DID estimates
Pre-policy (overall)	−0.004	−0.02, 0.01
Post-policy (overall)	−0.02**	−0.04, −0.004
Event study DID estimates aggregated by years since policy adoption
First year of participation	−0.01	−0.02, 0.01
Second year of participation	−0.01	−0.02, 0.01
Third year of participation	−0.02*	−0.04, 0.003
Fourth year of participation	−0.03**	−0.05, −0.004
Fifth year of participation	−0.05***	−0.08, −0.03

Source: Authors’ analysis of data from OCHIN, US state Departments of Education, National Center for Education Statistics, Food Research and Action Center, Robert Graham Center, and US Department of Agriculture, 2013–2019. The sample includes 149 052 distinct patients matched to 1085 schools followed from school year 2013–2014 through 2018–2019. BMI z-score was calculated using the 2022 updated CDC extended BMI growth charts.³³ The table shows overall aggregated pre- and post-policy difference-in-differences estimates and 95% CIs, and estimates aggregated by years since policy adoption.*P < .10, **P < .05, ***P < .01.

Abbreviations: BMI, body mass index; CDC, Centers for Disease Control and Prevention.

Robustness checks, secondary analyses

Schools that had at least 1 child who attended an OCHIN clinic were similarly likely to adopt CEP compared with all schools nationwide that were ever individually eligible to adopt CEP (Table S2). Community Eligibility Provision adoption was not associated with the proportion of children in a school visiting an OCHIN clinic (Figure S4). Including age-squared, age-by-sex, and age-squared-by-sex interactions resulted in a flat trend in the value of BMI at the 95th percentile of age and sex, indicating adequate control for any within-school compositional changes in age (Figure S5). Results were robust to dropping 14 schools with imputed ISP values (Table S3), controlling for using Medicaid in direct certification (Table S4), clustering at the child and school level (Table S5), and to limiting to only children in households with nonmissing income and incomes less than 185% FPL (Figure S6).

Secondary analyses of treatment effects by school-level ISP showed negative treatment effects for lower ISP schools (Figure S7). By grade, only elementary level had a large-enough sample and reasonable pretreatment trends; the sample size in middle, high, and “other” schools was substantially smaller and therefore limits interpretations by grade. Among elementary schools, results were consistent with the primary results (Figure S8). The cohort-specific effects (Figure 3) show that the 2014–2015 and 2016–2017 adopting cohorts are consistent with the primary results, while the 2015–2016 and 2017–2018 cohorts are less consistent with the overall pooled result and indicate potentially underlying heterogeneity in the treatment effects by the year of treatment adoption, with these treatment adoption years showing increases in BMIz in association with CEP.

Figure 3.

Treatment effect estimates of participation in the Community Eligibility Provision on BMI z-score by cohort and year. Source: Authors’ analysis of data from OCHIN, US state Departments of Education, National Center for Education Statistics, Food Research and Action Center, Robert Graham Center, and US Department of Agriculture, 2013–2019. The sample includes 149 052 distinct patients matched to 1085 schools followed from school year 2013–2014 through 2018–2019. BMI z-score was calculated using the 2022 updated CDC extended BMI growth charts.³³ The figure shows yearly difference-in-differences estimates by cohort (year schools first adopted the Community Eligibility Provision). Vertical dashed lines indicate when the policy was first adopted in that cohort. Asterisk (*) indicates the reference year for each cohort. Point estimates are shown as dark blue dots, and 95% CIs are shown by the light blue shaded area. Abbreviations: BMI, body mass index; CDC, Centers for Disease Control and Prevention.

Discussion

Using clinically measured height and weight data from a large sample of racially and ethnically diverse and predominantly lower-income children within 1085 schools across 12 states, we found evidence that adoption of UFM was associated with slower increases in BMIz compared with eligible, nonadopting schools. Effect estimates became more strongly negative with longer duration of CEP. Secondary analyses of cohort-initiation treatment effects suggested some evidence for potential heterogeneity.

Our primary findings that CEP participation was associated with a relative decline in child BMIz among a sample of predominantly low-income children are consistent with some, but not all, previous studies. They are consistent with a previous analysis in California that found that CEP was associated with declines in school obesity prevalence.³⁰ Similar to this previous study, our sample includes a large proportion of schools located in California, is predominantly children with lower incomes, and the largest race/ethnicity share is Hispanic/Latine children. We did not test for heterogeneity by income, but the vast majority of our sample had incomes less than 185% FPL, thus making our results also consistent with Andreyeva and Sun,²⁸ who found that CEP was associated with decreases in overweight prevalence among children with household incomes less than 185% of the FPL. Examining CEP in New York State, Rothbart and colleagues²⁷ also found evidence for CEP being associated with declines in obesity prevalence, albeit only among secondary schools, with some evidence this may be driven by secondary schools in rural areas. They found no impact on overweight or obesity for elementary schools or in more urban areas.

Our primary findings are less consistent with a study of the implementation of a UFM program in New York City, which found no association between CEP and BMI among children with incomes less than 185% FPL, but found evidence for a protective association among children with incomes greater than 185% FPL.²⁶ Our primary results are also not consistent with a longitudinal study of 1 large school district in the metro-Atlanta area, which found significant increases in BMI percentile in association with CEP adoption. The income distribution of that sample is similar to ours, but different in racial composition, with Black students comprising 74% of their sample.²⁹ Our cohort-specific analysis suggests an increasing BMI among some treatment-year cohorts; however, the key factors driving such heterogeneities are unclear.

Since our sample is nearly entirely composed of children who are lower income, our primary results and analyses among children with incomes less than 185% FPL are consistent with the idea that children who were previously eligible for FRP meals likely also benefit from UFM. Researchers have hypothesized that the lowest-income children may not experience benefits from UFM.^26,28 However, empirical evidence suggests that school meals participation increases for children both newly eligible as well as among children who were previously eligible (reviewed in Domina et al⁴⁰) and that removing the stigma (ie, that eating FRP school lunch is associated with poverty) and individual applications are additional pathways by which children who were previously eligible for school lunch may increase their participation in school lunch and experience changes in outcomes as a result. Our secondary analyses by school ISP found that results were stronger for lower-poverty schools, for which, paired with the fact that our sample is primarily low-income, we speculate that findings might be strongest among lower-income children attending lower-poverty schools.

Findings that the treatment effect estimates become more strongly negative over time may reflect multiple plausible mechanisms. Even though our sample was primarily different children measured each year, as years post-policy increase, it becomes more likely that the children who are measured in our sample have experienced UFM for multiple years and thus had more time to influence their BMI. Additionally, schools and families could learn over time in ways that affect the policy's impact on BMI. For instance, meals could become healthier as schools adjust to full implementation of nutrition standards of the HHFKA. Additionally, the culture of UFM could progressively increase the percentage of kids who participate in the program. Finally, if children initially eat double meals, once families become more widely aware of the free meal option, they may adjust behaviors outside of the classroom to be consuming only 1 meal.

While our primary models point to an average treatment effect of decreasing BMIz, we also found some evidence of heterogeneity of effects by treatment-year cohort. While allowing that heterogeneity by cohort is necessary for correctly specifying the model, this way of grouping schools does not necessarily identify salient characteristics responsible for treatment effect heterogeneity. Our primary estimates allow for heterogeneity in treatment effects at the lower level of school and we speculate that factors discussed previously in the literature reviewed above (such as grade, income, locale/region, rurality, race/ethnicity) as well as unmeasured factors such as the quality, healthfulness, and acceptability of the school meals and availability of alternatives (ie, fast-food or corner stores near school) may be driving these heterogeneities, but our study design is not conducive to examining those heterogeneities. Nevertheless, some cohort-specific estimates suggest that the CEP policy was associated with increases in BMIz. Future well-powered studies should continue to unpack the mechanisms that link UFM with child weight and, in particular, which characteristics might be associated with potentially negative unintended consequences such as increases in BMIz. Studies that incorporate school food quality, as well as qualitative and implementation studies, could shed light on mechanisms.

Limitations

We followed schools longitudinally, but measurements from children are largely cross-sectional. The sample of children is clinic-based rather than school-based; therefore, we did not capture the experience of all children in the included schools. We assigned children to their neighborhood school based on their address, age, and school enrollment policies. There is likely misclassification in this process, particularly for places that do not use residential address as the default for assigning schools or have a highly subscribed open-enrollment program. However, we have no reason to believe that this misclassification would be differential by CEP status among CEP-eligible schools. Some schools may be CEP-eligible based on district ISP but not included in our comparison sample. While this sample is not necessarily generalizable to high-income or privately insured children, within the sample we did not find evidence that school treatment status was associated with the probability of attending an OCHIN clinic. We additionally did not find evidence that schools with children attending an OCHIN clinic were associated with the likelihood of adopting CEP nationwide, suggesting that CEP was not associated with selection into the sample. We did observe some differences in child and school characteristics by treatment status and initiation date; however, event study models indicated that BMIz was evolving in parallel for both groups prior to treatment. The estimated treatment effect is diluted by a potentially large proportion of children who do not change meal source or payment as a result of the policy compared with the proportion who do change. We excluded a large proportion of schools in constructing a balanced sample of schools. While this decreased the generalizability of our sample, it increased internal validity.

Conclusion

This study builds on the small body of evidence suggesting that UFM are associated with relative decreases in childhood BMI, at least for some populations. It adds to the evidence that eliminating barriers to school meal participation may impact child health. At the same time, the salient factors that determine health impacts of UFM and any negative unintended consequences on BMI deserve further study.

Supplementary material

Supplementary material is available at Health Affairs Scholar  online.

Funding

Research reported in this manuscript was supported by the Eunice Kennedy Shriver National Institute of Child Health and Human Development of the National Institutes of Health under award number R01HD105666 and by PCORnet. PCORnet has been developed with funding from the Patient-Centered Outcomes Research Institute (PCORI) and conducted with the Accelerating Data Value Across a National Community Health Center Network (ADVANCE) Clinical Research Network (CRN). ADVANCE is led by OCHIN in partnership with Health Choice Network, Fenway Health, and Oregon Health and Science University. ADVANCE's participation in PCORnet is funded through the PCORI award RI-OCHIN-01-MC. Partial support for this research came from a Eunice Kennedy Shriver National Institute of Child Health and Human Development research infrastructure grant, P2C HD042828, to the Center for Studies in Demography and Ecology at the University of Washington (CSDE). Additional support for this research came from support to CSDE from the College of Arts and Sciences, the UW Provost, eSciences Institute, the Evans School of Public Policy and Governance, College of Built Environment, School of Public Health, the Foster School of Business, and the School of Social Work.

Data availability

Raw data underlying this article were generated from multiple health systems across the OCHIN network; restrictions apply to the availability and re-release of data under organizational agreements. Researchers interested in accessing the study data can find relevant information at https://ochin.org/research.