Social Risk Factors and Racial and Ethnic Disparities in Health Care Resource Utilization Among Medicare Advantage Beneficiaries With Psychiatric Disorders

Abstract

The intersection of social risk and race and ethnicity on mental health care utilization is understudied. This study examined disparities in health care treatment, adjusting for clinical need, among 25,780 Medicare Advantage beneficiaries with a diagnosis of a psychiatric disorder. We assessed contributions to disparities from racial and ethnic differences in the composition and returns of social risk variables. Black and Hispanic beneficiaries had lower rates of mental health outpatient visits than Whites. Assessing composition, Black and Hispanic beneficiaries experienced greater financial, food, and housing insecurity than White beneficiaries, factors associated with greater mental health treatment. Assessing returns, food insecurity was associated with an exacerbation of Hispanic-White disparities. Health care systems need to address the financial, food and housing insecurity of racial and ethnic minority groups with psychiatric disorder. Accounting for racial and ethnic differences in social risk adjustment-based payment reforms has significant implications for provider reimbursement and outcomes.

Introduction

Numerous studies have documented the excess medical comorbidity and mortality in patients with mental illness (Abe-Kim et al., 2007; Hert et al., 2011; Plana-Ripoll et al., 2019). Meanwhile, only 38% of adults with any mental illness received treatment in the previous year (Walker et al., 2015), with even lower rates of access to disorder-specific treatment for individuals living with anxiety disorders (24%) and substance use disorders (19%; Olfson et al., 2019). Rates of access to treatment are even lower than the general population for individuals from racial and ethnic minority backgrounds. Black, Latino, and Asian individuals are less likely to initiate and follow-up with treatment, compared with Whites (Abe-Kim et al., 2007; Cook et al., 2014). Patients with untreated or under-treated mental illness are less likely to start or maintain treatment for a physical condition, potentially resulting in the overuse of more expensive hospital services for acute symptoms (Thornicroft et al., 2017).

Eliminating racial and ethnic disparities in mental health care is important to achieving overarching goals for health equity as set forth in Healthy People (Pronk et al., 2021) and has the potential to improve workplace productivity (de Oliveira et al., 2023) and reduce health care costs (Cook et al., 2015). In patients with mental illness, racial and ethnic disparities in the management of physical comorbidity, for example, co-morbid diabetes, dyslipidemia, and hypertension (Mangurian et al., 2016; Nasrallah et al., 2006) can have further negative consequences in terms of health care expenditures (Mangurian et al., 2016), medical morbidity and mortality (Dickey et al., 2002; Thornicroft, 2013).

To identify modifiable targets for health equity interventions, it is important to understand differences in both the “composition,” or distributions by race and ethnicity, of the allowable and non-allowable covariates of mental health care, and how the “return,” or influence, of these covariates differs by race and ethnicity. Allowable covariate differences can be justified as being outside of the scope or responsibility of the health care system. Non-allowable covariate differences cannot be justified and are differences that health care systems should be held accountable for (Cook et al., 2012; Jackson, 2021; McGuire et al., 2006). As discussed below, it is important not to over-control for contextual factors which themselves reflect and reinforce racial and ethnic inequalities. Among these non-allowable covariates are health-related social needs (HRSNs), which are social conditions associated with higher disease prevalence, poorer disease management, and increased health care costs (Bensken et al., 2021, 2022). Modifiable HRSNs, such as food insecurity, housing instability, and lack of transportation, have increasingly become the target of interventions for health care systems, with the goal of improving patient outcomes (Berkowitz et al., 2019; Engelberg Anderson et al., 2020).

Since late 2019, Humana’s Bold Goal and Population Health Team has conducted social risk factor screenings in a large, nationally representative sample of patients enrolled in Medicare Advantage (MA), from which results were recently published (Long et al., 2022). In 2020, MA beneficiaries represented 46% of all Medicare beneficiaries, and Humana accounted for approximately 18% of all MA enrollees nationwide (Ochieng et al., 2023). Notably, Black and Latino beneficiaries are disproportionately more likely to be enrolled in a MA plan compared with White beneficiaries, and high-rated plans were relatively less available to racial and ethnic minority enrollees than White enrollees (Park et al., 2023). Because of this, MA is a program with potential to narrow or widen health disparities across multiple health domains (Johnston et al., 2022; Park et al., 2021). This study leverages this unique HRSN dataset combined with claims-based clinical and utilization data, to examine how race and ethnicity and social risk factors intersect to influence mental and physical health care resource utilization among those living with mental illness in an important nationally representative sample of MA beneficiaries. We hypothesized that, compared with White individuals with psychiatric disorders, minority individuals (Asian, Black, Hispanic, and Native American) with psychiatric disorders would have higher rates of social risk factors that would explain, in part, significant racial and ethnic disparities in health care utilization.

Conceptual Framework

We use the Institute of Medicine (IOM) disparities definition as defined in Unequal Treatment (Institute of Medicine, 2002) which defines racial and ethnic health care disparities as differences in utilization due to the operation of health care systems and legal and regulatory climate and discrimination but not clinical appropriateness and need and patient preferences (Figure 1). Importantly, the definition incorporates into the disparity calculation both interpersonal racial and ethnic discrimination (e.g., between the provider and patient) and systemic racism (“differences due to the operation of health care systems and legal and regulatory climate”), recognizing that societal policies, institutional practices, and cultural norms that have negative effects on individuals with low socioeconomic status (SES), underinsurance, food, housing, and financial insecurity and other social vulnerabilities, are non-allowable and systematically reinforce and perpetuate racial and ethnic inequity. Another contribution of the IOM committee is that they delineated a conceptual framework for how to measure disparities, implying that disparities analyses should adjust for so-called allowable covariates such as clinical need and appropriateness and patient preferences but should not adjust for non-allowable covariates representing categories (e.g., lower SES groups) systematically and negatively impacted by health care systems and related laws and regulations. In prior work and in the current study, this is operationalized by statistical adjustment of racial and ethnic differences due to allowable clinical need variables (differences that are “justified”) but not adjusting for non-allowable racial and ethnic differences due to SES and other social risk variables (differences in how health care systems treat individuals that are “unjustified”; Alegria et al., 2012; Cook et al., 2009, 2012, 2017; Walker et al., 2015; Zaslavsky et al., 2012). Other methods of measuring racial and ethnic disparities involve assessing the race or ethnicity coefficient, after adjusting for all other available covariates, but these methods adjust away certain differences (e.g., in how the health care needs of individuals with differing social risk statuses are met) that are non-allowable or unjust and should be allowed to contribute to measures of disparities (Cook et al., 2012; Duan et al., 2008; Jackson, 2021). Our dataset is unique in its richness of both clinical need and social risk factors available to serve as proxies of the differential treatment of health care systems. This allows us a unique opportunity to quantify how much of the racial/ethnic disparity is due to health care systems’ differential treatment of individuals with high social risk.

Figure 1.

Institute of Medicine Definition of Health Care Disparity (Institute of Medicine, 2002).

In addition to providing a framework for measuring disparity, the IOM framework provides guidance on how to meaningfully decompose the contributions of different factors (clinical need, the impact of systems, laws and regulations on vulnerable groups, and discrimination) to the overall racial and ethnic disparity calculation. We do so in this study, taking advantage of the unique combination of social risk factor survey responses and MA claims data. We identified the contributions of social risk to utilization disparities in two different ways, adapting the Kitagawa-Oaxaca-Blinder decomposition (Blinder, 1973; Kitagawa, 1955; Oaxaca, 1973): (a) by measuring a difference in “composition” of social risk by race/ethnicity (i.e., a difference in means; identifying differences by race-ethnicity in social risk and then assessing the associations of these social risk with utilization); (b) by measuring a difference in “return” on social risk by race/ethnicity (i.e., a difference in coefficients; identifying how the associations between social risk are exacerbated or diminished among racial/ethnic minority groups compared with the white group).

New Contribution

This is the first study to assess the intersectionality of social risk and race and ethnicity on health care utilization among MA beneficiaries with psychiatric disorder. The unique dataset in this study combines the rich clinical risk data of an insurance claims dataset with extensive social risk survey data (financial strain, food insecurity, housing insecurity, housing quality, loneliness, transportation issues, and utility issues) that allows for careful consideration of variables to be adjusted for (or not) in disparities calculations as well as allowing for a comprehensive assessment of the social risk contributors to the racial and ethnic disparity calculation.

Method

This cross-sectional, retrospective study of MA beneficiaries with a psychiatric diagnosis compares rates of social risk factors between racial-ethnic groups and examines racial-ethnic disparities in utilization using the IOM definition of health care disparity (Institute of Medicine, 2002). We analyzed the Accountable Health Communities (AHC) Health-Related Social Needs (HRSN) screening data (described in more detail below) merged to Humana MA insurance claims data. This study was approved as not human subjects research by the Cambridge Health Alliance Institutional Review Board.

Study Population

Individuals enrolled in Humana MA plans were included in the study based on the following criteria: (a) age 18 through 89 years old; (b) continuous enrollment for 12 months prior to survey completion date (referred hereafter as prior-year); (c) at least one medical claim (inpatient or outpatient) with an International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM) diagnostic code for a Diagnostic and Statistical Manual of Mental Disorders (5th ed.; DSM-5; American Psychiatric Association, 2013) psychiatric disorder during the prior-year; (d) full completion of the adapted AHC HRSN Screening Tool (Billioux et al., 2017) (English or Spanish); and, (e) availability of the patient’s race-ethnicity data. Patients were excluded from the study if (a) enrolled in any plan that is contractually excluded from research or (b) their race-ethnicity was other or unknown.

Race and ethnicity data were used in this study to characterize how the social needs of Humana’s MA population varied by race and ethnicity and to investigate racial and ethnic disparities in health care utilization. Race and ethnicity classifications were based on beneficiaries’ CMS enrollment data, as Asian, Black, Hispanic, Native American, and non-Hispanic White. These CMS enrollment data were derived from source data from the Social Security Administration and are limited in that they are not always accurate reflections of beneficiary-identified race and ethnicity classifications (Office of Inspector General, 2022), they combine race and ethnicity data, and they allow for only one response (Office of Inspector General, 2023). To improve upon the known underrepresentation of Hispanic individuals in Medicare enrollment data (Zaslavsky et al., 2012), beneficiaries who were not classified as Hispanic but completed the Spanish version of the AHC HRSN Screening Tool (Billioux et al., 2017) were re-classified as Hispanic. Beneficiaries were grouped into the following DSM-5¹⁸ diagnostic categories: anxiety (ICD-10-CM F06.4, F40.XXX, F41.X), bipolar and related (F06.33-F06.34, F30.XX-F31.XX, F34), depressive (F06.31-F06.32, F32.XX-F34.1, F34.81), dissociative and conversion (F44.XX), eating (F50.XX), neurocognitive (F01.XX-F05, G30.X), obsessive-compulsive and related (F42.X), schizophrenia-spectrum and other psychotic (F06.0-F06.2, F20.XX-F23, F25.X, F28-F29), sleep (F51.XX), somatoform (F45.XX), substance-related and addictive (F10. XXX-F19.XXX), and trauma- and stressor-related (F43.XX) disorders. Psychiatric diagnostic categories were not mutually exclusive to allow for comorbidity.

Social and Clinical Risk Factors

Social risk factor data were obtained from Humana’s multi-channel (interactive voice response call and/or short message service text and/or email, depending on contact information) AHC HRSN screenings between November 2019 and February 2020 using an adapted version of the AHC HRSN Screening Tool (Billioux et al., 2017) in a sample of patients enrolled in MA. The survey included questions about: financial strain, food insecurity, housing insecurity, housing quality, loneliness, transportation issues, and utility issues. These social risk data were merged to Humana Medicare Advantage enrollment and medical claims data. Enrollment claims data provided sociodemographic and plan information: age, gender, race-ethnicity, geographic region (Northeast, Midwest, South, or West), population density (urban, suburban, rural, or unknown), dual eligibility (both Medicare and Medicaid benefits), special needs plan coverage, low income subsidy status (Medicare beneficiaries with incomes below 150% of the poverty threshold), SES index (0–100 score summarizing area-level SES information covering occupation, income, wealth, education, and housing), and plan coverage start and end dates. Medical claims data provided diagnostic and utilization information on outpatient physician visits, tests and procedures, emergency department (ED) visits, and inpatient hospitalizations.

Allowable clinical risk variables adjusted for in disparity estimates included the CDMF Charlson comorbidity index (a claims-based chart review instrument for categorizing comorbidities based on ICD-10-CM diagnosis codes; Glasheen et al., 2019) and the RxRisk-V score (a pharmacy claims-based comorbidity index based on the identification of distinct medical condition categories using respective medication treatments; Sloan et al., 2003).

Outcome Measures

Prior-year utilization outcomes are from Humana claims data and include inpatient hospitalizations, ED visits, and outpatient visits based on place of treatment, revenue, and/or procedure codes. Utilization measures were further classified as mental or physical health-related; if a psychiatric diagnosis was observed in one of the first three diagnostic levels on a claim, that claim was considered mental health-related.

Statistical Analysis

Descriptive statistics were calculated to assess differences in baseline sociodemographic, clinical and social risk characteristics between racial and ethnic groups, allowing us to assess racial and ethnic differences in composition. Pairwise comparisons versus the referent race (i.e., White) in Table 1 and Figure 2 were conducted using Bonferroni-adjusted alpha levels of 0.0125 (0.05/4).

Table 1.

Racial and Ethnic Group Sociodemographic and Clinical Characteristics.

Variable	White (n = 20,864)	Asian (n = l29)	Black (n = 4,l54)	Hispanic (n = 558)	Native American (n = 75)
Sociodemographic
Age, M [SD], years	69.97 [9.13]	69.37 [9.86]	67.29 [9.3l]***	67.4l [l0.07]***	67.79 [9.65]
Age category, n (%)
18–34 years	28 (0.1)	l (0.8)	l6 (0.4)***	4 (0.7)***	0 (0)
35–49 years	491 (2.4)	5 (3.9)	l58 (3.8)***	30 (5.4)***	3 (4.0)
50–64 years	4,503 (21.6)	24 (l8.6)	l,2l0 (29.l)***	l38 (24.7)	22 (29.3)
65–79 years	12,948 (62.l)	84 (65.l)	2,432 (58.5)***	334 (59.9)	42 (56.0)
80–89 years	2,894 (13.9)	15 (ll.6)	338 (8.l)***	52 (9.3)*	8 (l0.7)
Female, n (%)	13,153 (63.0)	73 (56.6)	2,85l (68.6)***	347 (62.2)	49 (65.3)
Geographic region, n (%)
Northeast	585 (2.8)	3 (2.3)	84 (2.0)*	l7 (3.0)	2 (2.7)
Midwest	4,606 (22.1)	20 (l5.5)	720 (l7.3)***	74 (l3.3)***	7 (9.3)*
South	13,006 (62.3)	55 (42.6)***	3,l6l (76.l)***	356 (63.8)	4l (54.7)
West	2,667 (12.8)	5l (39.5)***	189 (4.5)***	lll (l9.9)***	25 (33.3)***
Population density, n (%)
Urban	11,978 (57.4)	ll0 (85.3)***	3258 (78.4)***	436 (78.l)***	33 (44.0)***
Suburban	5,779 (27.7)	l4 (l0.9)***	607 (l4.6)***	76 (l3.6)***	30 (40.0)***
Rural	2,603 (12.5)	3 (2.3)***	237 (5.7)***	26 (4.7)***	ll (l4.7)***
Unknown	165 (0.8)	2 (l.6)	18 (0.4)	8 (l.4)	0 (0)
Dual eligible, n (%)	4,209 (20.2)	37 (28.7)***	l,697 (40.9)	2l8 (39.l)***	2l (28.0)
Special needs plan, n (%)	1,308 (6.3)	l9 (l4.7)***	752 (l8.l)***	ll3 (20.3)***	4 (5.3)
Low income subsidy, n (%)	1,765 (8.5)	l3 (l0.l)	468 (ll.3)***	67 (l2.0)	5 (6.7)
Area-level SES index, M [SD]	52.81 [1.07]	53.72 [l.47]***	52.33 [0.96]***	52.74 [l.03]	52.68 [0.95]
Clinical
Unique psychiatric diagnosis count, M [SD]	1.84 [1.04]	l.53 [0.83]***	l.72 [0.97]***	l.72 [0.95]*	2.l2 [l.23]
Diagnosis category, n (%)
Anxiety	10,595 (50.8)	5l (39.5)*	l,6l2 (38.8)***	266 (47.7)	37 (49.3)
Bipolar	1,347 (6.5)	4 (3.l)	230 (5.5)	25 (4.5)	8 (l0.7)
Depressive	ll,496 (55.l)	59 (45.7)	2,0l4 (48.5)***	323 (57.9)	4l (54.7)
Dissociative/conversion	ll2 (0.5)	0 (0)	18 (0.4)	l (0.2)	0 (0)
Eating	81 (0.4)	0 (0)	17 (0.4)	2 (0.4)	0 (0)
Neurocognitive	1,348 (6.5)	6 (4.7)	293 (7.l)	29 (5.2)	7 (9.3)
Obsessive-compulsive	133 (0.6)	0 (0)	9 (0.2)*	4 (0.7)	0 (0)
Schizophrenia	474 (2.3)	5 (3.9)	l92 (4.6)***	ll (2.0)	8 (l0.7)***
Sleep	1,728 (8.3)	8 (6.2)	286 (6.9)*	45 (8.l)	8 (l0.7)
Somatoform	238 (l.l)	4 (3.l)	46 (l.l)	l0 (l.8)	l (l.3)
Substance	8701 (41.7)	46 (35.7)	l,974 (47.5)***	207 (37.l)	35 (46.7)
Trauma/ stressor	2,239 (l0.7)	l4 (l0.9)	434 (l0.4)	35 (6.3)***	l4 (l8.7)
Any suicidal ideation or attempt, n (%)	262 (l.3)	2 (l.6)	44 (l.l)	5 (0.9)	0 (0)
CDMF Charlson comorbidity index, M [SD]	4.84 [2.38]	4.7l [2.33]	5.09 [2.59]***	4.93 [2.40]	4.8l [2.50]
RxRisk V, M [SD]	7.02 [3.47]	6.57 [3.08]	7.45 [3.44]***	7.42 [3.20]*	6.8l [4.34]

Note. SD = standard deviation; SES = socioeconomic status; CDMF: Claims-based, Disease-specific refinements, Matching translation to ICD-10, Flexibility to allow use as a chart review instrument.

^*P < .0125.

^**P < .001.

^***P < .0001 vs. White.

Figure 2.

Participant Diagram.

Note. AHC = accountable health communities; HRSN = health-related social need; y = years.

Similar to prior studies (e.g., Alegria et al., 2012; Biener & Zuvekas, 2021; Cook et al., 2017; Creedon & Cook, 2016; Mahmoudi & Jensen, 2012), we assessed disparities using the IOM Unequal Treatment framework (Institute of Medicine, 2002), which defines racial and ethnic health care disparities as all differences in health care use except for differences due to clinical need, clinical appropriateness, and patient preferences. To use this framework, we adjusted for allowable variables related to clinical need (Cook et al., 2012; McGuire et al., 2006) as represented by binary DSM-5 (American Psychiatric Association, 2013) diagnostic categories, binary suicidal ideation, CDMF Charlson comorbidity index score (Glasheen et al., 2019), RxRisk-V score (Sloan et al., 2003), and binary special needs population category. We also adjusted for age, sex, geographic region category, and population density, considering them to be allowable covariates given the lack of modifiable health care system targets available to address differences due to these factors (Jackson, 2021). We did not adjust for other non-allowable covariates related to social risk and other variables highly associated with structural disadvantages faced by racial-ethnic minority groups, including binary dual eligibility for Medicaid and Medicare, binary low-income subsidy, continuous SES index score, and binary social risk categories of financial strain, food insecurity, housing insecurity, housing quality, loneliness, transportation issues, and utility issues.

Implementing the IOM definition of disparity calls for a three-step process that includes model estimation, adjustment of the distributions of variables related to clinical need so they are equivalent across racial-ethnic groups (but not distributions of variables related to social risk), and outcome prediction. In the first step, logit regression models (for any inpatient hospitalization, ED visit, or outpatient visit) and two-part models (logit model for any visit and negative binomial models for number of visits) were estimated, adjusting for race and ethnicity, and all covariates described above. Logit models take the general form of the following equation:

$$\begin{gathered} \mathrm{l}\mathrm{o}\mathrm{g}\mathrm{i}\mathrm{t}E\left[y_i\right]={\beta }_0+{\beta }_1\mathrm{*}\mathit{MinorityGrou}{p}_i \\ +{\beta }_2\mathrm{*}{\mathit{Social\_Risk}}_i+{\beta }_3\mathrm{*}\mathit{Clinical\_Need} \end{gathered}$$ (1)

where $\mathit{MinorityGrou}{p}_i$ represents a vector of indicator variables for Black, Asian, and Native American race, and Hispanic ethnicity, $\mathit{Social}{\mathit{Risk}}_i$ represents “non-allowable” covariates related to social risk and other variables highly associated with structural disadvantages, and ${\mathit{Clinical\_Need}}_i$ represents “allowable” covariates related to clinical need as described above. In the second step, to adjust solely for allowable covariates and to allow racial-ethnic differences in non-allowable social risk and other structural disadvantage variables to enter into the disparity estimation, we applied the rank-and-replace method to adjust for an index of clinical need variables in the model, following methods described in prior studies (Cook et al., 2009; McGuire et al., 2006). In the final step, predicted utilization for each racial and ethnic group was calculated using the adjusted index of clinical need and then disparities were estimated by subtracting predicted adjusted means. To estimate the variability associated with use of services and disparity estimates, we applied the bootstrap method to replicate each step of the disparity estimation process 500 times (Efron & Tibshirani, 1994).

To measure differences in the “return” of social risk covariates, we additionally estimated multiple linear regression models similar to the general form of above Equation 1, but using linear models and adding interaction terms to quantify the influence of interactions of race-ethnicity and social risk on prior-year binary utilization measures, adjusting for all covariates:

$$\begin{gathered} Y_{i\dot{t}}={\mathbf{\beta }}_0+{\mathbf{\beta }}_1\mathrm{*}{\mathit{MinorityGroup}}_i \\ +{\mathit{\beta }}_2\mathrm{*}{\mathit{Social\_Risk}}_i+{\mathit{\beta }}_3\mathrm{*}{\mathit{MinorityGroup}}_i\#\# \\ \mathit{Social\_Risk}+{\mathit{\beta }}_4\mathrm{*}{X}_{it}+e_{it} \end{gathered}$$ (2)

where ${\mathit{MinorityGroup}}_i$ represents a vector of indicator variables for Black race, Hispanic ethnicity, and Asian race, and $X_i$ represents a vector of all other model covariates (clinical need and covariates related to other structural disadvantages). We estimated a total of seven regression models, one for each ${\mathit{Social\_Risk}}_i$ variable (financial strain, food insecurity, housing insecurity, housing quality, loneliness, transportation issues, and utilities issues). Assessing the coefficient on the $\mathit{Social\_Ris}k\#\#{\mathit{MinorityGroup}}_i$ interaction term is similar to identifying the racial and ethnic difference in coefficients (or “returns”) of social risk variables in a Kitigawa-Oaxaca-Blinder (KOB) decomposition. (The KOB decomposition uses fully race-stratified models, thus including all possible interactions with race, a step that is not needed in this analysis). The regression coefficients of these interactions (β₃) were tested for significance and then, for interpretability, transformed back to the percentage-point scale (per 10,000 scale for inpatient hospitalization). Comparisons were made by racial and ethnic group, social risk, and the interaction of race-ethnicity and social risk. All analyses were performed using SAS Enterprise Guide v7.15.

Results

Of the 90,864 beneficiaries who responded to the AHC HRSN screening, 25,780 (28.4%) were eligible for this study. We excluded those without prior-year psychiatric diagnosis, incomplete survey responses, and no prior-year continuous enrollment in MA (Figure 2). The largest share of beneficiaries were classified in the data as White (n = 20,864; 80.9%), followed by Black (n = 4,154; 16.1%), Hispanic (n = 558; 2.2%), Asian (n = 129, 0.5%), and Native American (n = 75; 0.3%).

Sociodemographic and Clinical Characteristics

Racial and ethnic minority groups were significantly different than the White group in most sociodemographic characteristics (Table 1). Notably, Black (40.9%, p < .0001) and Hispanic beneficiaries (39.1%, p < .0001) had significantly higher rates of dual eligibility than White beneficiaries (20.2%). Asian beneficiaries (14.7%, p < .0001), Blacks (18.1%, p < .0001), and Hispanic individuals (20.3%, p < .0001) had significantly higher rates of special needs plan coverage than White beneficiaries (6.3%). Mean area-level SES index scores were significantly lower for Black (52.33, p < .0001) than White beneficiaries (52.81), but higher for Asian beneficiaries (53.72, p < .0001). These area-level SES differences were small, reflecting a standardized index of multiple area-level SES variables with narrow confidence intervals within each racial and ethnic group. Nonetheless, these differences do reflect significant differences by race and ethnicity (all greater than 3.5 standard deviations) in the characteristics of the neighborhoods where survey respondents live.

Regarding clinical variables (Table 1), the Asian (1.53, p = .0004), Black (1.72, p < .0001), and Hispanic (1.72, p = .004) groups had a significantly lower mean number of unique psychiatric diagnoses in the prior year than the White group (1.84). Black beneficiaries had significantly lower rates of anxiety (38.8%, p < .0001) and depressive disorders (48.5%, p < .0001) than White beneficiaries (anxiety 50.8%; depressive 55.1%), but significantly higher rates of substance-related and addictive disorders (Black 47.5% vs. White 41.7%, p < .0001). As reflected in mean RxRisk-V scores, Blacks (7.45, p < .0001) and Hispanic beneficiaries (7.42, p = .007) showed significantly higher comorbidity burden than White beneficiaries (7.02).

Social Risk Factors

Black (78.0%, p < .0001) and Hispanic beneficiaries (71.9%, p < .0001) self-reported significantly higher rates of having at least one social risk factor than Whites (60.6%, Figure 3). Black beneficiaries had significantly higher rates of financial strain (59.7% vs. 45.9%, p < .0001), food insecurity (48.0% vs. 28.3%, p < .0001), housing insecurity (12.1% vs. 8.1%, p < .0001), housing quality (34.4% vs. 21.4%, p < .0001), transportation issues (19.6% vs. 11.0%, p < .0001), and utilities issues (18.1% vs. 10.2%, p < .0001) than White beneficiaries. Similarly, Hispanics beneficiaries had significantly higher rates of financial strain (55.2% vs. 45.9%, p < .0001), food insecurity (40.5% vs. 28.3%, p < .0001), housing insecurity (12.2% vs. 8.1%, p = .0005), housing quality (28.3% vs. 21.4%, p < .0001), loneliness (15.5% vs. 11.7%, p = .0065), and transportation issues (16.1% vs. 11.0%, p = .0001) than White beneficiaries. Asian beneficiaries had significantly higher rates of housing insecurity (18.6% vs. 8.1%, p < .0001) than White beneficiaries. There were no significant differences in rates of specific social risk factors between Native American and White beneficiaries. These racial and ethnic differences in social risk factors are important determinants of racial and ethnic differences in utilization, given that social risk factors were significantly correlated with multiple utilization outcomes (see significant associations between social risk and utilization outcomes in Appendix).

Figure 3.

Rates of Social Risk Factors Across Racial and Ethnic Groups.

*p < .0125. **p < .001. ***p < .0001 vs. White.

IOM-Concordant Disparities in Health care Resource Utilization

Disparities in prior-year utilization were observed in Black and Hispanic beneficiaries when compared with White beneficiaries (Table 2). The following comparisons between the Black and White beneficiary groups were significant at the p < .05 level. Among patients with a prior-year ED visit, Black beneficiaries had .89 fewer ED visits than White beneficiaries (5.15 visits vs. 6.05 visits, respectively). Black beneficiaries had a lower percentage of mental health outpatient visits than White beneficiaries (60.63% vs. 61.57%; disparity = −0.94%); among those with a mental health outpatient visit, Black beneficiaries averaged 4.25 fewer visits than White beneficiaries (30.22 vs. 34.74 visits, respectively). Black beneficiaries had a lower percentage of inpatient hospitalization for physical health conditions than White beneficiaries (18.56% vs. 16.07%; disparity = −2.49%).

Table 2.

Racial and Ethnic Disparities in Prior-Year Mental and Physical Health Care Resource Utilization^a.

Variable	White		Asian				Black				Hispanic				Native American
	%/μ	95% CI	%/μ	95% CI	Disparityb	95% CI	%/μ	95% CI	Dis parityb	95% CI	%/μ	95% CI	Dis parityb	95% CI	%/μ	95% CI	Dis parityb	95% CI
Mental health
Inpatient hospitalization (any; %)c	3.20	[2.6, 3.8]	3.27	[2.2, 4.3]	0.04	[−1.3, 1.3]	3.03	[2.2, 3.8]	−0.17	[−1.1, 0.8]	3.55	[2.2, 4.9]	0.34	[−1.3, 2.0]	4.14	[−1.7, 10.0]	0.9	[−5.0, 6.8]
# (μ) given ≥ inpatient hospitalizationd	1.25	[1.1, 1.4]					1.08	[1.0, 1.2]	−0.16	[−0.4, 0.03]	1.17	[0.8, 1.5]	−0.08	[−2.6, 2.4]	1.84	[0.5, 3.2]	0.6	[−0.7, 1.9]
ED visit (any; %)c	8.17	[7.2, 9.2]	8.2	[6.9, 9.5]	0.05	[−1.6, 1.7]	7.36	[6.1, 8.6]	−0.82	[−2.4, 0.8]	7.06	[5.5, 8.6]	−1.12	[−3.1, 0.9]	10.21	[4.1, 16.3]	2	[−4.1, 8.1]
# (μ) given ≥ ED visitd	6.13	[5.2, 7.0]	6.58	[4.1, 9.1]	0.44	[−2.2, 3.1]	5.18	[4.7, 5.7]	−0.95*	[−1.8, −0.09]	5.59	[1.0, 8.8]	−0.54	[−2.8, 1.7]	12.94	[5.0, 20.9]	6.8	[−2.8, 16.4]
Outpatient visit (any; %)c	61.57	[60.1, 63.1]	61.97	[60.2, 63.8]	0.41	[−2.1, 2.9]	60.3	[58.4, 62.2]	−1.27	[−2.8, 0.2]	57.41	[55.1, 59.7]	−4.18*	[−6.4, −1.9]	61.56	[56.3, 66.8]	−0.02	[−5.7, 5.7]
# (μ) given ≥ outpatient visitd	34.78	[33.9, 35.7]	34.16	[31.6, 36.7]	−0.61	[−4.3, 3.1]	30.25	[29.4, 31.1]	−4.52*	[−5.4, −3.7]	30.63	[29.3, 31.9]	−4.14*	[−5.6, −2.6]	38.84	[25.5, 51.9]	4.07	[−11.3, 19.4]
Physical health
Inpatient hospitalization (any; %)c	18.57	[17.5, 19.6]	18.76	[17.6, 20.0]	0.18	[−1.6, 2.0]	16.1	[14.9, 17.3]	−2.47*	[−3.7, −1.3]	17.27	[15.5, 19.0]	−1.29	[−3.2, 0.6]	18.68	[13.8, 23.6]	0.09	[−5.0, 5.2]
# (μ) given ≥ inpatient hospitalizationd	1.44	[1.4, 1.5]	1.47	[1.3, 1.6]	0.03	[−0.1, 0.2]	1.39	[1.3, 1.5]	−0.04	[−0.09, 0.01]	1.42	[1.3, 1.5]	−0.01	[−0.1, 0.09]	1.44	[1.3, 1.6]	0.0	[−0.2, 0.2]
ED visit (any; %)c	42.65	[41.2, 44.1]	43.96	[42.3, 45.6]	1.34	[−0.8, 3.5]	44.17	[42.7, 45.7]	1.52*	[0.3, 2.7]	44.02	[42.4, 45.6]	1.39	[−0.2, 3.0]	43.82	[39.9, 47.7]	1.12	[−2.9, 5.1]
# (μ) given ≥ ED visitd	2.83	[2.7, 3.0]	2.83	[2.6, 3.1]	−0.002	[−0.3, 0.3]	2.68	[2.6, 2.8]	−0.15	[−0.3, 0.03]	2.75	[2.6, 2.9]	−0.08	[−0.3, 0.2]	3.13	[2.3, 4.0]	0.3	[−0.6, 1.2]
Outpatient visit (any; %)c	93.69	[69.8, 117.6]	90.59	[52.1, 128.4]	−3.06	[−19.1, 13.0]	90.25	[52.1, 128.4]	−3.43	[−17.7, 10.8]	90.28	[52.1, 128.4]	−3.44	[−17.9, 11.0]	94.77	[73.9, 115.7]	l.l	[−3.2, 5.4]
# (μ) given ≥ outpatient visitd	21.89	[21.7, 22.1]	21.82	[21.6, 22.0]	−0.07	[−2.1, 1.9]	18.83	[18.7, 19.0]	−3.06*	[−3.3, −2.8]	19.14	[19.0, 19.3]	−2.75*	[−3.3, −2.2]	25.34	[24.5, 26.2]	3.45	[−1.8, 8.7]

Note. ED = emergency department; SE = standard error; IOM = Institute of Medicine.

^a*p < .05.

^bDisparity = Minority%—White%.

^cIOM-concordant predicted rates and disparities based on logistic regression.

^dIOM-concordant predicted rates and disparities based on negative binomial regression.

The following comparisons between the Hispanic and White beneficiary groups were significant at the p < .05 level. A lower percentage of Hispanic beneficiaries had a prior-year mental health outpatient visit than White beneficiaries (57.28% vs. 61.57%; disparity = −4.29%). Among those with a mental health outpatient visit, Hispanic beneficiaries averaged 4.17 fewer visits than Whites (30.57 vs. 34.74 visits, respectively). A lower percentage of Hispanic beneficiaries had a prior-year physical health outpatient visit than White beneficiaries (98.19% vs. 98.67%; disparity −.49%). Among those with physical health outpatient visits, Black and Hispanic beneficiaries had fewer such visits than White beneficiaries (18.83 vs. 21.89 visits; B-W disparity = −3.06 and 19.09 vs. 21.89 visits; H-W disparity = −2.80) during the prior year. No other outcomes were significantly different between White and racial and ethnic minority beneficiaries.

Interactions of Race-Ethnicity and HRSNs on Health care Resource Utilization

Significant interactions of race and ethnicity and specific social risk factors were observed in rates of mental health-related outpatient visits (Figure 4). The difference-in-difference (DiD) estimate (22.31 per 10,000, p < .05) for the Black × “Utilities Issues” interaction with respect to the rate of a prior-year mental health-related inpatient hospitalization (per 10,000), showed a greater association between utility issues and any mental health-related inpatient hospitalization for Black compared with White beneficiaries. In this case, having issues with utilities was associated with relatively greater mental health inpatient hospitalization for Black beneficiaries. Asian beneficiaries reporting loneliness had a higher probability of a prior-year mental health-related outpatient visit than those without loneliness (Within Difference [WD] 44.94%, p < .001), a greater difference than the difference due to loneliness among White beneficiaries (WD 5.82%, p < .0001; DiD 39.12%, p < .05).

Figure 4.

Difference-in-Difference Estimates of Probabilities of Health care Utilization for Race and Ethnicity by Social Risk Factor.

*p < .05. **p < .001. ***p < .0001.

Note. Only statistically significant difference-in-difference estimates are illustrated in this figure. Other estimates are available upon request. HRSN = health-related social need/social risk factor.

Hispanic × HRSN DiD results show relative increases in any outpatient mental health visit among Hispanic beneficiaries with financial strain and housing insecurity but a relative decrease in any outpatient mental health visit among Hispanic beneficiaries with food insecurity. Specifically, Hispanic beneficiaries reporting financial strain had a higher probability of a prior-year mental health-related outpatient visit than those without financial strain (WD 18.59%, p < .05), a greater difference than the difference due to financial strain among White beneficiaries, WD −0.77% NS; DiD of 19.36%, p < .05. Hispanic beneficiaries reporting food insecurity had a lower probability of a prior-year mental health-related outpatient visit than Hispanic beneficiaries without food insecurity (WD −13.27%, p < .05), a greater difference than the difference due to food insecurity among White beneficiaries (WD 1.22% NS; DiD −14.48%, p < .05). Hispanic beneficiaries reporting housing insecurity had a higher probability of a prior-year mental health-related outpatient visit than those without housing insecurity (WD 21.81%, p < .001), a greater difference than the difference due to housing insecurity among White beneficiaries (WD 1.99% NS; DiD 19.83%, p < .05).

Discussion

This study is among the first to incorporate self-reported social risk factor data to examine racial and ethnic disparities in health care utilization among beneficiaries enrolled in MA with a psychiatric diagnosis. Racial and ethnic minority beneficiaries self-reported higher rates of at least one social risk factor than White beneficiaries, and higher rates of specific social risk factor issues. Disparities in mental and physical health care utilization were observed for Black and Hispanic compared with White beneficiaries, particularly related to outpatient visits.

Decomposition of racial and ethnic disparities in health care utilization found that social risk factors contributed to disparities through differences in composition (differences by race and ethnicity in social risk factors were identified, and those social risk factors are highly associated with utilization) and differences in “returns” on social risk (differences by race and ethnicity in the association between social risk factors and utilization were identified). Specifically, Black and Hispanic beneficiaries had higher rates of nearly all social risk factors compared with White beneficiaries, and these social risk factors were significantly associated with utilization (Appendix); this suggests racial and ethnic social risk differences are significant drivers of Hispanic-White and Black-White disparities in utilization.

Assessing differences in “returns,” we identified significant race-ethnicity by social risk factors interactions in some, but not other outcomes. Relative increases in any mental health outpatient visit among Asian beneficiaries endorsing loneliness and Hispanic beneficiaries with financial strain and housing insecurity could be interpreted as beneficial results indicating that health care systems are accommodating the increased clinical need associated with greater social needs. In other, more concerning results, difficulties with utilities were associated with relative increases in inpatient hospitalization (assuming these hospitalizations could have been prevented) and there were relative decreases in outpatient mental health visits among Hispanic beneficiaries with food insecurity. The latter group may be foregoing mental health services so they can use resources to provide needed food for their families. The association between severity of food insecurity, diet-related comorbidities, and service use seen in prior studies (Jia et al., 2021; Weiser et al., 2013) deserves further investigation.

Black and Hispanic individuals are less likely than White individuals to receive mental health care in community-based settings (Alegria et al., 2012; Dobalian & Rivers, 2008; Manseau & Case, 2014), and more likely to be treated through inpatient services (Dobalian & Rivers, 2008; Substance Abuse and Mental Health Services Administration, 2015). Our findings showed that Black and Hispanic MA beneficiaries with a psychiatric disorder did not use mental health outpatient services as much as White beneficiaries, and that fewer Hispanic beneficiaries with a psychiatric disorder used physical health outpatient services than their White counterparts. Among those who had a physical health outpatient visit, both Black and Hispanic beneficiaries had fewer visits compared with White beneficiaries. Cai and colleagues (2021) reported that racial and ethnic minority individuals, including Black and Hispanic individuals, had lower rates of outpatient service use across numerous physician specialties compared with White individuals; this may be explained by physician shortage and limited specialist referrals, distrust built up over time by a legacy of structural racism in health care systems and provider interactions, and reduced appointment scheduling in Medicaid and uninsured populations.

Racial and ethnic differences observed in utilization may be explained by a multitude of factors, including but not limited to racial and ethnic differences in social risk factors, clinically determined need for treatment, geographical location, and patient preferences (Manseau & Case, 2014; Padgett et al., 1994). Our analyses have an advantage over prior studies in that they assessed the association of social risk on utilization and included those non-allowable covariates in disparity calculations, while adjusting for many allowable covariates related to clinical need, following IOM guidance to implement a “fair” method of measuring disparities. However, similar to prior studies, data were limited by the absence of patient preferences for mental health treatment. Patient preferences should be adjusted for according to the IOM framework, however, adjusting for preferences elicited from survey measures is problematic given that patients are rarely fully informed about treatment options (Cook et al., 2012; Padgett et al., 1994) and because patient preferences are highly correlated with prior experiences of discrimination in the health care system (Sonik et al., 2020) (see Sonik et al., 2020; Progovac et al., 2020 for discussion and critique of considering differences due to patient preferences as allowable).

Our findings on the importance of social risk factors in predicting utilization of treatment for individuals living with mental illness have implications for risk adjustment of quality metrics across value-based payment programs among public and private payers. Adjusting performance scores for social risk factors will more fairly assess and reward providers that serve socially vulnerable populations and decrease incentives for provider organizations to avoid complex patients (Cuellar et al., 2022; Jaffery & Safran, 2021). In fact, social risk adjustment schemes are being considered across private, Medicare and Medicaid plans (Tran, 2020) and have been implemented in Massachusetts and Rhode Island Medicaid managed care plans. Importantly, our findings of differential returns by race and ethnicity on these social risk factors suggest the need to also consider interactions between race and ethnicity and social risk in risk adjustment schemes so as to not further disadvantage racial/ethnic minority patients.

Loneliness increased the probability of a mental health-related outpatient visit among Asians to a greater extent than among Whites. Although this result should be viewed cautiously because of the small sample size of the Asian cohort, it is worth noting that screening for and discussions about loneliness might be a strategy to increase persistently low engagement in mental health service utilization among Asian American individuals (Taylor & Nguyen, 2020). Similarly, utility issues were associated with an increased probability of a mental health-related inpatient hospitalization in Blacks versus Whites. According to the American Council for an Energy-Efficient Economy, Blacks spend 43% more of their income on energy bills than Whites (Drehobl et al., 2020). High energy costs may force individuals or families to use money allocated to medical care or prescriptions to keep utilities running, thereby leading to acute decompensation and subsequent inpatient hospitalization.

Results should be viewed in the context of the study’s limitations. First, Medicare enrollment data has been shown to misclassify race and ethnicity for some beneficiaries (Zaslavsky et al., 2012), particularly for Asian, Hispanic, and Native American beneficiaries. Despite this limitation, we felt it was very important to include non-Black minority groups given the paucity of equity research inclusive of these groups. Nonetheless, the sample sizes for the non-Black minority groups remained relatively small. Second, while the study population was from a large, nationally representative sample of patients enrolled in MA, results may not be generalizable to beneficiaries enrolled in traditional Medicare (TM), other MA plans, or populations with other types of insurance. This is especially important given the overrepresentation of racial and ethnic minority beneficiaries in lower quality MA plans. Investigation of HRSN among TM beneficiaries and those that churn in and out of MA plans is needed. Third, the administration of the AHC HRSN screening tool (Billioux et al., 2017) occurred after the period of time from which utilization data were derived. Although some of the screening questions asked about the respective social risk factors in real time, the duration of time in which the patient actually experienced the social risk factor is unknown due to the lack of serial assessments during the past year. There is evidence to suggest that some social risk factors may indeed be chronic (Nord et al., 2002). Other screening questions had an evaluation period of 12 months, which could have introduced bias given that prior-year utilization itself may have had an impact on an individual’s responses to the survey. For example, high out-of-pocket expenses may have contributed to financial strain. An examination of disparities in utilization post-survey administration would not have been possible due to the onset of the pandemic in March 2020. Fourth, recall bias may have occurred not only because of the recollection periods of the screening questions but also because of symptoms associated with psychiatric disorders (Herbolsheimer et al., 2018; Rhodes & Fung, 2004). Fifth, we conducted a complete case analysis to fully account for these specific social risk factors when investigating utilization disparities across racial-ethnic groups. This approach may have introduced selection bias given the proportion excluded. Finally, the use of administrative claims data is subject to misclassification or missing data.

Among individuals with a psychiatric disorder, social risk factors occur at different rates across racial and ethnic groups and influence racial and ethnic disparities in mental health-related utilization. Further calls to action are necessary to remediate persistent racial and ethnic disparities in the use of health care services for patients with mental illness, and social context must be considered in reforms that incentivize racial, ethnic, and socioeconomic equity. Routine screening and addressing unmet social risk factors are critical because they do affect health care service use and health outcomes. Further research is necessary to explore the causal relationships between social risk and health disparities, and to systematically investigate the real-world effectiveness of potential interventions aimed to alleviate one or more social risk factors.

Appendix.

Association of Social Risk Factors With Health Care Utilization^a,b.

Outcome	Financial strain	Food insecurity	Housing insecurity	Housing quality	Loneliness	Transportation issues	Utilities issues
Mental health
Any inpatient hospitalization	1.69 (1.41–2.02)*	1.75 (1.47–2.08)*	2.05 (1.62–2.59)*	1.34 (1.10–1.62)*	2.91 (2.40–3.54)*	2.53 (2.08–3.09)*	1.41 (1.11–1.80)*
Any ED visit	1.53 (1.37–1.71)*	1.76 (1.58–1.96)*	1.72 (1.46–2.01)*	1.19 (1.05–1.34)*	2.43 (2.13–2.77)*	2.16 (1.89–2.47)*	1.49 (1.28–1.73)*
Any outpatient visit	1.30 (1.24–1.36)*	1.38 (1.31–1.45)*	1.39 (1.28–1.52)*	1.18 (1.11–1.24)*	2.09 (1.93–2.26)*	1.33 (1.24–1.43)*	1.07 (0.99–1.16)
Physical health
Any inpatient hospitalization	1.39 (1.30–1.49)*	1.29 (1.20–1.38)*	1.17 (1.04–1.31)*	1.07 (0.99–1.16)	1.32 (1.19–1.45)*	1.53 (1.39–1.67)*	1.13 (1.03–1.25)*
Any ED visit	1.49 (1.42–1.57)*	1.63 (1.54–1.71)*	1.45 (1.33–1.58)*	1.32 (1.25–1.40)*	1.47 (1.36–1.58)*	1.84 (1.71–1.98)*	1.54 (1.42–1.66)*
Any outpatient visit	0.90 (0.71–1.15)	0.81 (0.63–1.04)	0.59 (0.42–0.84)*	0.81 (0.61–1.06)	0.82 (0.58–1.16)	0.61 (0.44–0.83)*	0.94 (0.65–1.36)

Note. ED = emergency department.

^a*p < .05.

^bOdds ratio (95% confidence interval) from univariate logistic regressions.