Abstract
Policy Points.
Social determinants of health are an important predictor of future health care costs.
Medicaid must partner with other sectors to address the underlying causes of its beneficiaries’ poor health and high health care spending.
Context
Social determinants of health are an important predictor of future health care costs but little is known about their impact on Medicaid spending. This study analyzes the role of social determinants of health (SDH) in predicting future health care costs for adult Medicaid beneficiaries with similar past morbidity burdens and past costs.
Methods
We enrolled into a prospective cohort study 8,892 adult Medicaid beneficiaries who presented for treatment at an emergency department or clinic affiliated with two hospitals in Washington, DC, between September 2017 and December 31, 2018. We used SDH information measured at enrollment to categorize our participants into four social risk classes of increasing severity. We used Medicaid claims for a 2‐year period; 12 months pre‐ and post‐study enrollment to measure past and future morbidity burden according to the Adjusted Clinical Groups system. We also used the Medicaid claims data to characterize total annual Medicaid costs one year prior to and one year after study enrollment.
Results
The 8,892 participants were primarily female (66%) and Black (91%). For persons with similar past morbidity burdens and past costs (p < 0.01), the future morbidity burden was significantly higher in the upper two social risk classes (1.15 and 2.04, respectively) compared with the lowest one. Mean future health care spending was significantly higher in the upper social risk classes compared with the lowest one ($2,713, $11,010, and $17,710, respectively) and remained significantly higher for the two highest social risk classes ($1,426 and $3,581, respectively), given past morbidity burden and past costs (p < 0.01). When we controlled for future morbidity burden (measured concurrently with future costs), social risk class was no longer a significant predictor of future health care costs.
Conclusions
SDH are statistically significant predictors of future morbidity burden and future costs controlling for past morbidity burden and past costs. Further research is needed to determine whether current payment systems adequately account for differences in the care needs of highly medically and socially complex patients.
Keywords: social determinants of health, health care costs, Medicaid, disease severity
Despite their access to medical care and comprehensive coverage, the Medicaid population exhibits large health inequalities that are primarily due to factors outside the health care system and that significantly impact health, namely, social determinants of health (SDH). 1 , 2 Most risk adjustment models (RAMs) that Medicaid agencies currently use to reimburse managed care and accountable care organizations for the care they provide to Medicaid beneficiaries are based on medical risk, namely, medical diagnoses, age, and sex, but not SDH measures.
Several studies have evaluated whether the addition of social risk factors to traditional RAMs better predicts health care costs, with inconsistent results. 3 , 4 , 5 , 6 For example, two studies found that SDH measured at the neighborhood level did not improve the prediction of health care costs after controlling for prior medical risks. 4 , 6 Ash and colleagues compared concurrent prediction models based on medical risks measured in the same time period as the health care costs and also found that the addition of individual‐level SDH factors did not improve the overall prediction of costs. They did note, however, that individual SDH were important to identifying underestimated costs for specific vulnerable groups, such as those with housing instability or living in a high economically stressed neighborhood. 5 Finally, Fitzpatrick and colleagues used SDH factors measured at both the individual and the neighborhood level to predict high‐cost health care users. After accounting for past disease severity, numerous individual (e.g., food insecurity, not owning a home) and neighborhood (e.g., residential instability, material deprivation) SDH factors increased the odds of becoming a high‐cost health care user during a five‐year follow‐up period. 3 The inconsistent results of past studies were likely influenced by SDH measurement issues such as which SDH were measured, whether the individual factors were considered as individual or aggregated predictors, and whether the SDH were measured at the individual and/or the neighborhood level. The inconsistency of past study results could also have been caused by whether the models were predicting costs in the same time period that the medical risks were measured (concurrently) or in the future (prospectively).
The purpose of our study was to evaluate the relationship between adverse social and economic circumstances measured at the individual level and future health care costs, as well as the extent to which SDH predict future disease burdens above and beyond past comorbidity burdens. We hypothesized that Medicaid beneficiaries with more social and economic adversities would have higher future annual Medicaid expenditures and higher morbidity burdens even after controlling for their past morbidity burdens and past health care expenditures.
Methods
Overview
We conducted a prospective cohort study of adult Medicaid beneficiaries to evaluate the relationship between their SDH and health care costs. We enrolled adult Medicaid members at the time of a health care visit and administered a SDH survey. We obtained a copy of their Medicaid claims data and merged it with their SDH information. We measured participants’ disease burden one year prior to study enrollment and one year post study enrollment, in accordance with the Adjusted Clinical Groups (ACG) system, using the Medicaid claims data for those periods. We also used the Medicaid claims to calculate the participants’ health care costs one year pre‐ and post‐study enrollment. We modeled future morbidity burdens and future health care costs separately as a function of SDH, age, sex, past morbidity burdens, and past costs. Finally, it is well known that adding other outcomes, like future morbidity burden, to the predictor side of a regression model risks obscuring the primary relationships of interest, in this case the relationship between SDH and future health care costs. Nonetheless, earlier studies have done exactly this and have found that SDH no longer predict cost. 5 To show the risks of this approach, we added future comorbidity to another model, anticipating that the effect of SDH would be diluted, since our conceptual model also predicted SDH to be a cause of future comorbidity.
Study Setting and Sample
During a 16‐month enrollment period, research assistants (RAs) prospectively screened 17,719 adult beneficiaries, identified 12,346 as eligible for the study, and enrolled 8,943 in the Medicaid Cohort of the District of Columbia (MCDC) study when they came for treatment at an emergency department or a clinic affiliated with one of two Washington, DC, hospitals between September 2017 and December 31, 2018. 7
To be eligible for the study, patients had to be between the ages of 18 and 64, had to be insured by the DC Medicaid program, and have access to a telephone. Patients were excluded if they were unable to understand consent, did not speak English, were too sick (i.e., triage acuity level 1), or were also insured by Medicare. The most common reason for the 5,373 who were ineligible (85%) was that they already had been enrolled in MCDC during a previous health care visit. Of the 3,403 eligible patients that the RAs did not enroll, 55% refused, 24% reported feeling too uncomfortable, and 21% completed their medical visit before the RA was able to approach them.
The 8,943 participants who enrolled in MCDC completed a SDH interview and gave us permission to obtain a copy of their DC Medicaid claims data for two years pre‐ and post‐study enrollment. We analyzed 8,892 participants, excluding 51 participants because they were not covered by DC Medicaid during at least one month of the one‐year post‐study enrollment period. This study was approved by the Institutional Review Board of George Washington University.
Social Risks
Our SDH assessment was based on the World Health Organization's conceptual model, which differentiates between the structural determinants of health inequities and intermediary SDH. The World Health Organization (WHO) model theorizes that structural determinants of health such as education, income, and occupation define a person's socioeconomic status (SES) in society. In turn, SES position influences intermediary social determinants, including material circumstances, health behaviors, and psychological factors. Our 35‐item survey consisted largely of validated questions and scales. We asked participants about the following structural determinants of health inequities: sexual orientation, highest level of education achieved, employment status, and employment duration. In addition, we asked them about intermediary determinants, including material circumstances (e.g., food insecurity, financial strain, housing instability), health behaviors (e.g., smoking, alcohol use), and psychological factors (e.g., marital status, loneliness). 7
Because SDH variables are correlated with one another, considering all SDH factors simultaneously in a regression model may miss important associations. To prevent this, we used latent class analysis (LCA) to identify four unique subgroups of participants who had similar response profiles to the SDH assessment variables. Group selection was based on quality of fit, classification diagnostics, and ease of interpretation. 7
Class 1 members (44%) reported the fewest social adversities and were most likely to be employed; we refer to them as “employed and fewest social risks.” Class 2 participants (20%) also had a high employment rate but reported financial strain (e.g., difficulty paying bills); we refer to them as “employed and high financial strain.” Class 3 (27%) consisted largely of participants who were unemployed with limited access to the internet at home and to a car for medical appointments; we labeled them as “unemployed and limited internet and car access.” Finally, class 4 participants (9%) reported the most social adversities; we refer to this class as “unemployed and many social risks.”
Medical Risks
We measured past and future morbidity burden according to the Johns Hopkins ACG® system (version 12.0). The ACG system uses administrative billing data (ICD‐10 diagnosis and pharmacy codes) over a defined period as well as age and sex to classify disease burden. 8 The ACG system groups diseases based on clinical factors (likelihood or persistence of a condition, severity of the condition, diagnostic certainty, etiology, and types of health care services required). 9 Persons with the same ACG category are considered to have a similar morbidity burden and are assigned the same relative risk score. In this study, we used the Medicaid claims that occurred up to one year prior to study enrollment to calculate the past morbidity ACG unscaled risk score; we used claims that occurred during the one‐year post‐study enrollment period to calculate the future morbidity ACG unscaled risk score. Risk scores greater than 1 mean that the study sample is sicker than the reference population, a nationally representative, commercially insured population. 10
Total Health Care Costs
We calculated the total annual health care costs for a one‐year period before study enrollment (past costs) as well as a one‐year period afterward, starting the day of study enrollment (future costs), using the adjudicated Medicaid claims. For participants enrolled continuously during the one‐year follow‐up period, we summed the allowed amounts for all claims paid by the Medicaid program for each participant over the 12‐month period. For participants not enrolled for the entire 12‐month period (8%), we imputed their costs for the missing months based on two years of Medicaid claims data (one year pre‐ and one year post‐study enrollment) using a linear, mixed‐effect model that included time (month), social risk class, ACG concurrent risk score, age, and sex (see technical appendix for details of the model).
Statistical Analysis
The main outcome for the study was future annual Medicaid costs during the one‐year follow‐up period. The main independent variables of interest were SDH, measured both in aggregate by social risk class and individually by separate SDH factors. The main covariates to be controlled for were morbidity risk, measured by the past and future ACG risk scores, and past expenditures. The conceptual framework we used to guide our statistical analysis is illustrated in Figure 1. In brief, there is a strong mechanism‐based correlation between morbidity burden and costs measured in the same time period (see technical appendix, part A). We hypothesized that even among persons with similar past morbidity burdens and past costs, SDH is an important predictor of both future morbidity burden and future health care costs. We also posited that SDH will no longer be a significant predictor of future health care costs among persons with similar future morbidity burdens.
Figure 1.
Conceptual Framework of Relationships Among Social Determinants of Health, Morbidity Burden, and Health Care Costs
The sample size for our study was estimated a priori based on the prevalence rates of four common chronic conditions among adult Medicaid beneficiaries, Medicaid cost data from a previous study, a power of 80%, and an alpha level of 0.05. Assuming prevalence rates of 10% to 20% for different social risk factors, we estimated a target sample size of 9,600 participants to detect cost differences of $2,500 or greater among those both with and without different social risk factors.
First, we examined the distribution of annual health care expenditures and noted that total costs were highly right‐skewed, with variances roughly proportional to the mean (see technical appendix for details).
Second, we assessed associations among social risk class, future annual Medicaid costs, and future morbidity burden. We used analysis of variance (ANOVA) to estimate the differences in mean morbidity burden and annual Medicaid costs among the different social risk classes, and a scatterplot and Pearson correlation coefficient to evaluate the relationship between morbidity burden and annual Medicaid expenditures.
Third, using linear regression, we modeled the future morbidity ACG score as a function of social risk class, past morbidity risk, age, and sex. We also ran two additional models, one that adjusted for past costs and one that adjusted for past morbidity burden and past costs.
Fourth, we employed heteroskedastic linear regression fitted by iterative weighted least squares (IWLS) to estimate the mean future annual Medicaid cost as a function of the predictor variables while accounting for the dependence of the variance of cost on mean cost. 11 To simplify the interpretation of the regression coefficients and to avoid transforming back to the original scale, we directly modeled cost rather than log‐transformed cost. We included an age and sex interaction effect in all the multivariable linear regression cost models based on findings from prior work. 5 , 12
To estimate the total effect of SDH on future annual expenditures, we modeled them as a function of the four SDH classes (model 1). Then we added past morbidity burden (model 2) to determine how much our estimates of SDH attributable costs would change when we controlled for past morbidity burden. Next, we replaced past morbidity burden with past costs (model 3) to compare the difference in SDH attributable costs if we relied on past costs versus past morbidity burden. Subsequently, we modeled future annual Medicaid costs as a function of social risk class, past morbidity burden, and past costs to determine if social risk class was a significant predictor of future Medicaid costs after controlling for both past morbidity burden and past costs (model 4). Finally, we modeled annual Medicaid costs as a function of social risk class and future morbidity burden (model 5) because we anticipated that the effect of SDH on future costs would be lessened by the presence of future morbidity burden as proposed by our conceptual model.
We also compared the impact of individual SDH characteristics on annual Medicaid costs during the one‐year follow‐up period by running multivariable linear regression models including an individual SDH factor only, as well as those including an individual SHD factor, past morbidity burden, and past costs.
Sensitivity Analyses
We repeated the main analysis of annual costs by social risk class, but instead of imputing the costs for participants with missing months of costs, we assumed that the costs for those missing months equaled zero. Second, we repeated the main analysis and top‐coded (i.e., Winsorized costs) annual costs that exceeded $200,000 (1.6% of the sample) at $200,000 so that we could determine whether our main results might have been distorted by a small number of large values (outliers). Third, instead of the ACG risk score, we repeated the main analysis of future annual Medicaid costs using the Chronic Disability Payment Scale, another risk adjustment measure developed specifically for the Medicaid population, in order to determine whether the influence of SDH on future health care costs differed significantly depending on the risk adjustment measure used. Fourth, we repeated the main analysis using a generalized linear regression model assuming a gamma distribution for the cost outcome and a log‐link to determine whether the results we obtained would differ using a model with more specific distributional assumptions.
Finally, we estimated the odds of becoming a high‐cost health care user during the one‐year follow‐up period by social risk class with and without adjustment for past morbidity burden, past costs, and future morbidity burden using a multivariate logistic regression model and following the same approach as we did with the individual SDH factors. To do this, we excluded the top 5% of our participants (n = 445) who were the highest‐cost users during the year prior to study enrollment. Then we determined the top 5% of our participants who were the highest‐cost health care users during the one‐year follow‐up period based on our sample's annual spending distribution so that we could evaluate the impact of SDH on future high‐cost use among participants who were not current high‐cost users. 3
All analyses were completed using R version 4.0.1 (http://www.r‐project.org/).
Results
The 8,892 participants were primarily female and black (see Table 1), with a mean age of 39. The majority were covered by the DC Medicaid program for the entire 12 months after study enrollment (93%). On average, the participants were sicker than the commercially insured reference population (median = 2.9, mean = 6.7, range 0 to 93). The average annual health care expenditures of our cohort was $22,404.
Table 1.
Descriptive Characteristics of Study Sample by Social Risk Class
| Social Risk Class | |||||
|---|---|---|---|---|---|
| Characteristic | Overall (n = 8,892) | Lowest Risk (n = 3,928) | Lower Intermediary (n = 1,792) | Upper Intermediary (n = 2,408) | Highest Risk (n = 764) |
| A. Demographic | |||||
| Mean age (SD) a | 38.8 (13.3) | 35 (12.3) | 38.4 (12.3) | 43.8 (13.9) | 43 (12.6) |
| Female a | 5,917 (67%) | 2,887 (74%) | 1,328 (74%) | 1,294 (54%) | 408 (53%) |
| Black a | 8,047 (91%) | 3,541 (90%) | 1,623 (91%) | 2,213 (92%) | 670 (88%) |
| Mean (SD) number of months insured by DC Medicaid program b | 11.6 (1.7) | 11.6 (1.7) | 11.7 (1.5) | 11.5 (1.8) | 11.5 (1.9) |
| Covered by managed care organization a | 6,720 (76%) | 3,424 (87%) | 1,425 (80%) | 1,438 (60%) | 433 (57%) |
| B. Morbidity burden risk score c , d | |||||
| Mean past score (SD) a | 5.0 (8.4) | 3.2 (6.3) | 4.0 (6.4) | 7.4 (10.2) | 9.0 (11.5) |
| Mean future score (SD) a | 6.7 (10.3) | 4.4 (7.3) | 5.7 (8.3) | 9.8 (12.9) | 11.8 (13.3) |
| C. Annual mean (SD) health care costs d | |||||
| Overall a | 22,404 (48,188) | 14,393 (36,422) | 18,633 (37,938) | 32,931 (61,053) | 39,259 (63,997) |
| Females aged 18‐24 (n = 1,094) | 12,503 (40,548) | 11,206 (40,414) | 10,594 (14,792) | 18,686 (57,002) | 12,632 (12,557) |
| Males aged 18‐24 (n = 322) | 10,025 (26,549) | 7,537 (17,273) | 10,078 (35,014) | 14,449 (37,319) | 14,075 (14,438) |
| Females aged 25‐44 (n = 3,031) a | 14,273 (31,038) | 11,291 (25,810) | 14,637 (30,638) | 19,314 (39,196) | 24,982 (43,228) |
| Males aged 25‐44 (n = 1,202) a | 18,761 (43,656) | 12,787 (39,931) | 18,216 (39,233) | 20,651 (40,024) | 37,482 (63,861) |
| Females aged 45+ (n = 1,792) a | 35,125 (61,510) | 24,803 (48,007) | 26,321 (48,098) | 45,274 (71,650) | 54,749 (78,772) |
| Males aged 45+ (n = 1,451) a | 36,908 (62,533) | 22,288 (43,155) | 28,122 (48,409) | 43,801 (70,260) | 48,429 (70,511) |
p < 0.05.
During the 1‐year post study enrollment period.
Measured by the Adjusted Clinical Groups (ACG) system score during the 1‐year period prior to study enrollment (past score) or the 1‐year period post study enrollment (future score).
Based on observed claims data.
These average annual costs increased from $14,393 for the lowest social risk class to $39,259 for the highest one (p < .0001) (see Figure 2A). We also observed a statistically significant relationship between social risk class and the future morbidity burden score (see Figure 2B). Participants in the lowest social risk class had a mean future morbidity burden score of 4.4, compared with 11.8 in the highest social risk class (p < .0001). The future morbidity burden was also highly positively correlated with annual future health care costs (see Figure 2C) (r = 0.78, p < 0.001).
Figure 2.
Bivariate Relationships Among Social Risk Class, Future Annual Medicaid Costs and Future Morbidity Burden [Colour figure can be viewed at wileyonlinelibrary.com]
Table 2 shows the models predicting future morbidity burden and future annual health care costs by social risk class. After adjusting for past morbidity burden, age, and sex, the mean future morbidity burden score was significantly higher in social risk classes 3 and 4 (1.26; 95% CI: 0.87 to 1.65 and 2.17; 95% CI: 1.58 to 2.76, respectively) compared with the lowest class. The results were similar when we adjusted for past costs: the two higher social risk classes had significantly higher morbidity burden scores compared with the lowest social risk class (see Table 2). Even after adjusting for both past morbidity burden and past costs, the future morbidity burden remained significantly higher for social risk classes 3 and 4 (1.15; 95% CI: 0.76, 1.54 and 2.04; 95% CI: 1.45, 2.63, respectively) compared with the lowest one.
Table 2.
Future Morbidity Burden and Future Annual Medicaid Costs (95% CI) by Social Risk Class for 8,892 Medicaid Participants
| Social Risk Class | ||||
|---|---|---|---|---|
| Lowest Risk (n = 3,928) | Lower Intermediary (n = 1,792) | Upper Intermediary (n = 2,408) | Highest Risk (n = 764) | |
| A. Future Morbidity Burden | ||||
| Total predicted morbidity burden risk score a | 4.36 (4.28, 4.44) | 5.71 (5.58, 5.83) | 9.76 (9.64, 9.87) | 11.78 (11.58, 11.99) |
| SDH attributable severity | ||||
| SDH only b | reference | 0.75 (0.22, 1.28) | 3.55 (3.04, 4.06) | 5.76 (5.02, 6.5) |
| SDH + past morbidity burden c | reference | 0.41 (0, 0.82) | 1.26 (0.87, 1.65) | 2.17 (1.58, 2.76) |
| SDH + past costs d | reference | 0.36 (‐0.09, 0.81) | 1.49 (1.06, 1.92) | 2.72 (2.09, 3.35) |
| SDH + past morbidity burden + past costs e | reference | 0.39 (‐0.02, 0.8) | 1.15 (0.76, 1.54) | 2.04 (1.45, 2.63) |
| B. Future Annual Costs | ||||
| Total predicted costs a | 14,906 (14682, 15130) | 19,156 (18800, 19512) | 30,882 (30558, 31206) | 36,961 (36394, 37528) |
| SDH attributable costs | ||||
| SDH only b | reference | 2,713 (678, 4748) | 11,010 (8660, 13360) | 17,710 (13258, 22162) |
| SDH + past morbidity burden c | reference | 786 (‐260, 1,831) | 2,938 (1699, 4176) | 5,909 (3408, 8410) |
| SDH + past costs d | reference | 650 (‐404, 1705) | 1,611 (467, 2756) | 4,282 (2008, 6555) |
| SDH + past morbidity burden + past costs e | reference | 713 (‐245, 1670) | 1,426 (377, 2475) | 3,581 (1415, 5747) |
| SDH + past morbidity burden + past costs + future morbidity f | reference | 4 (‐206, 215) | ‐33 (‐245, 179) | ‐50 (‐475, 376) |
Predicted values from a regression model that includes social risk class and age and sex interaction effect (model 1).
Regression coefficients estimated from a model that includes social risk class and age and sex interaction effect (model 1).
Regression coefficients estimated from a model that includes social risk class, age and sex interaction effect, and past morbidity burden (model 2).
Regression coefficients estimated from a model that includes social risk class, age and sex interaction effect, and past annual costs (model 3).
Regression coefficients estimated from a model that includes social risk class, age and sex interaction effect, past morbidity burden, and past annual costs (model 4).
Regression coefficients estimated from a model that includes social risk class, age and sex interaction effect, past morbidity burden, past annual costs, and future morbidity burden (model 5).
The mean future health care spending by participants in all of the higher social risk classes was significantly higher on average compared with the spending of the participants in the lowest social risk class ($2,713, $11,010, and $17,710, respectively) (see Table 2). Holding past morbidity burden and past costs constant, the costs for those in the highest two social risk categories were higher by $1,426 and $3,581 relative to those of the lowest social risk group. Finally, when we modeled future annual costs as a function of social risk class and future morbidity burden, adjusted for past morbidity burden and past costs, social risk class was no longer a significant predictor of future costs.
When we considered the relationship between individual SDH characteristics and future annual health care costs, we observed that the majority of the individual SDH factors were significant predictors of future annual Medicaid expenditures (e.g., less education, unemployment, abusive relationship, loneliness, food insecurity) when we did not control for past morbidity burden and past costs (see Table 3). If we did adjust for past morbidity burden and past costs, the influence of individual SDH factors would be reduced but still be meaningfully large (i.e., > $1,000 annually) and statistically significant (p ≤ 0.05) for several SDH factors (e.g., not married, limited or no internet access at home, not living in a house or apartment, worried about unstable housing).
Table 3.
Annual Cost Estimates of Individual Social Determinants of Health With and Without Adjustment for Past Morbidity Burden and Past Costs
| Social Determinants of Health | n (%) | SDH Factor Only a | SDH Factor + Past Morbidity Burden + Past Cost b |
|---|---|---|---|
| A. Structural Determinants | |||
| Less than high school degree | 1,533 (17%) | 6,893 (3,903, 9,882) | 95 (–1,013, 1,202) |
| Not working | 4,397 (49%) | 13,366 (11,569, 15,162) | 643 (–95, 1,381) |
| B. Intermediary Determinants | |||
| Not married | 7,955 (89%) | 2,518 (–309, 5,344) | 1,369 (175, 2,562) |
| Lives with children | 4,566 (51%) | –6,651 (–8,607, –4,695) | –957 (–1,755, –160) |
| Relationship with someone who threatened/physically hurt you | 2,684 (30%) | 3,537 (1,568, 5,506) | 52 (–716, 819) |
| History of being in jail/prison | 2,836 (32%) | 2,295 (208, 4,381) | 667 (–97, 1,432) |
| Feels lonely a good bit to all of the time | 1,630 (18%) | 5,624 (3,187, 8,061) | 954 (–80, 1,987) |
| Currently smokes | 3,629 (41%) | 2,783 (927, 4,640) | 266 (–408, 939) |
| Uses illicit drugs | 359 (4%) | 6,562 (1,046, 12,078) | 306 (–1,679, 2,291) |
| More than recommended weekly drinking | 383 (4%) | –2,480 (–6,700, 1,740) | 1,194 (–612, 3,000) |
| Exercises vigorously < 3 days per week | 4,198 (47%) | 7,044 (5,293, 8,795) | 786 (93, 1,479) |
| Trouble paying phone bill | 1,974 (22%) | –308 (–2,336, 1,720) | 118 (–725, 962) |
| Trouble paying utilities | 2,401 (27%) | 1,897 (–79, 3,873) | 563 (–232, 1,359) |
| Trouble paying rent/mortgage | 1952 (22%) | 1,011 (–1,077, 3,098) | 26 (–792, 844) |
| Food insecurity | 4,470 (50%) | 4,589 (2,636, 6,543) | 456 (–281, 1,194) |
| Limited to no internet access at home | 717 (8%) | 13,969 (8,754, 19,185) | 3,397 (1,116, 5,679) |
| No access to a car for medical appointments | 5,337 (60%) | –2,771 (–4,535, –1,006) | –369 (–1,098, 360) |
| Does not live in a house or apartment | 736 (8%) | 10,933 (6,854, 15,012) | 2,879 (795, 4,963) |
| Has lived at current place ≤ 1 year | 1,339 (15%) | 2,404 (13, 4,796) | 283 (–697, 1,262) |
| Unstable housing | 1,986 (22%) | 5,954 (3,709, 8,199) | 1,232 (198, 2,267) |
| Reports ≥ 2 housing condition problems | 1,365 (15%) | 2,211 (–221, 4,643) | 430 (–604, 1,464) |
Regression coefficients estimated from a model that includes the individual social risk factor and age and sex interaction effect (model 1).
Regression coefficients estimated from a model that includes the individual social risk factor, age and sex interaction effect, past morbidity burden, and past costs (model 2).
When we repeated the main analyses, the results were qualitatively the same when we assumed zero costs for the months without costs (see Supplemental Table S1). When we replaced the total annual costs with $200,000 for participants with annual costs higher than $200,000, the results were also similar to our main results (see Supplemental Table S2). When we used the Chronic Disability Payment System (CDPS) morbidity burden score instead of the ACG system, our results were similar (see Supplemental Table S3).
When we used generalized linear regression to model future annual Medicaid costs as a function of social risk class, past morbidity burden, and past costs, our results were qualitatively the same (see Supplemental Table S4). Finally, we obtained similar results when we modeled the odds of becoming a high‐cost user during the one‐year follow‐up period (see Supplemental Table S5). Given age and sex, the odds of becoming a high‐cost health care user were significantly higher for all the upper social risk classes compared with the lowest social risk classes. When we controlled for past morbidity burden and past costs, the odds of becoming a high‐cost health care user remained significantly higher for social risk classes 3 and 4 compared with the lowest social risk class. When we included future morbidity burden, social risk class no longer was a significant predictor of becoming a high‐cost user.
Discussion
In this adult Medicaid cohort, we demonstrated strong and meaningful relationships between the social determinants of health and future morbidity burdens, as well as SDH and future health care spending. Participants in the higher social risk classes with similar past morbidity burdens and past costs had greater morbidity burdens one year later compared with those in the lowest social risk class. Similarly, when we controlled for past morbidity burden and past costs, those participants with more social risks had higher annual Medicaid costs than did participants in the lowest social risk class. When we included concurrent morbidity in our cost prediction models, SDH was not a significant predictor of costs.
The results of our study are consistent with two bodies of evidence. First, it is well‐established that adverse SDH are linked to higher morbidity 13 , 14 , 15 , 16 and mortality. 17 , 18 , 19 , 20 After adjusting for past morbidity burden and past costs, we also found that future disease burden was worse among participants with more social adversities. Second, our results are also congruent with past studies that found an association between lower socioeconomic status (SES) and higher health care spending. 21 , 22 , 23 , 24 Even in a low‐income Medicaid population, we observed large differences in health care costs for adults with different types and severities of social risk.
In addition to the SDH gradient among the four social risk classes and annual costs, we noted high excess costs attributed to individual SDHs when not adjusted for past morbidity burden and past costs. These unadjusted results are similar to those found by Tarasuk and colleagues, who reported significantly higher annual health care spending by adults who were marginally to severely food insecure ($235 to $1,092) compared with those who were food secure. 25 Their cost estimates attributed to food insecurity are lower than ours ($4,589), most likely because our sample consisted of low‐income adults, whereas theirs was based on a nationally representative sample of adults living in Ontario, Canada, and they also adjusted for other SES factors. 25
When we included concurrent morbidity burden in our cost prediction models, social risk class was no longer a significant predictor of future costs. This result, along with the correlation between future morbidity and future costs and the finding that SDH predicts both future costs and morbidity, is consistent with a mechanism in which SDH impacts costs through greater illness severity. If this is true, then including concurrent morbidity in cost prediction models will lead to the erroneous conclusion that SDH does not affect future health care costs. When we measured costs and morbidity during the same time frame, our results suggested that the deleterious effect of adverse SDH on costs is largely captured by concurrent morbidity.
It also is possible that the morbidity burden measures we included (i.e., ACG and CDPS) in our models, which were developed to predict cost, are measures of cost and utilization as opposed to meaures of morbidity burden. In addition, those participants who become ill may have had worse social and economic circumstances and not vice versa. We acknowledge that there is a feedback effect among SDH, morbidity burden, and health care costs and that this can only be properly evaluated over a longer time horizon with multiple, repeated measures of all three. Either way, when the prediction models include the past ACG or CDPS scores, SDH remains a significant predictor of future costs, which suggests that there is value in including SDH in prospective cost prediction models.
The results of our study have important policy implications. Because SDH are an important driver of future health care costs, we must address these upstream, nonmedical factors that cause poor health. We found that concurrent morbidity largely captures both the medical and social risks associated with health care spending. This finding is worthy of future research because it is important that our payment systems adequately compensate health care providers who treat highly complex patients so that they have the needed resources to address their patients’ medical and social needs. 26 , 27 When using a managed care payment formula that estimates the capitation payment for the next year based on age, sex, and the previous year of morbidity burden, our results imply that the addition of social risk to the reimbursement model would improve the compensation to providers who treat a disproportionate share of beneficiaries with socioeconomic challenges.
At the next level, we must better integrate health and human services so that both systems can more effectively and efficiently serve vulnerable populations. 28 Some Medicaid programs have begun to partner with other sectors and/or community‐based organizations to address their beneficiaries’ nonmedical needs, but we need much more collaboration and evaluation of the efforts in this area. 29 Finally, at the highest level, we need policy interventions that focus on the powerful upstream factors that largely determine one's health (e.g., education, economic opportunity, the physical environment) before the health care system becomes involved. 30 Only when we have more generous and equitable public policies in the areas of housing, labor protections, income support, and education will we be likely to see the positive downstream benefits for health and health care costs that we seek. 31 , 32 , 33 , 34
Limitations
Our findings should be considered in light of the following limitations. First, the generalizability of our results is limited to an urban Medicaid population enrolled at the time of a health care visit, rather than a random, representative sample of adult Medicaid beneficiaries. Our sample did not include children, dual eligibles, or non‐English‐speaking beneficiaries. Our study cohort also was sicker than the general population of adult DC Medicaid beneficiaries, and we enrolled them in the catchment areas of two hospitals. The variation of SDH that we observed in our nonrepresentative sample is likely less than the variation in the overall Medicaid population. Similarly, our estimation of the impact of SDH on future costs may be smaller in a general population of Medicaid beneficiaries. The average health care costs of our study sample were almost double the average health care expenditures of the general DC Medicaid population during our study's time period.
Second, although we measured a wide array of SDH factors, we likely missed factors affecting health care costs such as social support, health literacy, and neighborhood deprivation. Third, although the majority of our SDH questionnaire was based on validated items and scales, several factors were not (e.g., exercise and alcohol use). Finally, because we evaluated the relationship among SDH, morbidity burden, and costs over a short time, we were unable to investigate the feedback effect of past expenditures and past morbidity burden on current social and medical risks.
Conclusions
A sizable proportion of health care costs is associated with social determinants of health. SDH are an important predictor of future health care costs when they are adjusted for past morbidity burden and past costs. The impact of SDH on future health care expenditures is largely accounted for by concurrent morbidity. Our results suggest that reducing health care costs in the Medicaid population will require cross‐sectoral collaborations and multilevel interventions aimed at eliminating the structural inequities that contribute to large health disparities in the United States.
Funding/Support: This study was funded by the National Institute for Minority Health and Health Disparities (NIMHD), grant number R01MD011607.
Conflict of Interest Disclosures: The authors do not have any other associations, financial or otherwise, to disclose besides our employer affiliations and funder.
Acknowledgments: The authors would like to thank April Grady, the associate director of the Division of Analytics and Policy Research, at the District of Columbia Department of Health Care Finance, for answering our questions and providing technical assistance related to the Medicaid claims data.
Supporting information
Table S1. Annual Medicaid costs by social risk class assuming that month(s) of missing costs equals zero
Table S2. Annual Medicaid costs by social risk class with costs top coded at $200,000
Table S3. Annual Medicaid costs by social risk class with chronic disability payment system (CDPS) risk score
Table S4. Ratio of expected costs by social risk class using a generalized linear regression model approach
Table S5. Adjusted odds ratio (95% confidence interval) of becoming a high ‐cost health care user during one year after study enrollment by social risk class for 8,447 Medicaid participantsa
References
- 1. Frieden TR. A framework for public health action: the health impact pyramid. Am J Public Health. 2010;100(4):590‐595. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. McGinnis JM, Williams‐Russo P, Knickman JR. The case for more active policy attention to health promotion. Health Aff (Millwood). 2002;21(2):78‐93. [DOI] [PubMed] [Google Scholar]
- 3. Fitzpatrick T, Rosella LC, Calzavara A, et al. Looking beyond income and education: socioeconomic status gradients among future high‐cost users of health care. Am J Prev Med. 2015;49(2):161‐171. [DOI] [PubMed] [Google Scholar]
- 4. Irvin JA, Kondrich AA, Ko M, et al. Incorporating machine learning and social determinants of health indicators into prospective risk adjustment for health plan payments. BMC Public Health. 2020;20(1):608. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Ash AS, Mick EO, Ellis RP, Kiefe CI, Allison JJ, Clark MA. Social Determinants of Health in managed care payment formulas. JAMA Intern Med. 2017;177(10):1424‐1430. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Schuler A, O'Súilleabháin L, Rinetti‐Vargas G, et al. Assessment of value of neighborhood socioeconomic status in models that use electronic health record data to predict health care use rates and mortality. JAMA Network Open. 2020;3(10):e2017109. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. McCarthy ML, Zheng Z, Wilder ME, et al. Latent class analysis to represent social determinant of health risk groups in the Medicaid cohort of the District of Columbia. Med Care. 2021;59(3):251‐258. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8. The Johns Hopkins ACG System: State of the Art Technology and a Tradition of Excellence in One Integrated Solution. Baltimore, MD: Center HSRaD; 2012. [Google Scholar]
- 9. Weiner JP, Tucker AM, Collins AM, et al. The development of a risk‐adjusted capitation payment system: the Maryland Medicaid model. J Ambulatory Care Manage. 1998;21(4):29‐52. [DOI] [PubMed] [Google Scholar]
- 10. The Johns Hopkins ACG® System Version 12.0 User Documentation. In: Weiner JP, Abrams, C , eds. Baltimore, MD: Johns Hopkins University; 2019: https://www.hopkinsacg.org/. Accessed April 1, 2019. [Google Scholar]
- 11. Chen J, Shao J. Iterative weighted least squares estimators. Ann Statistics. 1993;21(2):1071‐1092. [Google Scholar]
- 12. Wilder ME, Zheng Z, Zeger SL, et al. Relationship between social determinants of health and antihypertensive medication adherence in a Medicaid cohort. Circ Cardiovasc Qual Outcomes. 2022:15:e008150. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Martinson ML, Teitler JO, Plaza R, Reichman NE. Income disparities in cardiovascular health across the lifespan. SSM–Popul Health. 2016;2:904‐913. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Braveman PA, Cubbin C, Egerter S, Williams DR, Pamuk E. Socioeconomic disparities in health in the United States: what the patterns tell us. Am J Public Health. 2010;100 (Suppl. 1):S186‐196. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Bilger M, Kruger EJ, Finkelstein EA. Measuring socioeconomic inequality in obesity: looking beyond the obesity threshold. Health Economics. 2017;26(8):1052‐1066. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Coughlin SS. Social determinants of breast cancer risk, stage, and survival. Breast Cancer Res Treatment. 2019;177(3):537‐548. [DOI] [PubMed] [Google Scholar]
- 17. Sasson I. Diverging trends in cause‐specific mortality and life years lost by educational attainment: evidence from United States vital statistics data, 1990–2010. PLOS ONE. 2016;11(10):e0163412. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Case A, Deaton A. Life expectancy in adulthood is falling for those without a BA degree, but as educational gaps have widened, racial gaps have narrowed. PNAS USA. 2021;118(11):e2024777118. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Chetty R, Stepner M, Abraham S, et al. The association between income and life expectancy in the United States, 2001–2014. JAMA. 2016;315(16):1750‐1766. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Harris KM, Woolf SH, Gaskin DJ. High and rising working‐age mortality in the US: a report from the National Academies of Sciences, Engineering, and Medicine. JAMA. 2021;325(20):2045‐2046. [DOI] [PubMed] [Google Scholar]
- 21. Roos NP, Mustard CA. Variation in health and health care use by socioeconomic status in Winnipe.g., Canada: does the system work well? Yes and no. Milbank Q. 1997;75(1):89‐111. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Chen W, Lynd LD, FitzGerald JM, Sadatsafavi M. Influences of socioeconomic status on costs of asthma under universal health coverage. Med Care. 2016;54(8):789‐795. [DOI] [PubMed] [Google Scholar]
- 23. Report summary—the direct economic burden of socioeconomic health inequalities in Canada: an analysis of health care costs by income level. HPCDP. 2016;36(6):118‐119. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. de Boer WIJ, Buskens E, Koning RH, Mierau JO. Neighborhood socioeconomic status and health care costs: a population‐wide study in the Netherlands. Am J Public Health. 2019;109(6):927‐933. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Tarasuk V, Cheng J, de Oliveira C, Dachner N, Gundersen C, Kurdyak P. Association between household food insecurity and annual health care costs. CMAJ. 2015;187(14):E429‐E436. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Joynt Maddox KE. Social and behavioral determinants of spending. JAMA Intern Med. 2017;177(10):1431‐1432. [DOI] [PubMed] [Google Scholar]
- 27. Joynt KE, De Lew N, Sheingold SH, Conway PH, Goodrich K, Epstein AM. Should Medicare value‐based purchasing take social risk into account? N Engl J Med. 2017;376(6):510‐513. [DOI] [PubMed] [Google Scholar]
- 28. Fichtenberg C, Delva J, Minyard K, Gottlieb LM. Health and human services integration: generating sustained health and equity improvements. Health Aff (Millwood). 2020;39(4):567‐573. [DOI] [PubMed] [Google Scholar]
- 29. Kassler WJ, Tomoyasu N, Conway PH. Beyond a traditional payer—CMS's role in improving population health. N Engl J Med. 2015;372(2):109‐111. [DOI] [PubMed] [Google Scholar]
- 30. Nichols LM, Taylor LA. Social determinants as public goods: a new approach to financing key investments in healthy communities. Health Aff (Millwood). 2018;37(8):1223‐1230. [DOI] [PubMed] [Google Scholar]
- 31. Montez JK, Beckfield J, Cooney JK, et al. US state policies, politics, and life expectancy. Milbank Q. 2020;98(3):668‐699. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32. Bradley EH, Elkins BR, Herrin J, Elbel B. Health and social services expenditures: associations with health outcomes. BMJ Quality Safety. 2011;20(10):826‐831. [DOI] [PubMed] [Google Scholar]
- 33. Avendano M, Kawachi I. Why do Americans have shorter life expectancy and worse health than do people in other high‐income countries? Annu Rev Public Health. 2014;35:307‐325. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34. Taylor LA, Tan AX, Coyle CE, et al. Leveraging the social determinants of health: what works? PLOS ONE. 2016;11(8):e0160217. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Table S1. Annual Medicaid costs by social risk class assuming that month(s) of missing costs equals zero
Table S2. Annual Medicaid costs by social risk class with costs top coded at $200,000
Table S3. Annual Medicaid costs by social risk class with chronic disability payment system (CDPS) risk score
Table S4. Ratio of expected costs by social risk class using a generalized linear regression model approach
Table S5. Adjusted odds ratio (95% confidence interval) of becoming a high ‐cost health care user during one year after study enrollment by social risk class for 8,447 Medicaid participantsa