Years of evidence have confirmed the importance of socioeconomic factors, often described as social determinants of health, in producing health outcomes.1,2 Socioeconomic factors account for almost half the attributable risk of morbidity and premature mortality in the United States, compared with 30% from patient behaviors and 17% from deficits in clinical care or limited access to care.3 Healthy People 2020 now includes objectives addressing social determinants of health, with specific targets for economic stability, education, neighborhood, and built environment, and social and community context.4
Evidence of direct links between social determinants of health and increased utilization of high-cost health services varies by a medical condition; several studies of low-income patients with diabetes demonstrate the adverse effects of unstable housing and food insecurity. Low-income patients with diabetes are at increased risk of hypoglycemia-related emergency department (ED) visits and hospitalizations in the last week of each month when governmental nutrition assistance often runs out.5,6 This concern takes on increased importance since the proportion of patients with diabetes reporting food insecurity increased by 58% from 2005 to 2014.7 Compared with patients with diabetes and stable housing, those who report unstable housing have much higher odds of a diabetes-related ED visit or hospitalization within the prior 12 months.8 We are not aware of existing studies evaluating the effect of interventions to address social determinants specifically among high-cost, high-need patients with diabetes.
There is no consensus on the optimal interventions to address medical, behavioral, and social determinants of health among high-cost, high-need patients. Some recent evaluations of care coordination interventions have not identified reductions in utilization.9,10 While others have reported significant reductions in ED visit and hospitalization rates in Medicaid populations, these include an uncontrolled pre-post analysis11 and a comparison to patients who declined participation in the intervention,12 both of which are vulnerable to selection bias and therefore difficult to interpret.13,14 These single-institution studies also have limited generalizability, since local interventions rely on referrals or connections to a specific network of partnerships, resources, and services in that particular area.15 Given that the strength of community networks vary from place to place and since outcomes such as ED use by Medicaid beneficiaries can vary 20-fold across different communities,16 local interventions may have entirely different results if implemented in another area or region, particularly if the intervention delivery process is not standardized.
In this analysis, our team evaluated a national care coordination intervention for high-cost, high-need Medicaid beneficiaries with diabetes, implemented by UnitedHealthcare and Optum. The goal of the intervention was to provide “whole-person care,” integrating medical, behavioral and social support at a community level to address access barriers in addition to social determinants of health. Our work is part of the Natural Experiments for Translation in Diabetes 2.0 (NEXT-D2) Network, a research collaboration of 8 academic centers using rigorous quasi-experimental study designs to evaluate naturally occurring experiments in health care policy and practice, with a focus on diabetes-related outcomes.17 We evaluated whether a care coordination intervention to address social determinants of health changed rates of ED visits or inpatient hospitalizations among beneficiaries enrolled in the intervention as compared with a comparator group of similar Medicaid beneficiaries not offered the intervention.
Study Design (Intervention Group Versus Comparator Group)
We used an individual-level interrupted time series (ITS) with a comparator group study design. The intervention group was all Medicaid beneficiaries with diabetes who were successfully enrolled by Optum into the care coordination organization (CCO) intervention, regardless of the amount of contact with the program after initial enrollment. Beneficiaries were eligible for the CCO if they were (1) identified as high-cost, high-need (ie, in the top 5% of Medicaid spend) based on utilization over a rolling 12-month window and predicted to be in the top 5% of spend in the following 12 months by a risk model developed by the health insurer, or (2) admitted to the hospital and flagged to be at high risk for 30-day readmission. Our comparator group was UnitedHealthcare Medicaid beneficiaries with diabetes who were CCO-eligible but never contacted for CCO enrollment because they received care in clinics within the Accountable Care Communities (ACC) program. The ACC program is a less-intensive, practice-level intervention limited to ambulatory clinical care delivery (eg, enabling same-day scheduling, reminding providers to close care gaps by prescribing evidence-based medications, etc.) rather than addressing underlying socioeconomic issues. An evaluation of the ACC program showed no effect on ED visits or hospital utilization for patients enrolled in ACC practices.18 This study was approved by the University of California, Los Angeles Institutional Review Board #16-000276.
We analyzed data from January 2013 to June 2017 for Medicaid beneficiaries age 21 years or above in 15 states that had both CCO and ACC programs (Table 1). We limited our analyses to Medicaid beneficiaries who were enrolled through 1 of 4 specific programs: (1) Temporary Assistance to Needy Families (TANF); (2) Medicaid expansion; the (3) Supplemental Security Income (SSI) program without Medicare; or (4) dual Medicaid/Medicare eligibility (simultaneously enrolled in both a Medicaid plan and a Medicare/Dual Special Needs Plan with the health insurer).
Care Coordination Organization Intervention
The CCO model was developed by UnitedHealthcare and Optum in response to the complexity and multifactorial needs of vulnerable members with multiple chronic conditions. The program includes services provided by behavioral health clinicians, primary care physicians, and community-based social service organizations such as Meals on Wheels. These providers and organizations together represent the “care team” providing whole-person care to CCO enrollees. Optum subcontracts with individual community-based organizations to function as CCOs, based on a successful record of accomplishment of providing services in a specific community. Optum provides each CCO with training and access to a standard electronic care coordination platform. Using a claims-based algorithm, UnitedHealthcare notifies Optum of Medicaid beneficiaries who are eligible for the program, and Optum, in turn, works with the CCOs to perform patient outreach. The CCO typically designates a community health worker (CHW) to enroll the eligible beneficiary, administer a needs assessment survey, and connect the beneficiary with appropriate services based on their reported needs. For example, if the beneficiary reported food insufficiency, the CHW would connect him/her to local, established resources such as food pantries. The CHWs try to make contact with each enrolled member at least once per month to provide further assistance as needed. Approximately 26% of eligible UnitedHealthcare Medicaid beneficiaries were successfully enrolled in the CCO.
Study Months and Care Coordination Organization Enrollment (Real or “Synthetic”)
Because we define study time as months relative to CCO enrollment (defined as month 0), we created a “synthetic” CCO enrollment date for each beneficiary in the ACC comparator group to align study months between enrollees and nonenrollees, using the method described by Harvey et al.19 With the data aligned using this approach, we then examined changes in the time trends of monthly utilization before the CCO enrollment or synthetic enrollment date (the “pre” period, months −15 through −4) versus the “post” period, months 6 through 22. We defined a “transition” period of months −3 through 5, because beneficiaries were identified as CCO-eligible ∼3 months before enrollment and may have interacted with program representatives during this period, and because we estimated it would take 6 months for the program to achieve any changes in ED visits or hospitalizations.
Data included Medicaid beneficiary eligibility and demographics data, insurance claims for medical, pharmacy and laboratory services, practice-level characteristics to identify which practices were in the ACC intervention and when each ACC went online, and CCO program eligibility and enrollment information. We used the CMS Chronic Conditions Data Warehouse (CCW) definition to identify beneficiaries with diabetes, namely any of the following: (1) inpatient billing claim; (2) outpatient billing claims, an A1C value ≥6.5 and/or ≥1 prescription for insulin or an oral antihyperglycemic medication.20 We specified that these criteria must be met within 24 calendar months of the first record of eligibility within the health plan.
We defined all outcomes at the person-month level, meaning that each observation corresponds to utilization by an individual within a given month. We made this decision given the significant number of persons who lose or gain Medicaid coverage during a given year.21 We included all eligible person-months in the main analyses. Our outcome measures included indicators for whether an enrollee had any hospital admissions and any ED visits in the given month. Covariates of interest included an indicator for treatment (CCO) versus comparator (ACC) group and indicator and spline variables that measured changes in an outcome’s time trend (changes in level and slope) in the postperiod and transition periods relative to the preperiod. Indicator variables were used to measure the immediate impact of the CCO on outcomes (the change in the level of the outcome at the start of the given study period). Spline variables were used to measure any gradual changes over time (the change in the slope of the outcome during the given period). Specifically, a continuous variable for a time in months since CCO enrollment [CCO enrollment (real or synthetic) defined as month 0] controlled for the outcome’s linear preperiod “baseline” time trend. The indicator variable for the postperiod (1 for months 6 through 22; 0 otherwise) measured the discontinuity, or immediate change in the level of the outcome, at the start of the postperiod, relative to the level expected based on the preperiod trend. The spline variable for the postperiod (counting months in the postperiod, from 1 to 17; 0 for all months in the preperiod and transition periods) measured changes in an outcome’s slope (ie, the gradual monthly rate of change) from the preperiod to the postperiods. We defined indicator and spline variables to account for changes in the transition period but focused on comparisons between the preperiod and postperiods for reasons described above.
Our estimates of interest in this ITS design were the interactions between the treatment group indicator and the indicator/spline variables for each period.22,23 For example, the difference between treatment versus comparator in the change in level from pre-CCO to post-CCO [difference-in-difference (DID) level] was measured by interacting the treatment indicator with the indicator for the postperiod; the difference between treatment versus comparator in the change in slope from pre-CCO to post-CCO (DID slope) was measure by interacting the treatment group indicator with the postperiod spline variable.
Other covariates include sex, age group, race, language, 17 comorbidity indicators, state-by-year fixed effects, an indicator for whether a state had adopted Medicaid expansion in the given month, and seasonality.
Segmented regression analysis models the difference between treatment and comparator in changes in an outcome’s time trend between the preperiod and postperiods, controlling for the aforementioned covarates.24 Outcome measures were modeled using logistic regression. Each model adjusted for 3-digit zip code–level clustering using cluster-robust Huber-White SEs, to account for the nesting of months within people with zip code.25 We used P-value<0.05 as our threshold for determining statistical significance. In addition to estimating the difference between treatment versus comparator in the changes in each outcome’s time trend (DID level and DID slope estimates), we used the above models to predict a counterfactual estimate of expected utilization at 12 months after enrollment (7 mo after the end of the transition period) in the postperiod based on the preperiod time trend. The counterfactual provides an estimate of the presumed utilization if the preperiod had continued and can be compared with the utilization as modeled in the postperiod. In addition, we estimated 3 sensitivity analyses: (1) restricting the analytic sample to beneficiaries with at least 1 month of data in the preperiod and 1 month of data in the postperiod; (2) including patients who were CCO-eligible but not enrolled in the intervention group (intention-to-treat); and (3) restricting the analytic sample to beneficiaries in the 3 states with ∼70% of the ACC practices (Arizona, Tennessee, Washington).
The person-month demographic and clinical characteristics of the CCO-enrolled intervention sample (n=154,324 person-months) and ACC-assigned comparator sample (n=40,510) at baseline are presented in Table 1. The comparator sample included more person-months from older, Medicaid/Medicare dual-eligible beneficiaries than the intervention sample (Table 1). In the preperiod the intervention and comparator groups had a similar person-month risk of ED visits (22.7% vs. 20.2%) and inpatient hospitalizations (9.9% vs. 9.0%). Characteristics of the intervention and comparator samples stratified by Medicaid program type are presented in Supplementary Table 1 (Supplemental Digital Content 1, http://links.lww.com/MLR/B991, available online).
Among SSI Medicaid beneficiaries, we observed a larger projected decrease in ED visit risk among the intervention sample versus the comparator sample at 12 months postenrollment (Table 2, DID: −6.6%; 95% confidence interval: −11.2%, −2.1%). As shown in Figure 1, changes in both the level (−3.7%, P=0.016) and slope (−0.3%, P=0.02) for ED visits were significantly different for the intervention sample versus the comparator sample. We did not find a difference in hospitalization risk between the intervention and comparator samples of SSI Medicaid beneficiaries.
Among Medicaid expansion beneficiaries, we observed a greater decrease in hospitalization risk among the intervention sample versus the comparator sample at 12 months postenrollment (Table 2, DID: −5.8%; 95% confidence interval: −11.4%, −0.2%). As shown in Figure 2, while differences in the level change did not reach significance (−3.1%, P=0.14), there was a significant difference in the slope change for hospitalizations (−0.2, P=0.03).
We did not observe differences for the other tested comparisons at 12 months postenrollment, including no significant difference in the change of either ED visits or hospitalization for TANF beneficiaries (Fig. 1) and Medicare/Medicaid dual beneficiaries. For Medicare/Medicaid duals (Fig. 2), gradual monthly declines in ED visit and hospitalization risk were about half a percentage point more per month for CCO enrollees than for comparators (ED: DID slope=−0.52%, P=0.032; hospitalization: DID slope=−0.48%, P=0.027). In our sensitivity analyses limited to beneficiaries with at least 1 month of data in the preperiod and 1 month of data in the postperiod, we found similar effect magnitudes for the Medicaid SSI, expansion, and Medicare/Medicaid samples to the main analysis. Analyses among TANF enrollees were limited by low sample size. The results of our other 2 sensitivity analyses were consistent with the overall pattern of results, with the majority showing no statistically significant differences in outcomes between the intervention and comparison populations.
In an evaluation of a CCO intervention for high-cost, high-need Medicaid beneficiaries with diabetes, at 12 months postenrollment we found no evidence of reduced utilization with the CCO program in 6 of 8 comparisons. However, we observed larger reductions in the ED visit risk for beneficiaries receiving Medicaid through the SSI program and in the hospitalization risk for Medicaid expansion enrollees relative to comparators receiving a less-intensive, practice-level intervention. Strengths of this analysis include a multi-state sample with comprehensive administrative data to measure utilization outcomes, as well as a rigorous, quasi-experimental DID design with a comparison group selected to minimize selection bias.
While they need to be replicated in other datasets, our limited findings of decreased ED utilization for Medicaid SSI beneficiaries with diabetes and fewer hospitalizations for high-cost, high-need Medicaid expansion beneficiaries indicate potential benefits of care coordination interventions for these subgroups. To qualify for SSI without Medicare, adults must demonstrate that they have an impairment preventing them from performing “substantial gainful activity” for at least 1 year.26 These patients may have a higher burden of behavioral and social health needs than other Medicaid beneficiaries and often receive care in a poorly coordinated network of EDs, hospitals, psychiatric inpatient facilities, community clinics, jails and substance abuse clinics.27 Analyses of high-cost, high-need Medicaid expansion beneficiaries demonstrate similar complex needs and cross-sector service use.28 The CCO intervention may have helped provide these beneficiaries with a new, single point of initial contact with the health system. There are few studies of care coordination among high-cost, high-need Medicaid expansion beneficiaries. However, intention-to-treat analysis of an intensive care management program among Medicaid SSI beneficiaries in Washington State found no overall effects on either ED or hospital utilization at 24-month follow-up. The investigators did identify an increase in outpatient mental health service use and prescription drug costs, outcomes not measured in the current analysis.29
As above, it should be noted that there was no strong, consistent pattern of evidence across all outcomes and populations that the CCO program significantly reduced high-cost health care events. Our results suggest that identification and enrollment of high-cost, high-need Medicaid beneficiaries into a care coordination program may not always be sufficient to result in a change in ED and hospital utilization. We used a conservative study design with a comparator group who received an active although weaker practice-level intervention and additional follow-up time may be required to assess the full effects of this type of comprehensive program. In fact, interventions based on addressing social needs may potentially increase acute care utilization and costs if supports are difficult to access elsewhere, since any weak link in the continuum of community-partnership building, resource referral, and follow-up may drive patients back to existing patterns of interaction with the larger system. As of yet, there is little consensus on the factors that make social needs interventions more likely to be broadly successful, sustainable, and scalable.15,30
We believe that a future “per-protocol” analysis, evaluating the impact on Medicaid beneficiaries who received the complete CCO intervention as intended, will be an important complementary assessment to the results reported here. Such an analysis will help distinguish between issues of low uptake versus modest intervention effect size, might highlight more or less effective program components for cost savings, and may help quantify the potential impact of the intervention if recruitment and retention efforts are increased.13 In addition, future studies examining patient perspectives on this intervention will be invaluable, since evidence suggests that high-need, high-cost patients who perceive higher levels of trust and respect from CHWs and other team members may be more engaged in changing behavior patterns.15,31–33
An overall limitation of the published literature evaluating interventions that target social determinants is an inability to distinguish between different categories of unmet needs. For example, it is unclear whether connecting high-cost, high-need patients with housing supports, food insecurity or income support would be expected to result in similar effects. Since measured of unmet social needs such as unstable housing, lack of access to nutritious food, and inability to afford basic costs of living are not captured in claims data, it has been challenging to study subgroups of patients who fall into these categories. To address this issue, a working group including UnitedHealthcare and the American Medical Association proposed an additional 23 specific International Classification of Diseases (ICD) codes to the Centers for Medicare and Medicaid Services (CMS) to record different categories of unmet needs, which are currently released for public comment.34 If these codes are approved for use in Medicaid and Medicare, they will be enacted in fall 2020.
Our analysis has several limitations. First, since issues of unmet social need have accumulated over many years and can be slow to change, longer term follow-up may be required to identify the effects of care coordination interventions such as this one. Second, we did not match our samples for baseline utilization, and there may have been unmeasured differences between the intervention and comparison populations. However, the ITS approach allows the intervention and comparison groups to have different intercepts and slopes in the preperiod, so the estimated intervention effects should be consistent as long as the changes in intercepts and slopes from the preperiod to the postperiod would have been the same for both groups in the absence of the intervention.35 Finally, while the CCO intervention targeting high-cost, high-need beneficiaries was more intensive than the practice-level ACC intervention from which our comparison Medicaid beneficiaries were drawn, the latter group still may have received additional intervention services, leading to a null result.
In conclusion, we found that an evaluation of a national care coordination program designed by UnitedHealthcare to reduce utilization among high-cost, high-need Medicaid beneficiaries with diabetes was associated with decreased rates of ED visits and hospitalizations for some Medicaid programs but not others. Our findings are consistent with prior results from well-designed studies in this area. Given the oversized contribution of unmet social need and social determinants on health outcomes, finding effective and scalable approaches to address medical, behavioral, and social issues is essential to improve the performance of the health care system in the United States, particularly for high-cost, high-need patients.36
The authors appreciate the logistical support for this work provided by Carolina Vasquez and Arturo Gongora.
1. World Health Organization (WHO). Social determinants of health. 2012. Available at: www.who.int/social_determinants/en
. Accessed May 16, 2019.
2. Centers for Disease Control and Prevention (CDC). Social determinants of health: know what affects health. 2018. Available at: www.cdc.gov/socialdeterminants/index.htm
. Accessed May 16, 2019.
3. Park H, Roubal AM, Jovaag A, et al. Relative contributions of a set of health factors to selected health outcomes. Am J Prev Med. 2015;49:961–969.
4. Koh HK, Piotrowski JJ, Kumanyika S, et al. Healthy people: a 2020 vision for the social determinants approach. Health Educ Behav. 2011;38:551–557.
5. Seligman HK, Bolger AF, Guzman D, et al. Exhaustion of food budgets at month’s end and hospital admissions for hypoglycemia. Health Aff (Millwood). 2014;33:116–123.
6. Basu S, Berkowitz SA, Seligman H. The monthly cycle of hypoglycemia: an observational claims-based study of emergency room visits, hospital admissions, and costs in a commercially insured population. Med Care. 2017;55:639–645.
7. Walker RJ, Grusnick J, Garacci E, et al. Trends in food insecurity in the USA for individuals with prediabetes, undiagnosed diabetes, and diagnosed diabetes. J Gen Intern Med. 2019;34:33–35.
8. Berkowitz SA, Kalkhoran S, Edwards ST, et al. Unstable housing and diabetes-related emergency department visits and hospitalization: a nationally representative study of safety-net clinic patients. Diabetes Care. 2018;41:933–939.
9. Murphy SME, Hough DE, Sylvia ML, et al. Going beyond clinical care to reduce health care spending: findings from the J-CHiP community-based population health management program evaluation. Med Care. 2018;56:603–609.
10. Kangovi S, Mitra N, Norton L, et al. Effect of community health worker support on clinical outcomes of low-income patients across primary care facilities: a randomized clinical trial. JAMA Intern Med. 2018;178:1635–1643.
11. Wesson D, Kitzman H, Halloran KH, et al. Innovative population health model associated with reduced emergency department use and inpatient hospitalizations. Health Aff (Millwod). 2018;37:543–550.
12. Capp R, Misky GJ, Lindrooth RC, et al. Coordination program reduced acute care use and increased primary care visits among frequent emergency care users. Health Aff (Millwod). 2017;36:1705–1711.
13. Murphy SME, Hough DE, Sylvia ML, et al. Key design considerations when calculating cost savings for population health management programs in an observational setting. Health Serv Res. 2018;53(suppl 1):3107–3124.
14. Heckman JIH, Smith J, Todd P. Characterizing selection bias using experimental data. Econometrica. 1998;66:1017–1098.
15. Beck AF, Cohen AJ, Colvin JD, et al. Perspectives from the Society for Pediatric Research: interventions targeting social needs in pediatric clinical care. Pediatr Res. 2018;84:10–21.
16. Lowe RA, Fu R, Ong ET, et al. Community characteristics affecting emergency department use by Medicaid enrollees. Med Care. 2009;47:15–22.
17. Duru OK, Mangione CM, Rodriguez HP, et al. Introductory overview of the natural experiments for translation in Diabetes 2.0 (NEXT-D2) Network: Examining the Impact of US Health Policies and Practices to Prevent Diabetes and Its Complications. Curr Diab Rep. 2018;18:8.
18. Moin T, Harwood JM, Mangione CM, et al. Trends in costs of care and utilization
for medicaid patients with diabetes in Accountable Care Communities. Med Care. 2020;58:S40–S45.
19. Harvey R, Jankus D, Mosely D, et al. Random assignment of synthetic event dates to unexposed individuals in observational studies: an automated technique using SAS®. Midwest SAS Users Group; 2012.
20. Centers for Medicare & Medicaid Services. Chronic Conditions Data Warehouse. Condition categories; 2019. Available at: www.ccwdata.org/web/guest/condition-categories
. Accessed May 16, 2019.
21. Roberts ET, Pollack CE. Does churning in Medicaid affect health care use? Med Care. 2016;54:483–489.
22. Fretheim A, Soumerai SB, Zhang F, et al. Interrupted time-series analysis yielded an effect estimate concordant with the cluster-randomized controlled trial result. J Clin Epidemiol. 2013;66:883–887.
23. Lagarde M. How to do (or not to do)...assessing the impact of a policy change with routine longitudinal data. Health Policy Plan. 2012;27:76–83.
24. Wagner AK, Soumerai SB, Zhang F, et al. Segmented regression analysis of interrupted time series studies in medication use research. J Clin Pharm Ther. 2002;27:299–309.
25. Cameron AC, Miller DL. A practitioner’s guide to cluster-robust inference. J Hum Resour. 2015;50:317–372.
26. Social Security Administration. Understanding Supplemental Security Income (SSI) and Other Government Programs. 2019. Available at: http://www.ssa.gov/ssi/text-other-ussi.htm
. Accessed February 13, 2020.
27. New Freedom Commission Report: The President’s New Freedom Commission: Recommendations to Transform Mental Health Care in America. 2003. Available at: https://ps.psychiatryonline.org/doi/full/10.1176/appi.ps.54.11.1467
. Accessed February 13, 2020.
28. Shippee ND, Vickery KD. The complex needs of Medicaid expansion enrollees with very low incomes. Issue Brief (Commonw Fund). 2018;2018:1–10.
29. Bell JF, Krupski A, Joesch JM, et al. A randomized controlled trial of intensive care management for disabled Medicaid beneficiaries with high health care costs. Health Serv Res. 2015;50:663–689.
30. Taylor LA, Tan AX, Coyle CE, et al. Leveraging the social determinants of health: what works? PLoS One. 2016;11:e0160217.
31. Capp R, Kelley L, Ellis P, et al. Reasons for frequent emergency department use by medicaid enrollees: a qualitative Study. Acad Emerg Med. 2016;23:476–481.
32. Mautner DB, Pang H, Brenner JC, et al. Generating hypotheses about care needs of high utilizers: lessons from patient interviews. Popul Health Manag. 2013;16(suppl 1):S26–S33.
33. Berkowitz SA, Parashuram S, Rowan K, et al. Association of a care coordination model with health care costs and utilization
: The Johns Hopkins Community Health Partnership (J-CHiP). JAMA Netw Open. 2018;1:e184273.
34. Robeznieks A. New ICD-10 codes will help physicians tackle social barriers to care; 2019. Available at: www.ama-assn.org/practice-management/digital/new-icd-10-codes-will-help-physicians-tackle-social-barriers-care
. Accessed May 16, 2019.
35. Mora R, Reggio I. Treatment effect identification using alternative parallel assumptions. Universidad Carlos III de Madrid Working Papers. Madrid, Spain; 2012.
36. Schoen C, Davis K, How SK, et al. U.S. health system performance: a national scorecard. Health Affairs. 2006;25:w457–w475.