Comparing Care Coordination and Food Assistance Referrals to Improve Diabetes Outcomes: a Cohort Study

Abstract

Background:

Care coordination or food assistance referrals can both improve diabetes outcomes by addressing health-related social needs such as food insecurity. However, the evidence comparing these approaches is limited.

Objective:

To examine whether care coordination and food assistance referrals are differentially associated with diabetes outcomes.

Design:

Longitudinal, observational cohort study.

Participants:

Adults with type 2 diabetes seen in community-based health centers referred for care coordination or food assistance.

Main Measures:

The primary outcome was hemoglobin A1c (HbA1c). Secondary outcomes were systolic and diastolic blood pressure (SBP and DBP) and low-density lipoprotein cholesterol (LDL). The primary timepoint was 6 months after referral; 12, 18, and 24 months were secondary timepoints. Targeted minimum loss estimation was used to adjust for age, sex, race and ethnicity, language, income, health insurance, comorbidities, body mass index, census tract social vulnerability index, food insecurity, housing instability, transportation barriers, and baseline outcome levels.

Key Results:

2108 individuals were included (food assistance = 960; care coordination = 1148). Mean age was 55.2 years (SD: 12.4), 61.9% were female, and mean baseline HbA1c was 7.6% (SD: 2.1%). Results did not suggest clinically meaningful differences in outcomes when comparing food assistance referrals to care coordination referrals. For example, the estimated 6-month difference comparing food assistance to care coordination was 0.01% (95%CI −0.15 to 0.18) for HbA1c; −0.20 mm Hg (95%CI −1.62 to 1.22) for SBP; 0.78 mm Hg (95%CI −0.14 to 1.69) for DBP; and 2.78 mg/dL (95%CI −2.30 to 7.86) for LDL. Results were similar at other time points.

Conclusions:

This study did not find clinically meaningful differences in diabetes outcomes between adults who received a referral for food assistance versus care coordination. Both interventions have been proven effective in other studies, so clinics might decide which intervention to offer based on factors such as feasibility and patient preference.

Food insecurity—uncertain access to the food needed for an active, healthy life¹—is associated with worse diabetes outcomes such as glycemic control, blood pressure, cholesterol, and with microvascular and macrovascular complications.^2–8 Recognition of these associations underlies an increase in clinic-based interventions to address these issues as a means to improve clinical outcomes.^8–10 Although there are many possible ways to do this, two major types of approaches now have evidence bases suggesting their effectiveness relative to usual care.^2,8,11,12 The first, care coordination, involves care managers or community health workers working with individuals over time to meet health-related social needs.¹³ The effectiveness of care coordination interventions has been demonstrated in multiple randomized trials and other studies.^11,14–19 A second approach is food assistance—either providing a subsidy so individuals can purchase healthy food or directly providing healthy food.⁸ Food assistance also has a substantial evidence base, though it is not yet as developed as that for care coordination.^3,20–26

Understanding whether care coordination or food assistance better improves diabetes outcomes could inform the design of health-related social needs interventions—for instance, if one were superior on average or for some types of patients, a clinic might prefer to implement that approach. This is particularly relevant as many clinics do not currently have ways to address health-related social needs, but may want to, given their high prevalence. Since these approaches are different and require different resources to set-up and maintain, understanding whether differences in outcomes can be expected may help inform implementation decisions. However, there has been little prior comparative effectiveness research concurrently examining care coordination and food assistance. Randomized clinical trials could provide relevant comparative effectiveness evidence, but trials are slow and expensive to conduct, and may enroll non-generalizable samples that limit the applicability of the findings.^27–29 Therefore, randomized trials can often be usefully complemented by observational research involving larger and more generalizable samples.

To that end, we conducted an observational study, set in community-based health centers, to compare the effectiveness of care coordination and food assistance referrals on diabetes outcomes—specifically hemoglobin A1c (HbA1c, primary outcome), systolic blood pressure (SBP), diastolic blood pressure (DBP), and low-density lipoprotein cholesterol (LDL).

Methods

Study Design

This retrospective cohort study used a new-user, active comparison design to compare food assistance versus care coordination referrals for their association with diabetes outcomes.^30,31 We used a target trial emulation approach to study design, described below and in eTable 1.^29,32,33 Study data were collected from June 2016 to June 2024, and analyses were conducted from February 2025 to June 2025. The UNC institutional review board approved this study. Reporting followed STROBE guidelines for cohort studies.³⁴ Clinical trial number: not applicable.

Study Sample

Data for this study came from the electronic health record (EHR) of OCHIN, Inc (not an acronym), which provides a shared EHR to community-based health centers throughout the U.S. Using a previously validated algorithm, we identified adults (age > 18 years) with type 2 diabetes mellitus who had received a referral for either food assistance or care coordination.^33,35–39

Exposure

We previously validated an algorithm that categorizes referrals for care coordination and food assistance referrals for health-related social needs within EHR data.³⁹ That validation process involved conducting manual chart review to determine actions taken as part of a given referral, and then to categorize types of referrals. Care coordination referrals involved providing ongoing support (that is, repeated interaction rather than a single interaction) to an individual patient to address resource barriers, including food insecurity, in the context of clinical care. This could include both in-clinic support provided by care managers and out-of-clinic support provided by community health workers. This study focused on care coordination related to addressing health-related social needs, rather than clinical services likely diabetes self-management support. We designed the study this way as we viewed services such as diabetes self-management support as complementary with, rather than substitutes for, health-related social needs interventions, and thus did not believe comparing them was warranted. Food assistance referrals all shared the direct provision of food resources (either in-kind provision of food or provision of subsidies so that individuals could purchase food).

When developing this algorithm,³⁹ we found that referral documentation was sporadic, likely due to the lack of structured EHR fields for this purpose. For this reason, it was important to use an active comparator, new user design (comparing individuals with documented referrals, rather than comparing individuals with versus without referrals).³¹ This was because the set of individuals without documented referrals likely included individuals who had been referred but for whom that information was not captured in the EHR. Comparing individuals with versus without referrals would thus likely have been biased by misclassification.³¹

Outcomes and Time Points

The primary outcome was HbA1c.⁴⁰ Secondary outcomes were SBP, DBP, and LDL.^41,42 LDL was selected as an outcome over other lipid variables (e.g., triglycerides, total cholesterol, or high-density lipoprotein cholesterol) as it is a common target in diabetes management recommendations.^41,42 The primary time point was 6 months from initial referral. Secondary time points were 12, 18, and 24 months. These time points were selected based on past trials of the studied intervention types and the timeframes over which interventions might be expected, physiologically, to improve outcomes. ^{11,14–19,24–26}

Covariates

We considered several covariates, assessed at baseline, that prior work indicated could confound the association between referral received and study outcomes.^3–5,8 These were: date of initial referral, age, sex, race and ethnicity, primary language, income (expressed as a percentage of the federal poverty line in the year measured, which accounts for household size), health insurance (private, Medicare, Medicaid and other public, or uninsured), the Gagne ICD-10 code-based comorbidity index,⁴³ census tract social vulnerability index (SVI),⁴⁴ mean HbA1c, SBP, DBP, LDL, and body mass index (BMI) in the year prior to initial referral, and results of food insecurity, housing instability, and transportation barrier screening. Food insecurity, housing instability, and transportation barrier screening was conducted in clinics using several validated tools, most commonly the Protocol for Responding to and Assessing Patient’s Assets, Risks, and Experience (PRAPARE) or the Accountable Health Communities Health-Related Social Needs Screening Tool.^36,45–48

Statistical Analysis

The goal of our statistical analysis was to estimate the expected level of study outcomes had individuals received a care coordination referral or a food assistance referral, and to contrast those expected outcomes. In this observational study, the estimand is the analogue of what would be an intention-to-treat average treatment effect (ITT ATE) estimated in a randomized clinical trial.^29,32

Analyses used the date of initial referral for either care coordination or food assistance as the index date or ‘day 0’: the date that an individual became eligible for the study, was assigned an exposure status (i.e., either care coordination referral or food assistance referral), and study follow-up began.⁴⁹ Individuals were analyzed based on the referral they initially received (care coordination or food assistance), treated as a static exposure. In this dataset, no individual who received an initial food assistance referral subsequently received a care coordination referral, and no individual who initially received a care coordination referral subsequently received a food assistance referral. We used data up to one year prior to the index date for baseline covariate assessment.

For analysis, we used targeted minimum loss estimation (TMLE).^50–52 TMLE fits two models—one that estimates the relationship between the exposure and the study outcomes, conditional on measured covariates, and another that estimates the joint probability of exposure and remaining uncensored, conditional on measured covariates.⁵³ It then combines estimates from these models to produce a ‘doubly robust’ estimate of the expected outcome and the difference in expected outcomes between different exposures. TMLE contains elements similar to g-formula computation, propensity score weighting, and inverse probability of censoring weighting.^53,54 To minimize the risk of model misspecification, when fitting the models we used a ‘SuperLearner’ approach, which uses an ensemble of estimation methods, including ones suited for possible non-linearities and interactions in the data.^55–58 The ‘library’ of methods we used was: mean estimation, generalized linear models, multivariate adaptive regression spines, and random forest. TMLE analyses included the variables described in the exposures, outcomes, and covariates sections. 10-fold cross-validation was used for model fitting.

To help account for missing data (eTable 2), we used multiple imputation by chained equations, under a missing at random assumption.^59,60 We created 50 multiple imputation datasets, and combined estimates across datasets using Rubin's rules.⁵⁹

Because many advocate for providing food insecurity interventions without requiring positive food insecurity screening, our primary analyses included food insecurity status (food secure or food insecure) as a covariate but did not require that individuals report food insecurity at baseline to be included.^61–63 In sensitivity analyses, we repeated our main analyses among those who reported food insecurity at baseline. In another set of sensitivity analyses, we separately examined outcomes at the 6-month time point for the period prior to the COVID-19 public health emergency (index date on or before Sept 1, 2019) and the period after the COVID-19 public health emergency had begun (index date or after March 16, 2020).

The main analyses estimate an average association, but that could hide heterogeneity of associations by groups defined by different combinations of characteristics. To investigate this possibility descriptively, we also examined, for each outcome, the quintile of observations where the difference in outcomes if referred for food assistance versus care coordination was largest, and the quintile where the difference was smallest. We plotted the density of the distribution of estimated differences and summarized the differences in demographic characteristics between the quintiles using standardized mean differences.^35,64

Analyses were conducted in SAS version 9.4 and R version 4.4.3. Key packages in R were lmtp, mice, SuperLearner, earth, and ranger.^65–67

Results

2108 individuals were included in this study—1148 of whom received a food assistance referral and 960 of whom received a care coordination referral (Table 1). Participants lived in 7 states and were seen at 128 distinct clinical locations. The mean age of the cohort was 55.2 years (SD: 12.4), 61.9% were female, 57.4% reported their race as White, and 42.4% reported their ethnicity as Hispanic or Latino. Mean HbA1c in the year prior to initial referral was 7.6% (SD: 2.1), mean SBP was 129.7 mm Hg (SD: 12.5), mean DBP was 78.4 mm Hg (SD: 8.1) and mean LDL was 101.5 mg/dL (SD: 34.8).

Table 1:

Demographic Characteristics of Cohort at Time of Initial Referral

Characteristic	Overall	Care Coordination	Food Assistance	P-value
	N = 2108	N = 960	N = 1148
	Mean (SD) or N (%*)	Mean (SD) or N (%*)	Mean (SD) or N (%*)
Age at Initial Referral, years	55.2 (12.4)	54.9 (13.3)	55.4 (11.7)	0.41
Female	1305 (61.9)	559 (58.2)	746 (65.0)	0.002
Racial Identity				<0.001
American Indian/Alaska Native	22 (1.0)	15 (1.6)	7 (0.6)
Asian	67 (3.2)	13 (1.4)	54 (4.7)
Black	641 (30.4)	51 (5.3)	590 (51.4)
Multiple	23 (1.1)	11 (1.1)	12 (1.0)
Native Hawaiian or Other Pacific Islander	6 (0.3)	2 (0.2)	4 (0.3)
Not Reported	140 (6.6)	85 (8.9)	55 (4.8)
White	1209 (57.4)	783 (81.6)	426 (37.1)
Hispanic or Latino Ethnicity	893 (42.4)	399 (41.6)	494 (43.0)	0.79
Primary Language Other Than English	1246 (59.1)	578 (60.2)	668 (58.2)	0.37
Comorbidity Index	0.02 (0.23)	0.05 (0.33)	0.00 (0.09)	<0.001
Health Insurance				<0.001
Medicaid and Other Public	714 (33.9)	247 (25.7)	467 (40.7)
Medicare	494 (23.4)	272 (28.3)	222 (19.3)
Private	198 (9.4)	91 (9.5)	107 (9.3)
Uninsured	683 (32.4)	346 (36.0)	337 (29.4)
Household Income as Percentage of Federal Poverty Threshold	120.1 (214.8)	193.6 (280.3)	53.0 (84.2)	<0.001
Food Insecurity				<0.001
Food Secure	472 (22.4)	293 (30.5)	179 (15.6)
Food Insecure	672 (31.9)	262 (27.3)	410 (35.7)
No Recorded Assessment	964 (45.7)	405 (42.2)	559 (48.7)
Housing Instability				<0.001
Stable Housing	734 (34.8)	344 (35.8)	390 (34.0)
Unstable Housing	300 (14.2)	104 (10.8)	196 (17.1)
No Recorded Assessment	1074 (50.9)	512 (53.3)	562 (49.0)
Transportation Barriers				<0.001
No Transportation Barriers	730 (34.6)	353 (36.8)	377 (32.8)
Transportation Barriers	285 (13.5)	82 (8.5)	203 (17.7)
No Recorded Assessment	1093 (51.9)	525 (54.7)	568 (49.5)
Social Vulnerability Index at Census Tract Level	0.69 (0.25)	0.63 (0.24)	0.74 (0.25)	<0.001
Mean Body Mass Index in Year Prior to Initial Referral, kg/m2	33.85 (7.95)	34.33 (8.28)	33.45 (7.64)	0.01
Mean HbA1c in Year Prior to Initial Referral, %	7.56 (2.05)	7.76 (2.08)	7.42 (2.01)	<0.001
Mean Systolic Blood Pressure in Year Prior to Initial Referral, mm Hg	129.65 (12.46)	128.45 (12.54)	130.61 (12.32)	<0.001
Mean Diastolic Blood Pressure in Year Prior to Initial Referral, mm Hg	78.40 (8.06)	76.34 (8.48)	80.05 (7.31)	<0.001
Mean LDL Cholesterol in Year Prior to Initial Referral, mg/dL	101.50 (34.78)	95.99 (37.33)	105.14 (32.50)	<0.001

^*Percentages are calculated from entire sample, including observations with missing values

Comorbidity index in the study sample ranged from −1 to 2, with greater scores indicating greater comorbidity; social vulnerability index scores in the study samples ranged from 0.005 to 0.9995, with greater scores indicating greater risk; HbA1c = hemoglobin A1c; LDL = low density lipoprotein

At baseline, compared with those who received care coordination referrals, those who received food assistance referrals were more likely to report food insecurity (37.5% vs. 27.3%, p <.001), had lower HbA1c (7.4% vs. 7.8%, p < .001), but had greater SBP (130.6 mm Hg vs. 128.5 mm Hg, p < .001), DBP (80.1 mm Hg vs. 76.3 mm Hg, p < .001), and LDL (105.1 mg/dL vs. 96.0 mg/dL, p < .001).

Results from the target trial emulation did not suggest meaningful differences in outcomes when comparing food assistance referrals to care coordination referrals (Table 2, Figure 1, and eTable 3). For example, the estimated difference in HbA1c at 6 months comparing food assistance to care coordination was 0.01% (95% Confidence Interval[CI] −0.15 to 0.18; p=0.88); the estimated difference in SBP was −0.20 mm Hg (95%CI −1.62 to 1.22; p =0.78); the estimated difference in DBP was 0.78 mm Hg (95%CI −0.14 to 1.69; p=0.09) and the estimated difference in LDL was 2.78 mg/dL (95%CI −2.30 to 7.86; p=0.28). Results were similar at other time points.

Table 2:

Estimated Differences in Means Comparing Food Assistance to Care Coordination

Timepoint	Hemoglobin A1c		Systolic Blood Pressure		Diastolic Blood Pressure		LDL Cholesterol
	Estimated Difference in Mean, % (95% CI)	p-value	Estimated Difference in Mean, mm Hg (95% CI)	p-value	Estimated Difference in Mean, mm Hg (95% CI)	p-value	Estimated Difference in Mean, mg/dL (95% CI)	p-value
6 Months	0.01 (−0.15 to 0.18)	0.88	−0.20 (−1.62 to 1.22)	0.78	0.78 (−0.14 to 1.69)	0.09	2.78 (−2.30 to 7.86)	0.28
12 Months	0.05 (−0.14 to 0.23)	0.62	−0.41 (−1.96 to 1.15)	0.61	1.38 (0.37 to 2.39)	0.01	3.58 (−2.07 to 9.23)	0.21
18 Months	0.15 (−0.06 to 0.35)	0.15	−0.07 (−1.87 to 1.73)	0.94	0.56 (−0.44 to 1.56)	0.27	2.87 (−2.47 to 8.22)	0.29
24 Months	0.00 (−0.24 to 0.23)	0.97	−0.25 (−2.09 to 1.59)	0.79	0.51 (−0.56 to 1.58)	0.35	4.01 (−1.92 to 9.94)	0.18

Estimated difference in mean compares the estimated difference in the mean of the outcome between counterfactual scenarios in which individuals were referred to food assistance versus care coordination, with care coordination as the reference condition. Because lower outcomes are desirable, a positive value indicates estimated benefit for care coordination. The differences in means were estimated using a longitudinal targeted minimum loss estimation approach.

LDL = Low Density Lipoprotein

CI = Confidence Interval

Figure 1.

Circles represent adjusted mean outcomes (and 95% confidence intervals) for care coordination and food assistance for hemoglobin A1c, systolic and diastolic blood pressure, and low-density lipoprotein (LDL) cholesterol, estimated using targeted minimum loss estimation. Triangles depict unadjusted values. Adjusted values at baseline (0 months), demonstrate that the adjustment approach appropriately balanced outcomes means at baseline. Note: unadjusted results were offset slightly on the x-axis to improve readability.

In sensitivity analyses examining only those who reported food insecurity at time of initial referral (eTable 4), we similarly observed no meaningful differences when comparing food assistance to care coordination (eTable 5). For example, when comparing food assistance to care coordination in this cohort at 6 months, the estimated difference in HbA1c was −0.01% (95%CI −0.27 to 0.25; p=0.93); the estimated difference in SBP was −0.94 mm Hg (95%CI −3.20 to 1.32; p =0.41); the estimated difference in DBP was 0.28 mm Hg (95%CI −1.05 to 1.60; p=0.68) and the estimated difference in LDL was 3.80 mg/dL (95%CI −4.26 to 11.87; p=0.35). Results were similar at other time points. Results from analyses restricted to the pre- and post-COVID-19 public health emergency periods were similar to each other (eTable 6).

Descriptive analyses showed relatively little heterogeneity (eFigures 1–4 and eTables 7–10). For example, the quintile with the smallest difference when receiving a food assistance referral versus a care coordination referral had a mean estimated difference in HbA1c of −0.06% (SD: 0.04), indicating a slightly lower HbA1c if receiving a food assistance referral rather than a care coordination referral. The quintile with the largest difference when receiving a food assistance referral versus a care coordination referral had a mean estimated difference in HbA1c of 0.09% (SD: 0.04), indicating a slightly higher HbA1c if receiving a food assistance referral rather than a care coordination referral. Neither of these differences are likely to be clinically meaningful. When examining standardized mean differences, demographic differences between the quintiles were usually small, mostly less than 0.10. Results were similar for the SBP, LDL outcomes, with DBP outcomes showing the most heterogeneity.

Discussion

This retrospective observational cohort study of adults with type 2 diabetes seen in community-based health centers did not find meaningful differences between receiving a care coordination referral or receiving a food assistance referral on HbA1c, SBP, DBP, or LDL, over a 24-month follow-up period. The findings were consistent both overall and for those with documented food insecurity at the time of the referral. Moreover, there was little heterogeneity, across sociodemographic characteristics, in estimated differences in these outcomes when comparing food assistance versus care coordination referral.

This study’s results are consistent with and expand on past literature. A number of studies have found that care coordination programs such as those that use community health workers and care managers can improve diabetes outcomes.^11,14–19 For example, a landmark study of community health workers found important benefits with regard to improving biomarkers of diabetes control.¹⁵ Moreover, an emerging body of evidence suggests that food assistance programs such as food subsidies and in-kind provision of food can be useful for diabetes management.^3,20–24 For instance, recent randomized trials of food subsidy interventions showed that they effectively lowered HbA1c and blood pressure.^24–26 As there has been little comparative effectiveness evidence regarding these approaches, this study provides real world evidence on how these approaches compare in routine practice. One implication of these findings is that clinics considering which approach to offer might decide based on factors other than clinical effectiveness. These could include existing infrastructure, relationships with community-based organizations, funding constraints, or preferences of the patients served.^68–70 Of note, the results of this study are specific to comparing care coordination to food assistance referrals. They should not be interpreted as evaluating the impact of care coordination compared with usual care, or food assistance compared with usual care. Since those comparisons had already been made in prior studies^{11,14–19,24–26}, we did not repeat them here. Based on past literature, it is likely that study outcomes would have worsened in the absence of any intervention.^2,5,8

This study has important limitations which should be considered when interpreting its results. We used an ‘intention-to-treať approach where individuals were analyzed on the basis of receiving a referral. We did not have data on intervention adherence, and so we do not know, for example, whether or how individuals engaged in care coordination or actually received food assistance. If there was non-adherence, and particularly if non-adherence was differential between the two groups, this could affect study results. Secondly, our analysis focused on overall categories of interventions, but there could be variation within them such that similar-seeming interventions vary in effectiveness. For example, there may be more or less effective versions of care coordination programs. However, the approach used in this study allows us to analyze multiple real-world implementations of such programs, thus providing high generalizability for the findings. We view this as a trade-off between external validity, which this study emphasized, and internal validity (as emphasized in prior randomized trials). As a future direction for research in this area, it would be useful to conduct comparative effectiveness randomized clinical trials comparing specific versions of interventions within the categories studied here. For instance, such a trial might compare a community health worker intervention to a food subsidy, or a care manager intervention to a food pantry-based program. Next, we did not have data on educational attainment. Another limitation is that using a new-user study design, as we did, could impact the generalizability of the findings. However, a new-user design does help avoid biases related to not observing the start of the intervention.⁴⁹ Next, this study used electronic health record data, and so could be affected by data missingness or inaccuracies. We would not, however, expect this to be differential between the groups, and we did employ missing data methods to help account for possible impacts of missing data.⁵⁹ Next, it could be that care coordination generated subsequent food assistance referrals. Though we did not see any documented food assistance referrals among those who received care coordination, they may not have been recorded. If this occurred, unrecorded food assistance referrals may have mediated the impact of care coordination, contributing to the similarity in the study outcomes when comparing the two groups. Next, the study sample drew from individuals seen in the OCHIN network, which may not be nationally-representative. Moreover, resources available to individuals can vary from state to state. Next, there could be confounding by indication, since baseline HbA1c was higher for those referred to care coordination than food assistance. However, this was explicitly adjusted for in the TMLE analyses. Next, there could be uneven access to care coordination and food assistance referrals across clinics. Finally, this is an observational study, and so unmeasured confounding could bias study findings. However, we did adjust for a large set of possible confounders, and used analytic approaches to help limit the impact of confounding.

Despite these limitations, we believe the findings of this study are still informative. Because care coordination and food assistance referrals are both commonly used interventions, having data on the similarity of outcomes associated with their use helps clinics make decisions about how and when to implement these approaches.

Conclusions

This observational study did not find meaningful differences in diabetes outcomes between adults with type 2 diabetes who received a food assistance referral, compared with those who received a care coordination referral. As both intervention types have been proven effective in randomized trials, the results of this study suggest that clinics might decide which intervention to offer based on other factors such as feasibility within their specific context and patient preference. Social conditions, such as food insecurity, clearly impact diabetes management, and so having ways to address these issues has become an important part of contemporary diabetes management. This comparative effectiveness study suggests that there may be several comparable options for how to do so.