Making the Most of Limited Resources: Predicting Food Insecurity

ABSTRACT

Objectives: This research assesses the feasibility of developing an inferential model of individual-level food insecurity by linking health-related social need (HRSN) survey data with health care claims and area-level data on social drivers of health.

Study Design: Logistic regression modeling linking attested food insecurity to individual-level health care claims data and area-level characteristics.

Methods: Food insecurity attestations were derived from a social risk survey deployed by a payer to be representative of membership by age, gender, urbanicity, and Social Vulnerability Index quartile and all insurance types aside from dual-eligible. Predictor variables were sourced from claims data and publicly available area-level data. Models were run with numerous selections of predictor variables and either including insurance type as predictors or stratifying by insurance type, in addition to several experimental models. Performance was primarily assessed by several goodness-of-fit measures, primarily area under the curve (AUC).

Results: Model performance exceeded the 0.7 AUC threshold of acceptable performance for models including insurance type as a predictor, but not for insurance type–specific models, with the best-performing model including all available variables. This model classified no individuals in the commercial sample and 97% of the Medicaid sample as having food insecurity, compared with 19% and 65% in attestations, respectively.

Conclusions: The results suggest that insurance type is a strong predictor of food insecurity, with differentiating identification by insurance type highlighting the importance of additional HRSN screening among commercial members and further outreach among Medicaid populations. Use of multilevel modeling may strengthen analyses through increased ability to address consideration of sourcing, scale, and setting of data.

Am J Manag Care. 2026;32(5):280-287. https://doi.org/10.37765/ajmc.2026.89936

Takeaway Points

This research yields recommendations for identification of individuals with health-related social needs (HRSNs) for intervention, building on policies and measures from organizations such as CMS and the National Committee for Quality Assurance, as follows:

• Insurance type is a strong predictor of food insecurity.
• Insurance type is a potential proxy for household income and family size.
• Efforts to expand screening for HRSNs should simultaneously target individuals with commercial insurance while interventions for individuals with Medicaid plans are underway, allowing for a high-touch, upstream impact while maximizing resources.

Extant research has extensively shown that health-related social needs (HRSNs) negatively impact health care utilization and outcomes.^1-3 For instance, food insecurity, defined by the US Department of Agriculture as a household-level economic and social condition of limited or uncertain access to adequate food, is associated with increased risk of chronic disease in adults, developmental and mental health problems in children, and obesity in both.^4-8 Screening for HRSNs provides a means of identifying individuals who may experience a tangible improvement in health through intervention to address any present social needs. For instance, a health system in Ohio and Michigan screens hospitalized patients for food insecurity using a questionnaire and operates “food pharmacies,” which provide counseling from dietitians for food-insecure patients, although this requires physician referral.⁹

With an increasing focus on health equity and addressing HRSNs, organizations such as CMS and the National Committee for Quality Assurance have revised their policies and measures to encourage hospitals and health plans to screen for HRSNs.^10-12 Conducting large-scale surveys is time-consuming and cannot be easily expanded or applied at scale for actionable outreach. Current literature has demonstrated that 79.2% of hospital respondents screened patients for HRSNs and 66.1% for food insecurity, although this common approach to HRSN screening requires individuals to be present in a clinical setting and self-identify.^13,14 Further barriers to individual-level HRSN screening include the need for administrative resources and that individuals from marginalized identities may be reluctant to consent to screenings.¹ The validity of self-reports as a source, the setting of HRSN screening, and the scale of the issue all pose substantial hurdles to proper identification of those with food insecurity needs; thus, inferential statistical modeling holds potential in overcoming these challenges by estimating food security likelihood.

Prior research has shown that important factors associated with food insecurity include age, race, and gender, but also socioeconomic and behavioral health features.^15-17 In addition, insurance coverage itself appears to demonstrate significant impact on food insecurity.¹⁸

Several limitations exist within prior research, namely exclusive population samples and self-reported data, both in food security assessment and/or insurance status. This research intends to compare evidence of individual survey attestations of food insecurity to inferential predictions to overcome the challenges of sources, scale, and setting. The current research exhibits a representative sampling survey by membership by age, gender, urbanicity, social vulnerability, and insurance coverage.¹⁹ By linking survey data with existing administrative health care claims at the individual and aggregated levels, using publicly available data, we aim to demonstrate the successes and challenges in identifying those with food insecurity.

METHODS

Model Design

We performed several stepwise analyses using logistic regression where the outcome was defined as a binary flag for food insecurity, as defined by personal attestations (Table 1). Each model included demographic information as predictor variables. Subsequent models added selected area-level characteristics or health condition variables, and the final model included all variables. Please see the eAppendix (available at ajmc.com) for a detailed description of data sources. Models were run either including insurance type as a predictor or stratified by insurance type. Further models assessed interactions between each area-level characteristic and age, gender, and the Elixhauser Comorbidity Index (ECI), replacing median household income and percentage of the population completing less than high school with tract-level variables from county-level variables, log-transforming median household income, and replacing insurance type with other covariates. The performance of logistic regression models was also compared with supervised machine learning, using random forest, including all variables on an 80% random training sample of the available data, with 1000 trees and 5 variables randomly sampled as candidates at each split. Analyses were performed using SAS 9.4M7 (SAS Institute Inc) and R 4.2.0 (R Foundation for Statistical Computing).

Assessing Model Performance

Model performance was primarily measured using the area under the curve (AUC) to assess discrimination and predictive quality. Akaike information criterion (AIC) and pseudo-R² (Nagelkerke) were assessed to consider the balance of model fit and parsimony, and further detail on variation explained. AUC calculation and receiver operating characteristic (ROC) analysis were completed using leave-one-out cross-validation. For models of interest, confusion matrices were also generated to compare model predictions to survey outcomes considering prediction from a 0.5 likelihood threshold. OR estimates were provided for included variables.

Response Variable

The response variable was a binary flag of member survey responses indicating evidence of attestation to food insecurity. Please see Falconi et al¹⁹ and the eAppendix for further details on survey design, sampling methodology, and specific survey questions used to identify food insecurity. Individuals eligible for the survey were those with complete address, age, and gender information who were active members of the health plan conducting the survey, residing in 2 states selected for geographic and demographic diversity. The survey was conducted with a stratified quota probability sampling strategy to ensure representativeness by insurance type, gender, age, Social Vulnerability Index ranking, and urbanicity of residence, with representativeness achieved across all insurance types except dual-eligible. A detailed cohort disposition diagram is available for the commercial sample in Falconi et al.¹⁹

Predictor Variables

This study included 3 categories of predictor variables: demographics, health conditions, and area-level characteristics. Member demographic information included age, gender, and type of insurance, sourced from survey data, covering commercial, Medicare Advantage (MA), Medicaid adults, and dual-eligible (a subset of MA). The presence of health conditions was defined dichotomously, set as 1 if a respondent had either 1 inpatient or 2 outpatient claims with a relevant International Statistical Classification of Diseases, Tenth Revision diagnosis code for a given condition and 0 otherwise, as assessed retroactively 1 year prior to the survey completion date. ECI was sourced from the same health care database as health condition data. Area-level measures included percentage of high school completion, life expectancy, percentage with poor or fair health, median household income, percentage uninsured, and extent of school segregation based on extant literature. See eAppendix Table 1 for detailed variable definitions and measurement periods.

RESULTS

Population Characteristics

Food insecurity varied by insurance type (P < .05), with approximately one-fifth of commercial and MA members attesting to food insecurity (19.12% and 17.30%, respectively), whereas more than half of Medicaid and dual respondents (65.13% and 58.88%, respectively) indicated food insecurity. Please see the eAppendix Figure for further details and eAppendix Table 2 for summary statistics of all variables.

Model Results

All models were run either with insurance type as a predictor or stratified by insurance type. Among models including insurance type as a predictor, AUC ranged from 0.71 to 0.76, pseudo-R² ranged from 0.27 to 0.31, and AIC ranged from 3129 to 3391 (Table 2 [part A and part B]). As stepwise inclusion of area-level characteristics (models 2-10) and health characteristics (models 11-21) advanced, AUC and pseudo-R² increased, and AIC performed worse in models including area-level measures (models 2-10) compared with the baseline demographic model (1) and better than baseline in the health characteristics models (11-21). The model, including all available variables (22), had the most favorable values for each of the aforementioned metrics compared with other models.

An exploratory version of model 1, replacing insurance type with income category and family size as predictors, had comparable AUC performance to the standard model 1 with insurance type (exploratory: AUC = 0.74; pseudo-R² = 0.23; AIC = 3054; standard: AUC = 0.74; pseudo-R² = 0.27; AIC = 3341). Additional variations of model 1, replacing insurance type with Area Deprivation Index (ADI) national percentiles or state deciles, or each area-level or health characteristic variable, had worse AUC and AIC performance than the standard model 1 (Table 2 and eAppendix Table 2).

Compared with other predictors, dual vs commercial and Medicaid vs commercial insurance type had notably high estimates (in model 22, dual vs commercial: OR, 4.56; 95% CI, 2.83-7.33; Medicaid vs commercial: OR, 6.35; 95% CI, 5.10-7.91) (Table 3). The odds of being identified with food insecurity did not vary by MA vs commercial (OR, 0.92; 95% CI, 0.60-1.42). Model 1, which contained only insurance type, age, and gender, saw a right shift in OR for insurance type compared with model 22 (dual vs commercial: OR, 6.69; 95% CI, 5.90-8.62; Medicaid vs commercial: OR, 7.13; 95% CI, 5.90-8.62; MA vs commercial: OR, 1.06; 95% CI, 0.71-1.60) (Table 3). Overall significant parameters in model 22 were insurance type for Medicaid vs commercial and dual vs commercial as noted earlier, age (OR, 0.99; 95% CI, 0.98-1.00), poor/fair health (OR, <0.001; 95% CI, <0.001-0.44), school segregation (OR, 0.21; 95% CI, 0.06-0.78), gender (OR, 1.41; 95% CI, 1.09-1.57), chronic obstructive pulmonary disease (OR, 1.55; 95% CI, 1.06-2.26), and ECI (OR, 1.09; 95% CI, 1.02-1.17) (Table 3).

All models, including insurance type as a predictor, had AUCs above 0.7, but no models specific to insurance type met this threshold (Table 2 [A-C]). There was little variation in AUC across models as the number of variables increased. AUC increased from 0.74 in the demographics-only model (1) to 0.76 in the model including demographics, area-level, and health conditions (22), a 0.02 increase in AUC (Figure [A and B]). Similarly, models with increased numbers of variables were associated with a minimal increase in AUC in the commercial model compared with the demographics-only model 1 (AUC increase of 0.04) and were not associated with a difference in AUC in the Medicaid model (Figure [C and D]). Exploratory models, including tract-level variables, log-transforming income, or including interaction terms, all yielded negligible AUC improvement (<0.01). Similarly, random forest using all insurance types as one of the predictors (comparable to model 22), as well as only among commercial and Medicaid populations, found equal or lesser (within 0.05) AUC to comparable logit models. The models consistently performed better among commercial membership than Medicaid (Figure [C and D]).

Model 22 tended to classify individuals in the commercial sample as not having food insecurity (0% predicted food insecurity vs 19% in survey results) and individuals in the Medicaid sample as having food insecurity (97% predicted food insecurity vs 65% in survey results) (Table 4). These results lead to a ratio of type I to type II errors of 1.35 with insurance type as a predictor, 0 for the commercial-only model, and 32.53 for the Medicaid-only model (Table 4).

DISCUSSION

With this research, we set out to assess the successes and challenges in identifying food insecurity through inferential modeling addressing considerations of sourcing of attestations, scale required of screening, and setting of data collection. The results suggest that predictability of food insecurity models was notably dependent on insurance type. For instance, in models containing all potential claims-based and area-level predictors (model 22), inclusion of insurance type as a predictor increased AUC to 0.76—the best-performing model—compared with an AUC of 0.66 in the comparable model restricting to commercial members and an AUC of 0.54 in the model restricting to Medicaid (Figure [B, C, and D]). Across models 1 through 22, ORs consistently indicated that insurance type had the strongest association with food insecurity compared with other predictors. These ORs were lower in magnitude in model 22, where all available covariates were included, compared with model 1 with only 2 covariates; additionally, model 22 had the lowest AIC of all models, suggesting the importance of insurance type on attested food insecurity but highlighting the need to adjust for area-level characteristics and individual health conditions to address confounding and model oversimplification. Nevertheless, additional area-level social and health condition variables contributed little to predictive accuracy beyond insurance type when considering the modest increase in AUC between models 1 and 22, as well as reduced AUC and increased AIC in experimental versions of model 1, replacing insurance type with each predictor variable, as well as block group–level ADI data. These findings together highlight potential overlapping or correlated factors with insurance status and especially the relevance of insurance type as a predictor beyond simply serving as a proxy for social disadvantage or health conditions. Further, exploratory model results found that insurance type may still serve as a proxy for income and family size (2 strong predictors for food insecurity that are often not collected by health plans) in the context of food insecurity inferential modeling, given comparable AUC performance (0.74 in both exploratory and standard models) and improved AIC performance (3054 vs 3341, respectively).

In review of these models, it is essential to note that the predictors introduced were reflective of the 3 problems of source, scale, and setting. Inclusion of clinical parameters addresses considerations of sourcing by lacking reliance on member attestations and scale through the wide availability of electronic health record data. However, consideration of setting still needs to be addressed, given the requirement of presence in a clinical setting for these data to be collected. Modeling results support inclusion of these data, with decreasing AIC and increasing AUC and pseudo-R² as stepwise inclusion of individual health conditions increased in models 11 through 21. Inclusion of area-level characteristics addresses scale through consistent availability of data at a national level and setting by removing reliance on clinical settings for data collection. However, sourcing of these data continues to heavily rely on individual attestations downstream through means such as the American Community Survey. Although modeling results demonstrated increased AUC and pseudo-R² through stepwise inclusion of area-level characteristics in models 2 through 10, AIC increased, calling into question their relevance alone for improving model fit while considering parsimony. Model 22, including all available health conditions and area-level characteristics, had the highest AUC and pseudo-R² and lowest AIC of the 22 models, highlighting the complementary strengths of these sources.

The models with the highest AUC for predicting food insecurity included insurance type as a predictor variable. However, the best-performing model (22) did not identify any commercial members as food insecure, although 19% of commercial survey respondents indicated this HRSN. Similarly, it predicted that 97% of the Medicaid sample had food insecurity, whereas 65% of the Medicaid adult sample reported food insecurity in survey data. Although the results elucidate the differing variation in food insecurity by insurance type, they fail to capture the heterogeneity of need within each insurance type’s membership. While this research has identified several limitations in inferential modeling of food insecurity and HRSNs more broadly, critical actionable steps suggested from these results are to (1) target HRSN screenings in commercially insured populations, given that their heterogeneity of needs may not be fully represented in inferential modeling, and (2) take more active steps to target interventions within the Medicaid community, given the high baseline need and a population defined as having limited resources. Insurance type is an especially powerful lever for action due to its demonstrated relevance to food insecurity, universal access across health care settings, and potential to serve as a proxy for more proximate, less measured factors influencing food insecurity, such as income and family size. Although this article does not advocate for screening at a specific level, its context is especially appropriate for plans due to the accessibility of health-related data, particularly the parameters covered in the models presented here, the ability to take decisive action, and a vested interest in members’ well-being. For instance, a health plan was able to associate a medically tailored meals program for dual-eligible members with a 50% reduction in inpatient admissions and a 70% decrease in emergency department visits.²⁰ In terms of practical implementation, claims-based measures of social need are insufficient in assessing commercial populations’ HRSNs, identifying less than 2% of privately insured adults with HRSNs,²¹ highlighting the importance of implementation of self-reported surveys. Current examples of implementations of self-report surveys deployed by payers have been published by Elevance Health,¹⁹ a data source for this manuscript, and Blue Cross Blue Shield of Rhode Island.²² Stakeholders may especially be persuaded to adopt these measures due to the logit models utilized, allowing interpretations of each parameter for a convincing call to action, vs the black-box nature of machine learning models such as random forest, which had equal or lesser AUC performance.

Ethical and Equity Considerations

Although this article demonstrates that insurance type is a strong predictor of food insecurity, it is essential to avoid oversimplifying this nuanced topic into a single characteristic. Individuals qualifying for Medicaid are, by definition, a population with increased need, but a focus on this association may overgeneralize this population as a monolith with the same social conditions. Further, this line of thinking risks missing identifying individuals in commercial plans with food insecurity for intervention. The models presented here were designed to address these concerns by including diverse variables spanning HRSNs to capture as fully as possible individuals’ circumstances.

Limitations

HRSNs tend to be widely distributed, poorly measured, and long-term in nature, requiring a comprehensive approach to their observation. There remains a degree of heteroscedasticity in the models presented, arising from factors that are difficult or impossible to measure or quantify; for instance, housing stability may be influenced by interpersonal relationships or social support networks, which would not traditionally be captured in measures of housing stability. Future research should consider methods to address unobserved heterogeneity, such as methodological changes, like inclusion of fixed effects to control for time-invariant characteristics. Additionally, area-level data were matched to individuals at as granular a level as possible and as close to their survey attestations as possible, but the aggregate nature of these data limits their generalizability to the individual level. Generalizability is further limited because the data analyzed were designed to be representative of a health plan’s membership across 2 states, rather than the overall US population. As such, this has prompted several insights that we wish to elucidate. First, area-level social needs data and individual-level clinical health data should not be treated as proxies for an individual’s HRSNs. Self-attestation does not appear to be easily replicated at an aggregate level, as it glosses over the problem of ecological validity. This assumption precludes the possibility of observing effects of cross-dimensional factors. Others have suggested combining individual- and area-level data to focus health care and social service resources to address HRSNs, arguing that such approaches remain too far downstream to address drivers of poor health.¹⁴ However, this approach, while appropriate for member-level intervention, retains ecological problems, which inherently remain tied to observation. Without accounting for variation at multiple levels of analysis, area-level metrics may be blind to the proximate needs of individuals. Additionally, model 22 had a ratio of type I to type II errors of 1.35, which would ideally be less than 1, given the preference for a type II error over a type I error when considering the consequences of erroneously missing someone who indeed has food insecurity. ORs were well above 1 for dual vs commercial and Medicaid vs commercial for model 1, with reduced ORs and CIs, but still well above 1 in model 22 (Table 3). The MA vs commercial OR remained close to 1, with 95% CIs overlapping 1 across models. This demonstrates that inclusion of additional predictors accounted for part of the initial association between insurance type and odds of food insecurity, but the association remains strong for insurance type. Thus, the 97% predicted vs 65% actual food insecurity discrepancy among Medicaid patients is likely contributing to the high ORs noted in Table 3 and indicates a bias that leads to overclassification. This suggests the model may not capture heterogeneity within the Medicaid population, thereby limiting its generalizability beyond insurance type.

CONCLUSIONS

This exercise identified several best practices for HRSN screening and prediction. The relevance of insurance type in inferential modeling, but challenges in distinguishing need within each insurance type, highlight the actionable step of focusing screenings to identify those who could benefit from HRSN-related intervention among the heterogeneous commercially insured population, alongside active intervention in the more homogenous Medicaid population, allowing for a high-touch, upstream impact while maximizing resources. Further, the finding associating insurance type with family size and income may provide a greater sense of members’ social context without collection of survey data, allowing interventions to be further preemptively tailored. As the study of screening for and addressing HRSNs continues to evolve, this article serves as a firm marker on the journey to better understanding how we can best leverage data and resources to give our all at improving the health of others. 

Author Affiliations: Carelon Research (CS, PO, MJ, WC), Wilmington, DE; Elevance Health, Indianapolis, IN (JW), and Washington, DC (SA).

Source of Funding: No sources of external funding were used to assist in the preparation of this manuscript. Research was completed as part of the usual employment obligations.

Author Disclosures: The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

Authorship Information: Concept and design (CS, PO, MJ, WC, SA); acquisition of data (CS); analysis and interpretation of data (CS, MJ, WC, JW, SA); drafting of the manuscript (CS, WC, JW); critical revision of the manuscript for important intellectual content (CS, PO, MJ, WC, JW, SA); statistical analysis (CS); obtaining funding (WC); and supervision (PO, WC).

Address Correspondence to: Cory Silver, MPH, Carelon Research, 123 Justison St, Ste 200, Wilmington, DE 19801. Email: cory.silver@carelon.com.

REFERENCES

1. Berkowitz SA, Baggett TP, Edwards ST. Addressing health-related social needs: value-based care or values-based care? J Gen Intern Med. 2019;34(9):1916-1918. doi:10.1007/s11606-019-05087-3

2. Billioux A, Verlander K, Anthony S, Alley D. Standardized screening for health-related social needs in clinical settings: the Accountable Health Communities screening tool. NAM Perspect. Published online May 30, 2017. doi:10.31478/201705b

3. Ryan JL, Franklin SM, Canterberry M, et al. Association of health-related social needs with quality and utilization outcomes in a Medicare Advantage population with diabetes. JAMA Netw Open. 2023;6(4):e239316. doi:10.1001/jamanetworkopen.2023.9316

4. Hernandez DC, Reesor LM, Murillo R. Food insecurity and adult overweight/obesity: gender and race/ethnic disparities. Appetite. 2017;117:373-378. doi:10.1016/j.appet.2017.07.010

5. Gregory CA, Coleman-Jensen A. Food Insecurity, Chronic Disease, and Health Among Working-Age Adults. US Department of Agriculture Economic Research Service. July 31, 2017. doi:10.22004/ag.econ.261813

6. Nord M. Food Insecurity in Households With Children: Prevalence, Severity, and Household Characteristics. US Department of Agriculture Economic Research Service. September 2009. Accessed July 11, 2025.
https://files.eric.ed.gov/fulltext/ED508211.pdf

7. Rose-Jacobs R, Black MM, Casey PH, et al. Household food insecurity: associations with at-risk infant and toddler development. Pediatrics. 2007;121(1):65-72. doi:10.1542/peds.2006-3717

8. Huang J, Barnidge E. Low-income children’s participation in the National School Lunch Program and household food insufficiency. Soc Sci Med. 2015;150:8-14. doi:10.1016/j.socscimed.2015.12.020

9. Cacciaglia A. Food as medicine: addressing hunger in the community. EpicShare. September 13, 2021. Accessed December 9, 2024. https://www.epicshare.org/share-and-learn/food-as-medicine-addressing-hunger-in-the-community

10. Billioux A, Conway PH, Alley DE. Addressing population health: integrators in the Accountable Health Communities model. JAMA. 2017;318(19):1865-1866. doi:10.1001/jama.2017.15063

11. Reynolds A. Social need: new HEDIS measure uses electronic data to look at screening, intervention. National Committee for Quality Assurance. November 2, 2022. Accessed July 11, 2025. https://www.ncqa.org/blog/social-need-new-hedis-measure-uses-electronic-data-to-look-at-screening-intervention/

12. Ryder E. Foundational standards to support effective, sustainable advancement of equity. National Committee for Quality Assurance. 2024. Accessed July 11, 2025. https://www.ncqa.org/wp-content/uploads/NCQA-Health-Equity-Accreditation-Programs-Webinar_May-24_-2024.pdf

13. Ashe JJ, Baker MC, Alvarado CS, Alberti PM. Screening for health-related social needs and collaboration with external partners among US hospitals. JAMA Netw Open. 2023;6(8):e2330228. doi:10.1001/jamanetworkopen.2023.30228

14. Chen A, Gwynn K, Schmidt S. Addressing health-related social needs in the clinical, community, and policy domains. JAMA Netw Open. 2023;6(4):e239274. doi:10.1001/jamanetworkopen.2023.9274

15. Coleman-Jensen AJ. U.S. food insecurity status: toward a refined definition. Soc Indic Res. 2009;95(2):215-230. doi:10.1007/s11205-009-9455-4

16. Goldberg SL, Mawn BE. Predictors of food insecurity among older adults in the United States. Public Health Nurs. 2014;32(5):397-407. doi:10.1111/phn.12173

17. Steiner JF, Stenmark SH, Sterrett AT, et al. Food insecurity in older adults in an integrated health care system. J Am Geriatr Soc. 2018;66(5):1017-1024. doi:10.1111/jgs.15285

18. Daw JR, Underhill K, Liu C, Allen HL. The health and social needs of Medicaid beneficiaries in the postpartum year: evidence from a multistate survey. Health Aff (Millwood). 2023;42(11):1575-1585. doi:10.1377/hlthaff.2023.00541

19. Falconi AM, Johnson M, Chi W, Stephenson JJ, Overhage JM, Agrawal S. Health related social needs and whole person health: relationship between unmet social needs, health outcomes, and healthcare spending among commercially insured adults. Prev Med Rep. 2023;36:102491. doi:10.1016/j.pmedr.2023.102491

20. Food as medicine. Elevance Health Foundation. Accessed March 31, 2025. https://elevancehealth.foundation/food-as-medicine

21. Liss DT, Cherupally M, Kang RH, Aikman C, Cooper AJ, O’Brien MJ. Social needs identified by diagnostic codes in privately insured U.S. adults. Am J Prev Med. 2022;63(6):1007-1016. doi:10.1016/j.amepre.2022.07.009

22. Methodologies. RI Life Index. Accessed April 12, 2025. https://rilifeindex.org/2024-methodologies/