Introduction

Racial/ethnic disparities exist in healthcare outcomes in the US and are deemed unacceptable as they typically result in worse outcomes for minorities relative to whites⁽¹⁾. Various studies have provided an overview of the nature and extent to which there are racial/ethnic disparities in mental healthcare for minorities^(2,3,4). Other studies have found evidence of statistically significant racial/ethnic disparities in both outpatient mental health treatment^(5,6,7,8,9) and inpatient mental health treatment^{(10,11,12,13, 14, 15)}.

In this study, we investigate whether racial/ethnic disparities exist in US inpatient mental health treatment for adults aged between 18 and 64 using different measurement approaches: descriptive analyses, single-level regression analyses, and three-level regression analyses. We define racial/ethnic disparities as racial/ethnic groups having differing inpatient total charges (hereafter referred to as “charges’‘) associated with US inpatient mental health treatment, after adjusting for other factors, such as demographic, socio-economic, clinical, organizational, and geographical characteristics.

A report from the Institute of Medicine⁽¹⁾ makes a distinction between racial/ethnic differences versus racial/ethnic disparities in the quality of healthcare. According to their report, racial/ethnic differences can be attributed to: (A) access-related factors or clinical needs, patient preferences, and appropriateness of intervention; (B) the operation of healthcare systems and the legal and regulatory climate in which health systems function; and (C) discrimination at the individual, patient-provider level (bias, stereotyping, and uncertainty). Racial/ethnic disparities are assumed in the Institute of Medicine report to be due to only the second and third reasons mentioned above. The Institute of Medicine definition of racial/ethnic disparities would not adjust for access-related factors such as socio-economic status (both individual and per-capita) and geographic location. Like other studies of racial/ethnic disparities, we adjust for access-related factors^(10,12,13). If lower incomes inhibit racial/ethnic minority (hereafter referred to as “minority”) access to mental healthcare, this can result in a difference in charges that should be modeled as part of a study of racial/ethnic disparities. Also, if minorities are less likely to have private insurance than whites, this may limit a minority’s affordability of and/or access to mental health services and would also result in a difference in charges that should be modeled. We also model geographic location to adjust for medical practice variation, and model per-capita income to proxy the quality of the local healthcare system that could contribute to different charges per region.

Prior studies of racial/ethnic disparities have relied on either descriptive analyses or on inferences from a single-level regression model. There are potential difficulties with relying solely on these analyses. In healthcare data, observations are not independent, as the data are sampled from a hierarchy, where individuals receive mental healthcare in treatment centers (micro-units nested within macro-units). When observations are not independent, estimates of regression coefficients are less precise and standard errors may not be valid^(16,17). We contrast descriptive and single-level regression model analyses with a three-level regression model⁽¹⁶⁾ that statistically accounts for the hierarchy inherent in our data: individuals discharged from a hospital after diagnosis and/or treatment for a mental health disorder (hereafter referred to as “discharges’‘) are nested within hospitals, and the hospitals are nested within US counties.

This proposed multi-level regression technique of racial/ethnic disparities can be a useful analytical approach. Zaslavsky and Ayanian⁽¹⁸⁾ argue for a more integrated and comprehensive approach to studying disparities in healthcare, and specifically suggest multi-level regression modeling as a statistical method for describing effects at several levels of aggregation. They mention that multi-level regression models “can be used to model predictors of and unexplained variations in racial disparities, and to profile areas, hospitals, or health plans on the magnitude of disparities (p. 305).” By employing an extensive set of control variables to account for factors other than race/ethnicity that might influence charges, the three-level regression model enables the assessment of racial/ethnic disparities at three different levels of the model (individual, hospital, and county) to obtain a more complete summary of possible racial/ethnic disparities, which is a quantitative analysis of the type suggested by Zaslavsky and Ayanian⁽¹⁸⁾.

A further motivation for the three-level regression model arises from the healthcare literature, suggesting that macro-level factors aside from race/ethnicity can influence the treatment and charges associated with the mental healthcare of an individual. Previous studies have suggested that hospitals with higher quality health outcomes per individual have a higher volume of patients⁽¹⁹⁾. This is referred to as a “volume-quality relationship.” Other research points to volume-quality relationships in mental healthcare; the higher the number of individuals admitted to a hospital for a mental health disorder, the more skilled the physicians become in diagnosing and treating the disorder⁽²⁰⁾. Rice et al. provided an introductory account of multi-level regression models for the health economics literature and described some of the areas of health economics research that may benefit from their use, such as the research of medical practice variation⁽²¹⁾. Medical practice variation can vary by geographic region and can have a substantial impact on healthcare costs and spending with up to 70% of the difference in healthcare spending by region attributed to different clinical care for similar diagnoses^{(22, 23)}. Significant findings of medical practice variation in mental healthcare have been reported for the treatment of major depression⁽²⁴⁾ and mental health hospital utilization⁽²⁵⁾. Another factor that can contribute to differences in charges is the quality of the local mental healthcare system. The National Research Council found that differences in healthcare treatments and outcomes by race/ethnicity could result from minorities living in communities with poorer quality healthcare systems⁽²⁶⁾. If minorities lived in geographic areas with poorer quality healthcare systems as compared to the geographic areas lived in by whites, poorer minority mental health outcomes could arise⁽²⁷⁾.

Race/ethnicity variables are included in each level of our models to assess the net impact of factors such as discrimination and healthcare system characteristics that are difficult to directly measure, after including other factors such as socio-economic status, insurance status, and geographical location. Using the three-level regression model, it is possible to quantify the overall impact of discrimination and healthcare system factors on an individual (by basing inference on race/ethnicity at the discharge-level) and on groups of individuals of the same race/ethnicity (by basing inference on the race/ethnicity variables at the hospital- and county-levels). It is also possible to consider the simultaneous impact of race/ethnicity at the micro-level (race/ethnicity of discharge) and at the macro-level (racial/ethnic composition of the US county) on charges, as well as the impact of each of these measures separately on charges. As our results indicate, these measures of race/ethnicity at different levels of the model have offsetting impacts on charges. We also consider the potential impact of the hospital and county factors discussed above that may also influence the amount of charges: volume-quality relationship, medical practice variation, and quality of local mental healthcare.

Methods

Data

Variables

Statistical Analyses

Results

Descriptive Analyses

Table 1 contains summary statistics for all variables. Frequencies and percentages (based on the total number of discharges, hospitals, or counties depending on the specific variable) are provided for each categorical variable, and means and standard deviations are provided for each continuous variable. Variable names are in bold font, and any categories are in plain font.

From Table 1, 69.18% of all discharges were white, 19.18% were black, 7.83% were Hispanic, 1.26% were Asian/Pacific Islander, 0.25% were Native American, and 2.31% were of other race/ethnicity. The average percentage of white individuals who were discharged from a hospital in 2003 was 73.36%, was 12.86% for black individuals, was 9.36% for Hispanic individuals, was 2.02% for Asian/Pacific Islander individuals, was 0.23% for Native American individuals, and was 2.17% for individuals of other race/ethnicity. The average percentage of white individuals residing in a US county in 2003 was 75.51%, was 8.85% for black individuals, was 10.75% for Hispanic individuals, was 2.98% for Asian/Pacific Islander individuals, was 1.14% for Native American individuals, and was 0.77% for individuals of other race/ethnicity.

Figure 1

Table 1: Summary Statistics for all Variables.

Table 2 contains summary statistics for charges by race/ethnicity. Table 3 contains four discharge-level variables and provides summary statistics by race/ethnicity. In Tables 2 and 3, means and standard deviations are reported for charges for ease of interpretation; however, all ANOVA and post-hoc analyses used the natural logarithm of charges due to the skewness of the distribution of charges.

In Table 2, analysis of variance by racial/ethnic category showed that mean charges varied substantially by race/ethnicity (p < 0.0001), where post-hoc analyses showed that whites had significantly lower mean charges than blacks (p < 0.0001), Hispanics (p < 0.0001), Asians/Pacific Islanders (p < 0.0001), Native Americans (p = 0.008), and other races/ethnicities (p < 0.0001), indicating a racial/ethnic disparity. Median values are also reported where whites had the lowest median charge. For many of the categories in Table 3, minorities had significantly higher mean charges than whites, further indicating potential racial disparities.

Figure 2

Table 2: Summary Statistics for Charges by Race/Ethnicity.

Figure 3

Table 3: Summary Statistics for Mean Charges for Selected Discharge-Level Variables by Race/Ethnicity.

Table 4 contains the income quartiles and expected payer variables and focuses on the percentage of sample and number of discharges by race/ethnicity. Higher percentages of blacks, Hispanics, Native Americans, and other races/ethnicities were in the lowest income quartile, Income Q1, relative to whites and Asians/Pacific Islanders. Higher percentages of minorities were insured via Medicaid relative to whites. Lower percentages of blacks, Hispanics, Native Americans, and other races/ethnicities had private insurance coverage relative to whites and Asians/Pacific Islanders.

Figure 4

Table 4: Percentage Comparisons for Selected Discharge-Level Variables by Race/Ethnicity.

Regression Analyses

Regression model coefficient estimates for all variables in the single- and three-level regression models are provided in Table 5. Findings for racial/ethnic disparities at the discharge-level were similar for the two regression models. Using the single-level regression model, Asians/Pacific Islanders (p = 0.010) had charges that were about 5.7% higher than whites, blacks had charges that were approximately 2.1% lower than whites (p = 0.003), and Hispanics had charges that were about 3.9% lower than whites (p = 0.0001). Using the three-level regression model, charges were on average about 5.3% higher for Asians/Pacific Islanders than whites (p = 0.017) and Hispanics had charges that were on average about 3.5% lower than whites (p = 0.007). The regression coefficient estimate associated with black discharges was significant in the single-level regression model (p = 0.003) but not significant in the three-level regression model, and the findings for Native American and other race/ethnicity discharges were not significant in either model. Other significant discharge-level variables in both regression models were age with a positive relationship, male gender with a negative relationship, expected primary payer (only Medicare, Medicaid, No Charge), admission source (with the exception of Court/Law Enforcement in the three-level model), APR-DRG (with the exception of categories 758 and 760), and APR- DRG severity of illness.

The findings of the two regression models greatly differed at the hospital-level. All hospital race/ethnicity variables, except for the Asian/Pacific Islander racial/ethnic category, were significant in the single-level regression model, with p = 0.047 for blacks, and p < 0.0001 for Hispanics, Native Americans, and other races/ethnicities; albeit with different directions for the coefficient as black and Native American had negative coefficients while Hispanics and other race/ethnicities had positive coefficients. This result differed from the three-level regression model in which all race/ethnicity hospital-level variables were not significant. Also, percentage of mental health discharges, bed size, ownership, and teaching status were all significant in the single-level regression model while all corresponding hospital-level variables of these types were not significant in the three-level regression model (except for small bed size).

Analyses of racial/ethnic disparities were not similar for the two regression models at the county-level. Charges were higher in counties with higher percentages of Hispanics (p < 0.0001), Asians/Pacific Islanders (p < 0.0001), Native Americans (p < 0.0001), and other race/ethnicity (p = 0.0009) in the single-level regression model. In the three-level regression model, charges were higher in counties with higher percentages only for other race/ethnicity (p = 0.029). HPSA codes were significant in the single-level regression model but not in the three-level regression model. Other significant county variables in both regression models included per-capita income and location while US state had more state locations significant in the single-level regression model than the three-level regression model.

Figure 5

Table 5: Single- and Three-Level Regression Results.

Discussion

Various methods for analyzing racial/ethnic disparities in US inpatient mental health treatment yielded conflicting results. At the discharge-level, descriptive analyses suggested racial/ethnic disparities in that whites had lower charges than other minorities, including black and Hispanic discharges. These results agreed with findings from Husani et al.⁽¹³⁾ and Chumney et al.⁽¹⁴⁾, who both found higher charges for minorities relative to white discharges. Olfson et al.⁽¹²⁾ analyzed, among other outcomes, the impact of various factors on charges associated with prompt electroconvulsive therapy (ECT) treatment of recurrent major depression. The authors found using linear multiple regression analysis that factors significantly associated with higher charges included non-white race/ethnicity, older patient age, and Medicare insurance; these findings were consistent with the findings of our single-level analyses in Table 5 but not with our three-level analyses at the hospital- and county-levels.

The descriptive results obtained for Hispanics at the hospital- and county-levels were the opposite of the conclusions drawn from the single-level and three-level regression models at the discharge-level, where at the discharge-level the charges for Hispanics were on average lower than those of whites. Our findings conflict with those of Olfson et al.⁽¹²⁾, and with Horvitz -Lennon et al.⁽¹⁵⁾. Horvitz-Lennon et al.⁽¹⁵⁾ assessed disparities for black and Hispanic adults with schizophrenia using administrative claims data from the Florida Medicaid program. The authors found using generalized estimating equation models that minorities spent similar amounts on psychiatric inpatient spending (conditional on use), but had lower spending on psychotropic drugs, mental health, and all health (0.9-70% lower than whites). Our regression model results at the discharge-level suggest that blacks (in the single-level model) and Hispanics had lower charges than whites, indicating racial/ethnic disparities in the opposite direction as those suggested from the descriptive analyses.

There are a few reasons why we believe our regression findings differ. First, our regression models employed extensive discharge, hospital, and county variables; the three-level regression model also employed multiple random errors that accounted for residual variability of the outcome at each level. We were able to explain more of the variation in charges than in the previous studies. More variables and random errors produce less confounded regression coefficient estimates, allowing the race/ethnicity regression coefficients to more precisely measure the impact of race/ethnicity on charges. Second, for blacks, Hispanics, and Asians/Pacific Islanders, prior literature suggests that after adjusting for observable factors that might affect health outcomes, racial/ethnic disparities could be attributed to unmeasured cultural or “attitudinal’‘ differences^(1,6,35). We could not model these attitudinal factors, so they were omitted effects that potentially biased all of our regression coefficient estimates⁽³⁶⁾. Third, we considered multiple mental health disorders, and not just major depression or schizophrenia as in the studies by Olfson et al.⁽¹²⁾ and Horvitz-Lennon et al.⁽¹⁵⁾.

From the regression analyses, we found differing results regarding racial/ethnic disparities in inpatient mental health treatment at the hospital- and county-levels. Using the single-level regression model, we observed racial/ethnic disparities in that charges were higher in hospitals with higher percentages of Hispanic and other race/ethnicity discharges, and were lower in hospitals with higher percentages of black and Native American discharges. The hospital race/ethnicity findings for Hispanics differ from that observed for Hispanic discharges in the single-level regression model where Hispanics had lower charges. In addition, hospital race/ethnicity findings for Native Americans and other races/ethnicities were significant in the single-level regression model; the discharge-level Native American and other race/ethnicity findings were not significant. The lower charges for blacks at the hospital-level could be explained by statistical discrimination⁽²⁷⁾, and the higher charges for Hispanics at the hospital-level could be explained by language barriers⁽¹⁾. However, none of the hospital-level race/ethnicity categories were significant in the three-level regression model which we believe better accounts for the variability at each level of the model (discharge, hospital, and county), suggesting that the results of the single-level regression model at the hospital-level should be interpreted with caution. The findings from the three-level regression model analysis would not advocate allocating financial resources for the mitigation of discrimination or healthcare system factors such as language barriers.

Using the single-level regression model at the county-level, we found racial/ethnic disparities in that an increase in the number of Hispanics, Asians/Pacific Islanders, and other races/ethnicities resulted in an increase in the amount of charges; whereas an increase in the number of Native Americans resulted in a decrease in the amount of charges. A possible interpretation of these findings is that individuals living in counties with high proportions of Hispanics, Asians/Pacific Islanders, and other races/ethnicities receive more intensive inpatient mental health treatment than individuals in counties with low proportions, possibly due to healthcare system factors such as a lack of availability of and access to preventive mental health services: these specific minorities are unable to receive sufficient preventive care and thus require more intensive inpatient treatment⁽¹⁾. This interpretation assumes that inpatient total charges (accounting cost) is an adequate proxy for inpatient total cost (economic cost). Finkler⁽³⁷⁾ notes that if an analyst desires to proxy inpatient total cost via inpatient total charges, the quality of the proxy will depend on the magnitude of the gap between charges and cost, which depends not only on the magnitude of the gap between accounting cost and economic cost, but also on the loading fees included in charges. Our data was not detailed enough to determine whether charges were an adequate proxy for intensity of treatment, and we cannot make any definitive statements regarding racial/ethnic disparities in intensity and/or quality of US inpatient mental health treatment. Also, as with the hospital-level race/ethnicity findings, the county-level race/ethnicity findings should be interpreted with caution as only the county-level other race/ethnicity category was significant in the three-level regression model; the three-level regression model analysis does not suggest allocating financial resources to increasing the availability of and access to mental health services for minorities.

Comparison of p-values between the two regression models revealed that the regression coefficient estimates associated with blacks and Hispanics were more significant in the single-level regression model than they were in the three-level regression model. The significance of racial/ethnic disparities in the single-level regression model may be over-reported if one believes in the statistical preference for a three-level model. The over-reported significance of regression coefficient estimates was most evident at the hospital-level. An analysis of racial/ethnic disparities based on the hospital measures of race/ethnicity using the single-level regression model would erroneously conclude that there were hospital racial/ethnic disparities, as the nesting of discharges within hospitals within counties was ignored. This finding is important regarding policy, as the three-level regression model would not advocate allocating substantial financial resources to minimizing racial/ethnic disparities at the hospital- and county-levels, as was discussed above.

The three-level regression model further allowed us to model hospital and county factors that could contribute to differences in charges. At the hospital-level, we only found significant evidence of a volume-quality relationship for US inpatient mental health treatment for small bed size but not medium bed size as proxied by the percentage of mental health discharges from a hospital in 2003. Medical practice variation and quality of the local healthcare system were both significant as proxied by the rural/urban variable and per-capita income. Specifically, charges were about 21.3% lower in rural US counties than in urban US counties (p = 0.014).

However, US state variable indicators had mixed results where 8 were significant and 9 were not significant. Also, in the three-level regression model, New Jersey (NJ) was the only significant US state indicator with a positive regression coefficient. While we cannot definitively state why this particular state of the US appears to have higher charges, the literature for medical practice variation identifies the following as potential explanations: variation in finance structures that may create between-doctor differences for economic incentives⁽³⁸⁾; physician characteristics such as age and skill⁽²²⁾; patient characteristics such as preferences for treatment⁽¹⁾; some diseases that are more prone to medical practice variation⁽³⁹⁾; and regional/organizational factors such as hospital processes, norms, standards, and culture⁽¹⁾. Per-capita income was positively correlated with higher charges in the three-level regression model (p = 0.011). For US inpatient mental health treatment, differences in clinical practice by region and the overall affluence of the region can influence charges. These interpretations assume that the chosen variables were good proxies.

Our study is only applicable for US inpatient mental health treatment, with inferences conditional on access to treatment. As the Nationwide Inpatient Sample provided individual information per hospital discharge as opposed to per individual; we could not track multiple hospital admissions for the same individual and consequently could not incorporate any longitudinal information into our analyses. One can also question some of the results of our analyses that assume statistical independence for the statistical analysis, as the same individuals may have been analyzed multiple times as independent cases. We do not believe that this is a major concern of duplication for most of our sample, but the possibility does exist. Furthermore, all individual-level variables considered in this study are not assessed prior to admission; in particular, clinical characteristics such as APR-DRG and APR-DRG severity of illness are as recorded upon discharge and not upon admission to the hospital. This is a limitation of our analyses as it is possible that clinical characteristics could differ between admission and discharge and may be influenced by the same dynamics that lead to disparities. We also did not explicitly control for secondary diagnoses in our analyses, although they were indirectly assessed by APR-DRG severity of illness. Finally, it is possible that omitted variables, such as stigma and other cultural and attitudinal differences, potentially biased our results.

In conclusion, racial/ethnic disparities were found at the discharge-, hospital-, and county-levels using the single-level regression model with offsetting impacts on charges. However, many racial/ethnic disparities were not found at the discharge-level in the three-level regression model with only a disparity of greater charges for Asians/Pacific Islanders, none were found at the hospital-level and only for other race/ethnicity at the county-level. Our three-level regression model findings would not suggest allocating substantial financial resources for the education of physicians to avoid stereotyping, or the mitigation of physician bias or statistical discrimination; they would also not suggest allocating financial resources to increasing the availability of and access to mental health services for minorities. Future research will consider mitigating omitted variables bias, as well as extensions of the model itself.