Waiting for the Weekend – The Adoption and Proliferation of Weekend Feeding (“BackPack”) Programs in Schools

1 Introduction

Backpack or weekend feeding programs, which support food insecure schoolchildren by providing them packs of food for the weekend, have grown from a single school in Arkansas in 1995 to a nationwide patchwork of programs. Feeding America (FA) reports serving more than 450,000 U.S. children annually through their “BackPack” (BP) programs (Morello 2019);^[1] the total number served through all programs is unknown but was estimated to exceed 800,000 in 2015 (Fram and Frongillo 2018). The limited research in this area suggests that weekend feeding programs, like government-provided, school-based programs, improve student food security and academic outcomes (Burke et al. 2021; Ecker and Sifers 2013; Fiese et al. 2020; Kurtz, Conway, and Mohr 2020; Laquatra, Vick, and Poole 2019; Ryan et al. 2023; Shanks and Harden 2016; for a survey see Kurtz, Conway, and Mohr 2022). However, in contrast to school-based feeding programs that are administered and funded by the government, backpack programs are administered solely through private charities and their partners. This feature raises the question of which needy children and schools end up being served by these programs and which do not – and the reasons for and implications of those differences.

This work provides a first look at the initiation and spread of backpack programs and undertakes an in-depth investigation into the factors influencing this spread. In addition to need, community/school resources and proximity, our data permits an unusually rich exploration into the role of race/ethnicity and religion as well. It also provides the first evidence, to our knowledge, of the current state of these programs and how they have evolved since FA began offering them in 2006. That BP programs have been linked to improvements in both food security and academic outcomes among students underscores the need to understand and acknowledge their presence. Not only may these programs be a critical unobserved variable in empirical studies of other educational policy interventions, but their decentralized, private provision raises questions of racial/ethnic and socioeconomic equity.

To tackle these questions regarding a program for which little data exists and thus little is known, we focus on the rapid spread of Feeding America’s BackPack program in a 17-county region of northwest North Carolina served by one FA affiliated foodbank. The North Carolina program that we study requires each school to have sponsorship from a local community partner. Therefore, program adoption is dependent on charitable giving by the surrounding community and is not determined solely by the local foodbank or the school itself. We observe the characteristics of the students, schools, and surrounding communities, which we then match to program-level data on the pattern and extent of program adoption during a 6-year period when this area transitioned from having no programs to more than one-quarter of ‘eligible’ schools having one. Besides the largely donor-driven approach of the local food bank, our data also covers a time period before other charitable organizations began taking on these programs, which our update at the end of this paper finds are now widespread.

The nature of our data and our research question preclude a sharp identification strategy and the associated argument for estimated causal effects. While we try to rule out alternative explanations and possible biases via robustness checks and falsification tests, the variation is primarily cross-sectional and there is no treatment versus control group to exploit. With this caveat in mind, our empirical results suggest that student need and school location are crucial factors to program adoption. Our investigation also contributes new evidence and insights into two factors commonly explored in charitable behavior research – racial/ethnic diversity and religion.

Unlike most research, we distinguish the racial/ethnic composition of the donor (community) from the recipient (school); moreover, the racial/ethnic diversity of the study area in North Carolina provides substantial variation in and between these measures. We adapt Vigdor’s (2004) theoretical framework of the effects of race/ethnicity on voluntary contributions to a public good to the case of charitable giving when the racial/ethnic characteristics of donors and recipients differ. This adapted model produces a ‘hybrid’ fragmentation index (HFI) that takes account of the differences between the donor and recipient. The growing racial/ethnic diversity of the US population skews toward the young (Frey 2019), which means that the racial composition of schools is increasingly different from that of the surrounding communities and suggests that distinguishing between the two matters.

Exploring the role of religion when the donor is a community and the recipient is a school requires a more granular measure of places of worship (PW) than the county-level or even more aggregated measures typically used (Finke and Bader 2011). To our knowledge, ours is the first economic study to use PW GIS data from Environmental Systems Research Institute (ESRI), which provides the exact location (latitude and longitude) of every PW. Such refined data allow us to identify places of worship in close proximity to a school, a measure not observable in more commonly used county-level data. This proximity measure reveals substantial variation within counties.

The remainder of the paper proceeds as follows. The next section provides background information on past research as well as the geographic study area and the mechanics of the BP program. Section 3 provides a descriptive analysis of the program’s spread, comparing the observed spread with one that prioritizes need, and exploring factors that explain the divergence. Having found that need – alone – does not explain the spread the program, Section 4 develops a conceptual framework and empirical strategy with which to investigate other factors associated with the program’s spread, with an emphasis on the role of race/ethnicity. Section 5 reports results of a discrete-time survival analyses and a wide range of robustness and falsification checks and alternative specifications. Section 6 addresses the question of what has happened to the remaining unserved schools since the end of our sample period, providing an update on the current state of backpack programs in the area. The complicated answer to that question as well as the main takeaways from our empirical analyses coalesce and suggest directions for future research in the concluding section.

2 Background

We first place this research in a broader context, highlighting connections to past studies of both child nutrition programs and charitable giving. We then provide a discussion of the special features of the BP program and our area of study, all of which informs our conceptual and empirical approaches.

2.2 BackPack Programs in Northwestern North Carolina

Our study area consists of 17 Northwest North Carolina counties that are served by a single foodbank, the Second Harvest Food Bank of Northwest North Carolina (SHFB).^[2] This region, depicted in Figure 1, includes agricultural piedmont communities, Appalachian counties, and the urban centers of Greensboro, Winston-Salem, and High Point. The diversity of the region, along with the inclusion of urban areas, allows us to explore different factors, including differences in race/ethnicity, religiosity, income, and population density that might affect program adoption. Median household incomes, as well as measures of racial/ethnic diversity, in these North Carolina counties are typical of many US counties (see Kurtz, Conway, and Mohr 2020).

Figure 1:

17-County region of NW North Carolina served by SHFB, with school location and rate of adoption.

To bring the BP program to a school during the time frame of our study, SHFB required that the school meet an eligibility requirement that at least 50 % of students be economically disadvantaged (ED).^[3] However, schools observed below this cut-off were sometimes served as well. This fact suggests that need was not the only driving factor in adopting these programs and leads us to explore alternative ways of defining possibly eligible schools in our analyses.

The donor-driven nature of SHFB’s BP program helps explain this pattern of adoption and makes it especially suitable for studying charitable behavior.^[4] To participate, the school needed a community partner organization to finance the cost of the program. The community partner would commit to sponsoring at least 50 students and assume the responsibility for packing, storing, and delivering food packs to the school. The food was provided by SHFB. School employees assisted with the distribution of packs to students. During our sample period, the cost of food was about $5 per pack, so the minimum financial commitment for a community partner was $10,000 to provide 50 packs for the 40 weeks of a school year. While we do not know the identities of community partners during our study period, information about current (2021) partners, as well as conversations with SHFB personnel, indicate that community partners were predominantly churches.

As is evident from Figure 1, SHFB BP programs expanded rapidly during our study period. Participation increased from nine programs in our analysis sample that initiated in the 2008–09 school year (which for simplicity we refer to as 2009) to 80 schools with a program by 2014 (2013–14 school year). Towards the end of the sample period, weekend feeding “backpack” programs provided by other organizations began to emerge and/or evolve from an existing BP program. Two notable programs that grew towards the end of the sample were the Out of the Garden and Samaritan Kitchen of Wilkes County, serving exclusively Guilford and Wilkes counties, respectively. These programs differed in many ways, including eligibility and size requirements and the food provided. Program initiation information is less complete, making them challenging to include in our analyses. In Section 5, we explore the robustness of our results to the existence of these programs and find they make little difference to our findings.

The emergence and subsequent growth of other programs (see Section 6) is one of several reasons for ending our empirical analyses in 2014. Other reasons include changes to SHFB’s eligibility requirements and changes in the coding of the school-level data used, both of which occurred around 2014. The federal Community Eligibility Provision of the Healthy, Hunger-Free Kids Act of 2010 permitted the adoption of universal free lunch programs at many analysis sample schools beginning in 2014 and otherwise led to substantial changes in nutritional programs serving school-aged kids (see Kurtz, Conway, and Mohr 2022). Most importantly, as Section 6 also highlights, the vast majority of SHFB BackPack programs were initiated by 2014. After 2014, the foodbank instead started emphasizing in-school and mobile food pantries. The 2009–14 period (i.e. school years 2008–09 to 2013–14) permits a tight focus on understanding how these programs began and accelerated, while our update in Section 6 underscores their subsequent growth and thus the need for future work.

3 Data and Descriptive Analyses of BackPack Program Adoption

Our main analyses link primary data about BP program participation and programmatic initiation years collected by SHFB to school-level administrative data maintained by the North Carolina Education Research Data Center (NCERDC), community characteristics from the US 2010 Census, and precise locations of PWs from ESRI’s 2013 US Institutions data. SHFB staff confirmed for us that they were the first weekend feeding program in the area; the program staff initially targeted primary schools because there were fewer concerns about stigma associated with food packs, and during this time the program administrators tried to adhere closely to the eligibility criteria (a school’s program must serve at least 50 students and at least 50 % of students at a participating school are ED). These data sources allow us to identify exactly when a school initiated its BP program, distinguish between the racial/ethnic make-up of the school and the community, and permit us to explore the impact of a granular measure of religiosity. The Data Appendix provides details on these sources.

Our main analyses are limited to a balanced panel of schools that were likely to have been under consideration for a BP program during the initial years of the program. These are schools with complete data for 2009–2014 that don’t change location, don’t serve any grades higher than 8th grade, and where the proportion of ED students exceeded 0.40 for 2008, the year before the program’s existence (2009+).^[5] The analysis sample consists of 274 schools, 80 of which (29 %) had adopted a BP program by 2014. This sample comprises 54 % of all public schools in the 17-county sample area and includes all but 6 schools that participated in a BP program during our study period.^[6] Table 1 provides the number of schools observed in each of the 17 counties and reports the number of schools that had a BP program by year.

Table 2 provides summary statistics for the covariates used in empirical analyses, further broken down by whether the school adopted a program by 2014. At the top of the table is information on the school’s need – the proportion of ED students; a nontrivial proportion of schools with a student body below 0.50 ED in 2008 eventually adopted the program. The rest of the table is organized into the different groups of factors beyond need that may be related to adoption. BP adopting schools have a larger number of ED students (ED#), indicating possible economies of scale in providing the program, are more likely elementary schools, and tend to be in denser, poorer communities that are geographically closer to SHFB than schools that had not adopted by 2014.

Table 2 reveals that adopting schools are less predominantly white and have higher concentrations of all other racial/ethnic groups than those that have not adopted by 2014. The bottom of the table reports our measure of religiosity. With its reliance on county-level measures, past work provides little guidance in defining religiosity with more geographically precise data. Our primary measure is a count of PWs within a 3 km (10 km) radius of schools located in urban (rural) census tracts. In robustness checks we draw larger (5 km and 20 km) or smaller (1 km and 5 km) circles and explore county-level measures. Adopting schools have a larger count of PWs.

These simple comparisons ignore the strong correlation between need, as proxied by the proportion or percentage of ED students (ED%), and other attributes. Moreover, the richness of our data permits us to compare the observed program adoption pattern with a counterfactual scenario that prioritizes the highest ED schools – i.e. what a social planner giving preference to high-need schools would presumably choose. This counterfactual scenario specifies the same rate of adoption (e.g. 9 schools adopt in 2009 and 8 more schools adopt in 2010) but orders the schools so that those with the largest proportion of ED students initiate the program first; i.e. the 9 schools with the highest proportion ED are assigned to adopt in 2009, and the next 8 schools in terms of need are assigned to adopt in 2010. Comparing this need-based ordering with the actual adoption year identifies which schools adopted earlier than as specified by their ED%.

This exercise reveals that the observed adoption process deviates substantially from one driven solely by need. At the start of the program, 21 schools were made up entirely of ED students (ED% = 1) and thus would have been assigned to receive the program first. Yet, none adopted in the first year (2009) and only two did in 2010; five still had not received it by the end of the study period 2014. In contrast, the schools that actually adopted in 2009 had an average ED% of 0.64, which is below the overall sample average of 0.66. The average ED% of newly adopting schools rose over time but it was only in the last year that it was close to the 0.90 average ED% of the schools who would have received the program by 2014 under a prioritization based on need alone.

To explore the factors associated with an ‘early’ adoption, we create a dummy variable equal to 1.0 for any school that adopted the program earlier than its ED% would have dictated. We then estimate via probit how this ‘early’ adoption variable is associated with the school and community characteristics from Table 2. While we include the number of ED students in the school to account for (dis)economies of scale considerations in offering the program, we do not include ED% as an explanatory variable, since it is used to construct the measure of early adoption. The right side of Table 2 reports the estimated marginal effects of these variables. Other than proximity to SHFB, the factor associated with ‘early’ adoption is the racial composition of the school (nonwhite schools are less likely to adopt early); the association with PWs is positive but small and statistically insignificant. This simple analysis suggests that factors besides need helped drive the adoption of this donor-driven program and that race/ethnicity is a likely candidate.

4 An Empirical Model of BP Program Adoption

Given SHFB’s process, the decision to bring a program to a school is driven primarily by donors’ preferences. To model this decision, we generalize a model of voluntary contributions that focuses on racial/ethnic fragmentation. We adapt the model to the specifics of our data by devising an empirical strategy and estimating equation.

4.1 Conceptual Framework

Our conceptual framework adapts and generalizes Vigdor’s (2004) model of how the racial/ethnic characteristics of a community influence the decision to contribute to a public good. Vigdor (2004), like Vigdor (2002), assumes that people exhibit within-group affinity for beneficiaries of the same race/ethnicity. His model aggregates from the individual to characterize the collective contribution to a public good: in his case the response rates to the Census. We don’t observe individual donors and must also aggregate to the community level to describe the behavior of a representative donor. A key difference between BP programs and the case studied by Vigdor is that we model charitable giving in which the racial/ethnic characteristics of the donor (community) differ from the racial/ethnic characteristics of recipients (schools). Note that these differences could occur either due to demographic trends – “that the nation’s diversity is percolating from the ‘bottom up’ as the white population ages” (Frey 2019) – or due to residential sorting which could lead to segregation (e.g. Caetano and Macartney 2021; Macartney and Singleton 2018; Monarrez 2023). The latter may be more feasible in communities with multiple schools, a possibility we explore in our empirical analysis.

Consider the decision of a representative donor, i, who belongs to racial group r, resides in community c, and is contemplating sponsorship of a BP program at a local school, s. Let σ _rs denote the racial share of group r in school s. The probability the donor supports sponsorship of a BP program is represented by a linear function:

(1) Pr Y ircs = 1 = β 1 S s + β 2 C c + γ X irc + θ r + δ r σ r s

Here, Y _ircs  = 1 indicates that donor i supports a sponsorship decision; S _s , C _c , and X _irc respectively denote school, community, and donor characteristics that affect a donor’s willingness to support sponsorship. The costs of providing the program are not modeled directly but are captured by the variable vectors S _s and C _c to the extent they are affected by school and community characteristics. The parameter θ _r allows baseline giving to vary by racial/ethnic group, and δ _r  > 0 indicates a preference for supporting a school with pupils of the same race/ethnicity as the donor. Equation (1) is similar to the first equation in Vigdor (2004). However, the inclusion of the school-specific measures, S _s and σ _rs , extends Vigdor’s model to distinguish the characteristics of the donor from the characteristics of the recipient.

Like Vigdor (2004), we derive the sponsorship decision of an aggregated representative donor, whose decision depends on the average probability of support for program sponsorship from all donors in the community. This average probability is derived by taking the weighted sum of each group’s probability of support using σ _rc , the racial/ethnic shares within the community. Following Vigdor in assuming that δ _r  = δ is the same for all racial/ethnic groups, this weighted sum can be expressed as:

(2) Pr ̄ Y ircs = 1 c s = β 1 S s + β 2 C c + γ X ̄ c + ∑ θ r σ r c + δ ∑ σ r c σ r s

Equation (2) describes the probability that a community-school pair receives a BP program, expressed in a way that can be estimated empirically. Coefficient estimates of the vectors β, γ, and θ capture the effects of school and community characteristics, and community racial/ethnic shares.^[7] The right side of Equation (2) is similar to the equivalent expression in Vigdor (2004, Eq. (3)), but now generalized to account for both community and recipient characteristics. The most novel element of Equation (2) is the last term, which suggests modifying a common measure of racial/ethnic diversity, the fragmentation index, to capture the disparity between donor and recipient.

The fragmentation index, 1 − ∑ σ r 2 , is used by “virtually every empirical contribution” on ethnic or cultural diversity (Bossert, D’ambrosio, and La Ferrara 2011, p. 723). In the special case where donor and recipient population shares are identical σ r c = σ r s , the last term in Equation (2) reduces to an additive inverse of this fragmentation index. This special case holds when donors and recipients are the same, as in the case of contributions to a local public good, but it does not hold generally for charitable giving.^[8] Our adaptation of Vigdor’s model employs a new fragmentation index, F = 1 − ∑σ _rc σ _rs , which we refer to as a hybrid fragmentation index (HFI). Instead of capturing the diversity within a given population, the HFI is largest when there are differences across populations.

Like the standard racial/ethnic fragmentation index, the HFI ranges from 0 to 1. In contrast to the standard index, which equals zero if there is only one group and approaches 1 as the number of equal-sized racial/ethnic shares becomes very large, the HFI equals 0 if the entire community and school are of the same racial/ethnic group. The HFI equals 1 if all pupils at the school belong to one racial/ethnic group while the entire community belongs to other groups. The HFI decreases as community members shift from a group that is less prevalent in the school to one that is more prevalent. The same logic applies to shifts in the school’s population; if a school’s largest group is Hispanic, the HFI decreases as the Hispanic share in the community rises.

When the racial/ethnic composition of the donor and recipient are identical, all 3 fragmentation indices (donor, recipient, hybrid) are equal. The measures diverge as the compositions differ, as they do in our sample; the correlation between the school fragmentation index and the census tract fragmentation index is only 0.10. Figure 2 plots the proportion of white residents in the census tract against the proportion of white students in each school and includes a 45° line – the line of equality between the two – for reference. This plot reveals a positive but far from exact relationship. Census tracts tend to have a higher proportion white than the schools that reside in them. These differences seem largely driven by age, which is consistent with national demographic trends (Frey 2019). Comparing the racial composition of adult (over age 20) and child (age 14 and under) populations within census tracts confirms this tendency in our sample. The proportion white is higher for adults (70 %) than for children (57 %).

Figure 2:

Relationship between racial composition (%white) in the community and the school. Note: The proportion white from the community (census tract) is from the 5-year average 2013 ACS and the proportion white from the school is the sample average value (2009–2014) from the NCERDC school-level data.

5 Empirical Results from Discrete Time Survival Analyses

Table 3 reports the results of estimating Equation (3) for our main sample and several variants. To ease interpretation, we report the estimated marginal effects for each variable and p-values for the joint statistical significance of each variable group – school (S), community (C), location (L) and race/ethnicity. The first column reports the results for our main sample of 274 schools. As expected, our measures of need are highly statistically significant; however, the reported marginal effects absorb the effect of each ED variable shown in Equation (3) and obscure the predicted effect of proportion ED over the full range. We therefore plot the predicted effects from this model as proportion ED increases from 0.40 to 1.0 in Figure 3. All three coefficients are strongly, individually statistically significant and combine to predict that the probability of adopting increases with ED% much faster as the school nears the 50 % threshold. After reaching the threshold, the effects of ED% are much more modest and the probability of adopting actually drops when ED% exceeds 50 %. Observably less needy schools that approach eligibility are more likely to get the program than schools with moderately larger needs. In preliminary analyses, we investigated alternative specifications for need (e.g. excluding one or both terms beyond ED in Equation (3) or specifying a quadratic function in ED); results are largely robust. In all cases, ED is strongly statistically significant and positive and estimates consistently imply a change around 0.50.

Figure 3:

The predicted probability of BP adoption by proportion of ED students. Note: Predicted probabilities are constructed from the parameter estimates of Equation (3), using the specification reported in Table 3, column 1. The estimated coefficients on ED%, ED50 and their interaction are each statistically significantly different from zero.

Surprisingly, neither the individual effects nor the joint effect of the school’s (recipient) characteristics are statistically significant. The community (donor) characteristics matter little as well, except for the county unemployment rate, suggesting that communities especially hard hit by the Great Recession adopted these programs sooner. Location is strongly related to adoption, with schools that are closer to SHFB being more likely to adopt. The key variables of interest, race/ethnicity and religiosity, have the expected signs but do not rise to conventional levels of statistical significance.

Before investigating these results more thoroughly, especially with respect to race/ethnicity and religion, we first explore whether results are sensitive to how our sample is constructed. The next two columns of Table 3 report results when we alternatively narrow and then broaden our definition of ‘eligible’ schools, which in our main sample is defined as exceeding 40 % ED in 2008. Column 2 limits the sample to those schools with an ED% that exceeds 50 % in the previous school year, which leads to an unbalanced panel. Column 3 expands the sample to include schools that had an ED% over 40 % at any time, which leads to 59 more schools being included. Neither variation substantively affects the results, although expanding the sample to include questionably eligible schools (column 3) tends to yield estimated effects of similar signs but less statistical significance, consistent with including a noisier sample. An exception is the proximity, whereby proximity to another BP school replaces proximity to SHFB as statistically significant.

To address the possibly confounding impact of the alternative backpack programs that emerged towards the end of the sample, we drop schools that we know adopted a different backpack program. Results are quite similar (column 4).

Lastly, we revisit the concern that our conceptual model best fits the situation of a single donor and single recipient – i.e. a census tract with only one potentially eligible school. In tracts with more than one school, the donor’s choice is more complicated as they can now decide which school(s) to support. The last two columns in Table 3 report results when the sample is split into schools that are the only eligible school in the census tract (“lone”) versus those where there is more than one (“multi”). This distinction reveals stronger differences, especially for race/ethnicity. The HFI is now strongly negative and statistically significant for lone schools, suggesting that a school with a racial/ethnic composition different from the community is less likely to receive the program. Lone schools in communities with a larger proportion of non-Hispanic black residents are also less likely to receive a BackPack program. In contrast, the race/ethnicity and HFI variables are neither individually nor jointly statistically significant for multi schools.

One explanation for this difference between lone and multi is that our conceptual framework and/or data may not fully capture the multi-school situation. Recall that our conceptual framework and associated race/ethnicity measures assume one donor and one recipient. Choosing among multiple recipients suggests a more complex behavioral setting. Data may also play a role. In a census tract with multiple schools, the community may sort into smaller sub-communities associated with each school. However, such residential sorting suggests larger differences between each school’s racial/ethnic composition and that of the broader community. Instead, the observable characteristics are quite similar across multi and lone schools, including the hybrid HFI (see Appendix Table A1). This similarity lends support to a behavioral, rather than data, difference between these two settings.

For multi schools, what matters is being an elementary school and close proximity to the SFHB. There is suggestive evidence that having other BP schools in the census tract and proximity to PWs are likewise positively associated with adoption, but they are not statistically significant. These results suggest that when faced with several potential recipients, proximity matters and that donors prefer serving younger children.

5.2 Investigating the Role of Race/Ethnicity

Our main specification is motivated by the conceptual framework and its inclusion of the racial/ethnic attributes of the donor (community) and the HFI. Our data, with characteristics for both the donor and recipient, allows us to explore several variations of the model; for example, using only the community (donor) characteristics and fragmentation index, as most past research has done.

First, however, it is fair to ask how much variation in race/ethnicity is present in the data and whether it differs between lone and multi schools. Figure 2 has already confirmed substantial differences between school and community compositions in terms of the proportion white; the evidence is similar when split into lone versus multi schools. Figure 4 disaggregates race beyond white and nonwhite by displaying the degree of variation across the different racial/ethnic groups for lone versus multi schools.^[11] The figures display strong independent variation for both types of schools in that (1) it is not a simple breakdown of white and black, as there are nontrivial proportions of Hispanics and other groups, and (2) the relative size of each nonwhite group varies in a nonsystematic way as the percent white increases. These plots suggest that our data have sufficient variation to credibly identify the effect of race/ethnicity and that there are not strong differences in the racial composition of lone and multi schools. The lack of statistically significance for multi schools is therefore unlikely to be due to a lack of variation.

Figure 4:

Racial/ethnic composition of schools, ordered by percent white. Note: Figures are constructed using the sample average values of NCERDC school-level data (2009–14). Schools are ordered from lowest %white to highest %white.

We explore the robustness of the main model’s results with respect to race/ethnicity by estimating several logical variations, summarized in Table 4. For the sake of brevity, we report detailed regression results for the key variables only and summarize the rest via joint tests. For each subsample, the first column repeats the results from the main model (Table 3, last two columns) for comparison. The next two columns follow the more standard approach when only the race/ethnicity of the donor (or, alternatively, the recipient) is observed; i.e. the racial/ethnic composition and fragmentation index refer to the same entity, the census tract or, conversely, the school. The last column reports the broadest model which includes the racial/ethnic composition of both the census tract and the school, alongside the hybrid fragmentation index that captures the differences between the two.

Table 4:

Discrete survival, estimated marginal effects – different measures of race/ethnicity.

Panel A: lone school in tract		(1)	(2)	(3)	(4)
		Main (from Table 3)	Community frag index	School shares and school frag index	Add school shares to main model
Community	Prop. black non-Hispanic	−0.141*	−0.089		−0.103
		(0.083)	(0.071)		(0.071)
	Prop. Hispanic only	0.057	0.186		0.085
		(0.153)	(0.205)		(0.126)
	Prop. other race	0.206	0.188		0.492
		(0.236)	(0.294)		(0.504)
School	Prop. black non-Hispanic			−0.185*	−0.092
				(0.096)	(0.106)
	Prop. Hispanic only			−0.202	−0.171
				(0.175)	(0.196)
	Prop. other race			−0.143	−0.517
				(0.252)	(0.479)
Fragmentation	Fragmentation index	−0.221***	−0.248**	−0.089	−0.135
		(0.064)	(0.109)	(0.106)	(0.085)
		Joint tests of statistical significance (p-values reported)
	School vars	0.31	0.47	0.20	0.19
	Community vars	0.29	0.10	0.56	0.37
	Location vars	0.00	0.00	0.00	0.00
	Race/ethnicity vars	0.00	0.00	0.14	0.00
Panel B: multiple schools in tract		(1)	(2)	(3)	(4)
		Main	Community	School shares	Add school
		(from	frag	and school	shares to
		Table 3)	index	frag index	main model
Community	Prop. black non-Hispanic	−0.098	−0.089		−0.189
		(0.110)	(0.116)		(0.172)
	Prop. Hispanic only	−0.301	−0.241		−0.445
		(0.238)	(0.268)		(0.347)
	Prop. other race	−0.461	−0.246		−0.594
		(0.303)	(0.346)		(0.434)
School	Prop. black non-Hispanic			0.025	0.167
				(0.126)	(0.211)
	Prop. Hispanic only			0.138	0.268
				(0.209)	(0.333)
	Prop. other race			0.508	0.568
				(0.336)	(0.549)
Fragmentation	Fragmentation index	0.137	0.058	−0.115	0.004
		(0.130)	(0.118)	(0.092)	(0.120)
		Joint tests of statistical significance (p-values reported)
	School vars	0.00	0.00	0.00	0.00
	Community vars	0.01	0.00	0.00	0.00
	Location vars	0.00	0.00	0.00	0.00
	Race/ethnicity vars	0.54	0.94	0.29	0.74

1. Notes: All models include year dummy variables as well as all the variables listed in column 1 of Table 3 (unless noted). Robust standard errors, clustered at county-level, in parentheses. ^*** p < 0.01, ^** p < 0.05, ^* p < 0.1.

These exercises yield similar results to the main model in that racial/ethnic composition is associated with BP adoption for lone schools but not multi schools. For lone schools, however, a few subtle differences emerge that highlight the importance of including the race/ethnicity of both the donor and the recipient. In particular, using a fragmentation index that captures the difference between the donor and recipient – the HFI we propose in Section 4 – leads to a more statistically significant impact of fragmentation and race/ethnicity overall. If only community race/ethnicity is included (column 2), fragmentation still matters but not the community’s shares. Conclusions are even more different if we had instead relied only on the recipient (school) characteristics as a proxy (column 3); race/ethnicity is no longer jointly statistically significant and it is the school shares, not fragmentation, that seem to matter. Finally, including the race/ethnicity shares of both the community and school (column 4) diminishes the individual effects of our main model but the main story remains.

6 The Current State of Backpack Programs in Northwest NC

At the end of our sample (2014), almost three-quarters of the sample schools still did not have a SHFB BP program, and other charitable organizations were beginning to offer them. These two facts raise the question of what the weekend feeding program landscape in these 17 counties looks like today and whether the same factors may have helped to shape it. In this section, we make a modest attempt to address both questions.

Despite the apparent ubiquity of backpack programs and the presence of several academic and charitable North Carolina organizations that are dedicated to studying and alleviating child hunger, no centralized source of data exists for weekend feeding programs. While FA maintains records on its programs, there are now many other sources of weekend food assistance that include programs initiated by the schools themselves, partnerships between individual churches and schools, and independent organizations developing their own versions of backpack programs. In fact, the growth in SHFB’s BP program has dramatically slowed since the time of our analyses.

Our tight focus on 17 counties and 274 schools made it feasible for us to identify the current status of the remaining non-backpack schools, as described in greater detail in the Data Appendix. Briefly, we began with SHFB and other charitable organizations and academic study groups targeting food insecurity to build a comprehensive list of schools known to have had initiated a backpack program by 2022.^[13] For those schools not appearing on the list, we contacted the schools directly until we received a response.

Of the 194 schools that had never had a SHFB BP program by the last year of our sample (2014), 10 subsequently received one from SHFB, three later closed or were reconfigured, and another 166 received a backpack program from some other organization.^[14] This evidence therefore suggests a continuation of two patterns emerging at the end of the sample: (1) SHFB programs were slowing and so our sample period captures most of the BP programs initiated, and (2) other organizations were taking them on. Nearly all schools received a backpack program at some point. These other backpack programs vary substantially in terms of the number students served and food bag contents. Of the remaining 15 schools without a program, 7 report having a food pantry on school grounds.

Figure 5 highlights the locations of the 15 schools that are still without a backpack program, further split into if they have a pantry or not, contrasted with all other schools in our sample. These 15 schools are not geographically isolated. Nearly all are located near another school with a backpack (or BP) program and many are reasonably close to SHFB. As location does not appear to explain this pattern, we compare the other key characteristics of these schools versus the other types.

Figure 5:

Updated backpack program status and location of schools in 2022.

Table 5 summarizes the observation counts and key characteristics of need, race/ethnicity and religiosity from the last year of our sample (2014). For completeness, we report differences between ‘lone’ and ‘multi’ schools, as well as 2022 values for school race/ethnicity and need (publicly available through the State of North Carolina School Report Card Data). The patterns between 2014 and 2022 are similar, although schools have a lower ED% and are less white in 2022 (providing further evidence of the growing racial/ethnic diversity of the young). The high prevalence of backpack programs means that the number of observations in some columns of Table 5 is quite small and comparisons are an illustrative, exploratory exercise only. With that caveat, several intriguing patterns emerge.

Schools that received the program during our sample period (by 2014, column 4) had a higher ED% than other types of schools, consistent with SHFB’s guidelines. The rest of the schools, regardless of program status, are similar, suggesting that ED% plays less of a role after 2014. The racial/ethnic compositions of schools without a program continue to differ in a way that is consistent with our empirical analyses. Using either 2014 and 2022 measures, the lone schools without a program (whether or not they have a pantry) have larger nonwhite populations than any of the other types of schools. For example, the 3 lone schools without a program or pantry (column 1) are 27 % white in 2022, compared to 49 % white in the 142 lone schools with programs (column 3). In contrast, little difference exists among multi schools; the 5 multi schools without a program or pantry (column 1) are 49 % white compared to 46 % in the 114 multi schools with programs (column 3). The HFI follows the same pattern; schools, especially lone schools, without programs are more highly fragmented between school and community than schools with a program. Finally, the remaining multi schools without programs have fewer PWs nearby than those with programs, echoing the weakly suggestive findings from our empirical analyses.

7 Concluding Remarks

To our knowledge, this is the first empirical study to investigate the factors that affect how weekend feeding programs for school children – facilitated and funded by nongovernment entities – emerge and spread. It also is the first to provide a detailed update on the prevalence of backpack (weekend feeding) programs. Our area of study, 17 counties in northwest North Carolina, spans urban centers and rural areas, is racially and ethnically diverse, and has a high density of places of worship (PWs), making it uniquely well-suited to investigate commonly-studied factors of charitable giving – racial/ethnic diversity and religion. The richness of our data on school and community race and ethnicity leads us to revisit Vigdor’s (2002, 2004 interpretation of the fragmentation index and create a ‘hybrid’ index that measures differences between the donor and recipient. We also locate and use a more granular measure of PWs for measuring religiosity. Both innovations make a difference.

This work provides a view into the birth and rapid spread of a BackPack (BP) program initiated by a single foodbank with a clear, donor-driven policy. A descriptive analysis shows stark racial/ethnic differences in the schools that received the program early or late, relative to the school’s level of need. Our primary empirical investigation estimates a discrete time survival model. As expected, the results suggest that schools with greater student need and a close proximity to the foodbank or to other schools with BP programs were most likely to adopt earlier. However, race/ethnicity and, perhaps, religiosity also matters. Communities whose racial/ethnic composition differ the most from their school are less likely to provide such programs – i.e. the HFI has a negative effect. This is true in communities with a single high need school, a situation closest to our conceptual framework. Religiosity is a prominent factor in charitable giving research, and we are aware that many actual BP sponsors were religious organizations, yet our findings are weakly suggestive at best and only when there are multiple schools in the community. While our more granular measure – the number of PWs within a certain distance of the school – differs substantially from the county-level aggregates typically used, it is not the reason for the lack of a strong effect. Using county-level measures leads to even more mixed and, at most, suggestive findings.

About three-quarters of high need schools still did not have a SHFB BP program at the end of our formal analyses (2014). Our exploratory investigation into what has happened since yields key insights that deserve future investigation on a wider geographic scale. First, the vast majority of schools in our study area have now experienced having a backpack program; however, the first and biggest provider (Feeding America) is little responsible for this expansion as it initiated few programs after 2014. Instead, a disparate array of organizations, with differing criteria and program features, have taken over. These new programs are a critical, yet unobserved feature of the nutritional programs provided by schools. Our work reveals the challenge of obtaining comprehensive information about them. This exploration also suggests that the factors affecting provision have evolved such that need and proximity (i.e. to foodbanks, and to other schools with programs) may no longer play a role, while racial/ethnic diversity within and across communities and schools may continue to impede adoption. These patterns are based on a small number of non-backpack schools; they, therefore, are suggestive at best.

More generally, this study has at least two limitations that inform the conclusions we draw and suggest directions for future research. First, the empirical variation in our data is primarily cross-sectional and there is no corresponding control group, which precludes identifying causal effects. We address this concern with our many robustness checks and falsification tests, but nonetheless must temper our conclusions accordingly.

Second, while we believe that the diversity of northwestern North Carolina makes our findings likely applicable to other settings, the generalizability of our results is open to question. We address the concern both with our exploratory update (a new time period) and by comparing our findings to the conclusions of the only other study of backpack program adoption to our knowledge. Shanks and Harden (2016) also study the initiation and rapid expansion of Feeding America BackPack programs during the same approximate time period (2008–2012), but their focus is on Montana and they use an entirely different methodology (structured interviews). Similar to our study, about one-quarter of eligible schools adopted the program by 2012 and schools close to a food bank had higher rates of participation. Shanks and Harden (2016) also note the complexities of including programs that operate independently of Feeding America. They conclude that school eligibility (also defined as >50 % ED), school resources (e.g. a point person or advocate at the school), and, above all, the availability of financial resources from the community determine whether a school adopts and sustains a BackPack program. However, the study does not address the roles of religion or racial/ethnic diversity, so we are unable to draw comparisons there.

Recent research findings that backpack programs successfully reduce food insecurity and improve other child outcomes underscore the need to understand which children and schools receive weekend food assistance and why. Our study suggests that race/ethnicity may play a role and provides a new measure that captures differences between the donor and the recipient, both of which warrant additional investigation.