SNAP Participation and Emergency Department Use

Among those eligible to participate, households with greater levels of food hardship disproportionately self-select into SNAP.^4,5  To attend to this bias, we estimated a bivariate probit model with instrumental variables.^5,9  To simultaneously estimate the hypothesized mediation and moderation relationships, we embedded this model within a structural equation modeling framework. We detail our data, measures, and approach in the sections below.

We analyzed data from the 2016 to 2019 iterations of the National Survey of Children’s Health (NSCH), a nationally representative sample of noninstitutionalized children in the United States aged 0 to 17.⁴²  These were the first iterations to include a food hardship measure. Children with SHCN and children aged 0 to 5 are oversampled in the survey, and weights are provided to allow for nationally representative estimates.⁴²  After previous studies assessing the associations between SNAP and food hardship, we limited our analytic sample to children living in households with income <150% of the federal poverty level to focus on families most likely to be eligible for SNAP benefits.^5,9  This resulted in a final unweighted sample of 24 990 children, 6644 with and 18 346 without SHCN.

The exposure of interest was a bivariate indicator of whether a household had any SNAP participation in the previous year. The mediators were household food hardship and parent-reported child health status. Exposure to household food hardship over the previous year was measured through a single-item food insufficiency question (“Which of these statements best describes your household’s ability to afford the food you need during the past 12 months?”) with 4 possible responses: (1) “We could always afford to eat good nutritious meals,” (2) “We could always afford enough to eat but not always the kind of food we should eat,” (3) “Sometimes we could not afford enough to eat”; or (4) “Often we could not afford enough to eat.” Household food hardship was flagged as present if any of the latter 3 responses were chosen, given the associations between these responses and worse health and health care outcomes among children.²⁷  This definition incorporates marginal food sufficiency and food insufficiency,³  and it is roughly equivalent to a measure incorporating both low and very low food security.³ 

The 5 potential parent-reported child health statuses were dichotomized as excellent versus very good, good, fair, or poor; this dichotomization has been found to be a more relevant predictor of unfavorable health care outcomes among children than adult groupings of excellent with very good and good.²⁷  SHCN status was the moderator of interest, defined in the validated NSCH instrument as having heightened health care needs in at least 1 of 5 areas due to an underlying condition expected to last for at least 1 year.⁴³  The ultimate outcome was any emergency health care utilization. This was indicated by either 1 or “2 or more” visits to a “hospital emergency room” in the previous year (vs no visits).

We adjusted for individual, parent/caregiver, household, and state-level covariates. These included the child’s age in years, sex (male/ female were the only options in the survey), and race/ethnicity (aggregated to non-Hispanic white, non-Hispanic Black, non-Hispanic other, or Hispanic), whether any parent/caregiver was employed and whether any had a high school degree, household income as a percentage of the federal poverty level and whether there was a smoker in the household, and state-level unemployment, GDP, and income per capita.⁹  Adjusting for state, year, and state-by-year variables did not significantly alter model findings (Supplemental Tables 4 and 5), so we did not include them in the final model for parsimony.

For instrumental variables, we follow previous studies^5,9,30  in using state-level SNAP policies tracked in the US Department of Agriculture Economic Research Service SNAP Policy Database.⁴⁴  Benefits are federally funded, but states control policies related to ease of benefits access and maintenance. These policies are significantly associated with program enrollment. We, therefore, used as instruments the 2 policies that had the strongest associations with SNAP participation in our sample without being directly associated with the mediators or ultimate outcome: whether a state uses broad-based categorical eligibility for SNAP (through which certain households are deemed eligible on the basis of qualifying for other social programs)⁴⁵  and the proportion of nonearning, nonelderly SNAP units that have 1- to 3-month recertification periods (ie, a high administrative burden for participants; a SNAP unit is an individual or group of people living together who participate in SNAP).⁹ 

We conducted bivariate comparisons of child, parent/caregiver, household, and state-level covariates, SNAP participation, household food hardship, excellent health status, and any emergency health care use by SHCN status. For our fully adjusted model, given the binary nature of our exposure and outcome variables and the presence of covariates, we used the instrumental variables discussed above in the context of a bivariate probit model.⁴⁶  To assess the hypothesized mediation relationships, we estimated this model within a structural equation modeling frame in which household food hardship and health status mediated relationships between SNAP participation and emergency health care use. This involved modeling a latent variable for which SNAP participation and household food hardship were the measurement variables and the parameters for the relationships between this latent variable and each measurement variable were constrained to be equal to one another.⁴⁷  We modeled SNAP participation (and therefore the interaction of SNAP participation and SHCN status) as operating through household food hardship both on conceptual grounds (ie, SNAP is intended to reduce food hardships, and previous studies have revealed evidence for this effect^4,5,8–14 ) and for parsimony. For interpretability, we calculated predicted probabilities (accounting for both direct and indirect relationships when relevant) on the basis of the results of our model.

Because of model complexity, incorporating survey weights proved computationally infeasible. To assess potential bias from excluding weights, we compared weighted and unweighted results from simpler versions of our model (ie, an unadjusted version and versions with varying subsets of our covariates). Parameters from each did not meaningfully differ. At most, parameters in weighted versions had marginally larger magnitudes, indicating a slight bias toward the null from using an unweighted full model. Another potential concern was that computational difficulties in weighted models were caused by significant correlations between instrumental variables and residuals in the bivariate probit portion of the model because weights could theoretically magnify these correlations and cause the model to not converge. However, we assessed this possibility and found no significant correlations that could lead to such an issue. Relying on these checks, our data conformed with scenarios in which previous analyses have suggested that unweighted models can be used without a significant risk of bias,⁴⁸  and we, thus, proceeded with an unweighted final model. As such, for consistency, we also provide unweighted bivariate comparisons, which also did not qualitatively differ from weighted ones. Stata (Version 16) was used for all estimations.

Child, household, and parent/caregiver characteristics for our analytic sample are provided in Table 1. Compared with children without SHCN, children with SHCN were more likely to live in a household with someone who smokes cigarettes and have no employed parent or caregiver, despite being more likely to have at least 1 parent/caregiver with at least a high school level of education. Children with SHCN were also more likely to be in a household that participated in SNAP, be in a household that experienced food hardship, have less than excellent health status, and have had at least 1 emergency department visit in the past year (Table 2).

We present results from our final model graphically in Figure 1 (full model results provided in Supplemental Table 4). We provide adjusted predicted probabilities of each outcome variable (household food hardship, excellent child health status, and any child visit to a hospital emergency department) based on these results in Table 3. SHCN status was associated with a greater likelihood of experiencing household food hardship, nonexcellent health status, and at least 1 emergency department visit in the past year. Conversely, for all children, SNAP participation was associated with a lower likelihood of experiencing household food hardship, nonexcellent health status, and at least 1 emergency department visit in the past year. The magnitude of the associations between SNAP and each outcome was larger for children with versus without SHCN: 29% larger for household food hardship (−22.0 percentage points [pp] vs −17.1pp), 23% larger for excellent health status (+9.7pp vs +7.9pp), and 79% larger for visits to a hospital emergency department (−7.7pp vs −4.3pp; Table 3).

We found that household SNAP participation was associated with a lower likelihood of emergency health care use among children, that this relationship was mediated by a lower likelihood of household food hardship and a higher likelihood of excellent parent-reported child health status, and that the magnitude of these associations was larger among children with SHCN. The association of SNAP with a 31% reduction in the likelihood of experiencing household food hardship in the overall sample (an 18.3pp reduction from 59.7%) was similar to associations of SNAP with 31%⁵  and 38%⁹  reductions in the likelihood of household food hardship found in previous studies using similar methods (as measured by having either low or very low food security, a measure roughly equivalent to our measure of having either marginal food sufficiency or food insufficiency^3,49 ).

This suggests that our use of the single-item food insufficiency measure available in the NSCH, although not offering as much nuance as lengthier food insecurity tools,³  was unlikely to have altered our main findings. However, exploring the effects of using different food hardship measures may offer additional insights. For example, among all SNAP-participating households (not only those including children), Ratcliffe and colleagues⁵  found that SNAP was associated with a smaller reduction in the likelihood of more severe food security compared with food insecurity generally, indicating that more extreme food hardship experiences may be less elastic than milder ones. Future studies with additional food hardship measures may allow even richer assessments of associations between SNAP participation and downstream health and health care outcomes than have been found to date,⁵⁰  especially for children.

Still, we find evidence for meaningful associations with these downstream measures within our data. The conceptual model for the relationship between food hardship and negative health care consequences proposes mechanisms through which cycles of hyper and hypoglycemia drive long-term chronic disease development and short-term exacerbations of these conditions,³²  the latter of which are likely applicable to children with chronic conditions even if not developed over long timeframes. If the associations in the present data have a causal component, then it is most likely that the shorter-term pathways dominate what we detected.

However, even if SNAP alleviates harmful cycles of deprivation and sustenance with regard to food, the structural factors driving food hardship disparities in the first place remain intact. For example, the flow from racism to disparities in housing, health care, education, and other facets of life,^51–53  and the resulting long-term patterns of embodied stressors that increase risks for chronic disease development, are unlikely to be unwound by SNAP alone (even if SNAP yields benefits unrelated to food through substitution effects⁹ ). This is consistent with findings that most SNAP applicants seek benefits after spikes in food hardship experiences, indicating it is sought to prevent additional rather than initial harms.^4–7  In addition, shorter-term pathways are likely especially dominant among children for whom chronic disease onset is typically less related to long-term factors than for adults. Still, the potential benefits of SNAP alleviating shorter-term exacerbations are significant. Each exacerbation or insult to health that is not prevented can contribute to significant long-term harm,^54,55  particularly for children.^56–58 

Our finding of larger effect sizes among children with SHCN is consistent with previous work suggesting that the health status of these children is especially sensitive to food hardship.²⁷  It also emphasizes the potential importance of the shorter-term pathways through which SNAP benefits may improve health. Given the heightened material hardships experienced by children with SHCN generally,⁴⁰  it was conceivable that the benefits of SNAP participation would be muted for this group as the damage done by other concurrent challenges lessoned its potential effects. That this was not observed suggests that the immediate relief provided by SNAP may be strong enough to yield health gains even in the face of other long-term harms.

The current study has limitations. Although our modeling approach was designed to reduce bias caused by self-selection into SNAP and other potential confounding, the cross-sectional nature of our data still limits our ability to make causal inferences and rule out unmeasured sources of variation. Also, it is possible that parent-reported data could be imprecise, particularly for relatively subjective questions such as general health status. However, past research has revealed that parent-reported poor child health is associated with increased health care utilization,⁵⁹  a relatively more objective measure. We found the same association here, which serves to validate this child health status measure. The stigma associated with reporting poorly rated health status (as well as SHCN status, food hardship, and emergency health care) could produce bias, but stigma leading to underreporting of these conditions would have tended to bias our results toward the null. Additionally, our dichotomization of health status is generally a more conservative approach, and therefore more likely to lead to null results. However, we identified important relationships with this variable, which strengthens the study.

The simple measure of any SNAP participation over the past year limited the nuance of our analyses as well, although, again, the imprecision here would tend to bias results to the null. Relatedly, SNAP exists within an ecosystem of public benefit programs that are likely correlated. Although the instruments we chose were both significantly correlated with participation in SNAP but not with participation 3 other programs measured in the data (the Special Supplemental Nutrition Program for Women, Infants, and Children, Temporary Assistance for Needy Families, and the National School Lunch Program; Supplemental Table 6), the data did not ask about other programs, like child Supplemental Security Income, which may affect household food hardship as well.⁶⁰  Additionally, although food hardship is associated with leading causes⁶¹  of pediatric emergency visits (eg, respiratory disorders,^36,62  injury,⁶³  mental health⁶⁴ ), our inability to distinguish reasons for emergency health care visits prevented more nuanced analyses.

Our use of unweighted analyses also raises the specter of sampling error, nonresponse bias, and design effects; however, the fact that estimates did not differ between our weighted and unweighted versions of simpler versions of our model alleviates this concern. Additionally, concerns of overly tight standard errors in our final model were reduced given that the weighted models did not result in larger standard errors. Lastly, although we modeled multiple relationships, we ultimately measure an oversimplified version of reality. For example, the relationship between health status and emergency health care use is likely bidirectional. Future work with longitudinal data will likely be needed to further investigate such complexities.

We find evidence SNAP is associated with reduced food hardship, improved health status, and reduced emergency health care use among low-income children, with larger effects among children with SHCN. These relationships are important to weigh amid federal policy discussions determining SNAP benefit levels and state policy discussions determining the relative administrative ease or difficulty of SNAP access. They suggest access to SNAP may have additional benefits for children. These findings also suggest SNAP may have benefits beyond food hardship relief, through possible reductions in emergency health care use and associated healthcare spending. In addition, given the larger effects among children with SHCN, efforts to increase their access may be especially beneficial.

We thank Dr William Crown of Brandeis University for his thoughtful contributions to our approach to sensitivity analyses.

CI

confidence interval

NSCH

National Survey of Children’s Health

pp

percentage points

SHCN

special health care needs

SNAP

Supplemental Nutrition Assistance Program