Social Determinants of Health and Geographic Variation in Medicare per Beneficiary Spending

Key Points Question How much variation in Medicare per beneficiary spending across counties was associated with social determinants of health (SDoH)? Findings In this cross-sectional study, SDoH were associated with 37.7% of variation in price-adjusted Medicare per beneficiary spending between counties in the highest and lowest quintiles of spending in 2017, including both direct contributions and indirect contributions through other factors. SDoH’s direct contribution accounted for 5.8% of the variation after controlling for patient demographic characteristics, clinical risk, and supply of health care resources. Meaning These findings suggest that addressing SDoH is important for reducing geographic spending variation and improving the value of health care.


eAppendix 1. Selection of SDoH Variables
We used the National Academy of Medicine (NAM) conceptual framework as the basis for selecting SDoH measures. 1The NAM conceptual framework specifies five categories of SDoH associated with Medicare spending, including (1) socioeconomic position, (2) race and ethnicity composition, (3) social relationships, (4) overall residential and community context, and (5) gender.We identified 87 publicly available county-level SDoH measures from literature and web searches (eTable 1).SDoH measures identified in our searches, but not publicly available were excluded from this study (e.g., walking score, transit score, self-reported financial burden, and self-reported financial barriers to medication).After we identified publicly available SDoH measures, we mapped each of the 87 SDoH measures to one of the four NAM's conceptual framework categories and the subcategories therein (e.g.income, insurance etc. under socioeconomic position).We made two changes to the conceptual framework after this step.First, we did not use gender as one of the SDoH measures as we considered gender as part of demographics.Second, NAM's conceptual framework considered healthcare resources to be part of the Residential and Community Context category; however, we used healthcare resources separately to be consistent with previous literature, [2][3][4][5] which emphasized the importance of the supply of healthcare resources to regional spending variation.
After mapping of SDoH measures to the NAM conceptual framework, we qualitatively screened SDoH measures that were conceptually similar under the same category or across subcategories.For example, % receiving public assistance income, % receiving supplement security income, and % receiving food stamp/snap in the "Income" subcategory all capture poverty.We therefore only included % residents in poverty (based on federal poverty threshold).We selected up to two SDoH variables for each conceptually similar measure for further consideration.
The qualitative screening generated a total of 13 SDoH measures, generally with one measure in each subcategory (except for marital status and living alone which we did not include in any measure as they conceptually overlap with social relationships), including six for socioeconomic position (median household income, % of residents in poverty, % of residents who are uninsured, unemployment rate, % of residents without a high school degree, and food environment index), three for race and ethnicity composition (% of residents who are nonwhite, % of residents who are non-citizen, and % residents with limited English proficiency), one for social relationships (number of membership associations per 1,000 population), and three for overall residential and community context (% of households with severe housing problems, % of residents with access to exercise opportunities, and % of housing units in rural areas).Finally, we tested the correlation between SDoH measures within each category (eTables 2-4).For each group of measures that captured similar concepts and were highly correlated (i.e., correlation coefficient over 0.7), 6 we selected the variable that was most commonly used in the literature.Therefore, we dropped % of residents in poverty given its high correlation with median household income (eTable 2).We also dropped % residents with limited English proficiency as it is highly correlated with % of residents who are non-citizen (eTable 3).We tested the correlation between the remaining 11 SDoH measures and included them in the analysis (eTable 5).We subsequently adopted a more detailed race/ethnicity classification and replaced the % resident who are non-White with % Hispanic, % non-Hispanic Black and % non-Hispanic with another race.Therefore, our final analyses include 13 SDoH measures.

. Sources of social determinants of health measures used in this study
Notes: 1 Food environment index equally weights two indicators of the food environment:(1) Limited access to healthy foods, which estimates the percentage of the population that is low income and does not live close to a grocery store.(2) Food insecurity, which estimates the percentage of the population that did not have access to a reliable source of food. 2 Other races include American Indian and Alaska Native, Asian, Native Hawaiian and Other Pacific Islander, and other races. 3Social Associations measures the number of membership associations per 10,000 population. 4Severe housing problems is the percentage of households with one or more of the following housing problems: (1) Housing unit lacks complete kitchen facilities; (2) Housing unit lacks complete plumbing facilities; (3) Household is overcrowded; or (4) Household is severely cost burdened. 5Access to exercise opportunities measures the percentage of individuals in a county who live reasonably close to a location for physical activity, defined as parks or recreational facilities.Individuals are considered to have access to exercise opportunities if they reside in a census block that is within a half mile of a park, or reside in an urban census block that is within one mile of a recreational facility, or reside in a rural census block that is within three miles of a recreational facility.More information about these measures could be found at: https://www.countyhealthrankings.

Regression models to examine the contribution of patient, supply, and SDoH characteristics to geographic variation in per beneficiary Medicare total spending
We first categorized counties into quintiles based on their price-adjusted per beneficiary Medicare spending in 2017 and calculated the differences in mean price-adjusted per beneficiary Medicare spending between each higher spending quintile (quintiles 2-5) and quintile 1.We then followed previously developed methods to examine the extent to which the variation in priceadjusted per beneficiary Medicare spending across quintiles could be explained by (1) patient demographics, (2) patient clinical risk, (3) supply of health resources, and (4) SDoH.
To assess the total contribution of each group of characteristics to geographic variation in Medicare spending, we first ran a linear regression model where the outcome variable is the price-adjusted per beneficiary spending and explanatory variables are one of the four groups of characteristics above.
In this equation,   represents the price-adjusted per beneficiary Medicare spending in each county ,   is a vector of independent variables (e.g., demographics or clinical risk). represents the coefficients estimating the relationship between per beneficiary Medicare spending and the independent variables.  represents the error term.This model is weighted by the number of fee-for-service patients in each county.
After estimating model (1) using OLS, we estimated the predicted value of the outcome  �  given the independent variables and estimated coefficients  ̂ and calculated the residual for each county as   =   - �  .  represents the per beneficiary spending that is not explained by independent variables.We then calculated the mean per beneficiary spending across all counties as  � = ∑   3,038 =1 . Finally, the adjusted per beneficiary spending for each county was calculated as  � _ =  � +   ,which removes variation in   explained by   .The adjusted variation in per beneficiary spending was calculated as the differences in mean  � _ among counties in quintiles 2-5 and mean  � _ among counties in quintile 1.If the independent variables in the regression model (1) could explain the variation, we would expect a narrowed variation across quintiles.The share of the variation explained by the independent variable was calculated as one minus the ratio of variation in adjusted spending  � _ to that in price-adjusted per beneficiary spending   , times 100.This model and the estimation process were repeated for four times to calculate the total contribution of each group of characteristics.
We note that this approach is analogous and yields similar results to the R-squared statistic of the regression with   as the dependent variable and   as independent variables (Figure 3 and Table 2).The current approach has the benefit of allowing us to flexibly present changes in spending in terms of dollar amounts of counties in different spending quintiles.
To estimate the direct contribution of each group of characteristics, we ran a single model using all characteristics as independent variables.
=  +  1 ℎ  +  2    +  3   +  4   +   (2)  Similar with model 1, we first calculated the residual for each county as   =   - �  after estimating model (1) using OLS.We then sequentially replaced each group of characteristics using their means across all counties and estimated the predicted per beneficiary spending  �  given the independent variables and estimated coefficients  1 � −  4 � .Finally, the adjusted per beneficiary spending is calculated as  �   =  �  +   .Similarly, the adjusted variation in per beneficiary spending was calculated as the differences in mean  � _ among counties in quintiles 2-5 and mean  � _ among counties in quintile 1.The share of the variation explained by the independent variable was calculated as one minus the ratio of variation in adjusted spending  � _ to that in price-adjusted per beneficiary spending   , time 100.This process was repeated four times to calculate the direct contribution of each group of characteristics.
eTable 7. Summary of regression models and their purposes  Notes: For each quintile, the share of variation associated with each set of characteristics was estimated when controlling for other characteristics.Demographics include age, age squared, age cubed, and gender; supply characteristics include the following measures per 1,000 population: primary care physicians, specialists, hospital beds, skilled nursing facility beds, home health agency aides, registered nurses employed by hospices, and ambulatory care centers.SDoH include median household income, % who are uninsured, unemployment rate, % without high school degree, food environment index; % who are Hispanic, % of non-Hispanic black, and % of non-Hispanic with another race (i.e., American Indian and Alaska Native, Asian, Native Hawaiian and Other Pacific Islander, and other races), % who are noncitizen, social associations per 1,000 population, % with severe housing problems, % with access to exercise opportunities, and % of housing units in rural areas.

eFigure 2. Contribution to variation in price-, age-, gender-, and race-adjusted per beneficiary spending between quintiles 2-5 and quintile 1, using Dartmouth spending as outcome
Notes: For each quintile, the share of variation associated with each set of characteristics was estimated when controlling for other characteristics.Supply characteristics include the following measures per 1,000 population: primary care physicians, specialists, hospital beds, skilled nursing facility beds, home health agency aides, registered nurses employed by hospices, and ambulatory care centers.SDoH include median household income, % who are uninsured, unemployment rate, % without high school degree, food environment index; % who are Hispanic, % of non-Hispanic black, and % of non-Hispanic with another race (i.e., American Indian and Alaska Native, Asian, Native Hawaiian and Other Pacific Islander, and other races), % who are non-citizen, social associations per 1,000 population, % with severe housing problems, % with access to exercise opportunities, and % of housing units in rural areas.Notes: For each quintile, the share of variation associated with each set of characteristics was estimated when controlling for other characteristics.Demographics include age, age squared, age cubed, and gender.Supply characteristics include the following measures per 1,000 population: primary care physicians, specialists, hospital beds, skilled nursing facility beds, home health agency aides, registered nurses employed by hospices, and ambulatory care centers.SDoH include median household income, % who are uninsured, unemployment rate, % without high school degree, food environment index; % who are Hispanic, % of non-Hispanic black, and % of non-Hispanic with another race (i.e., American Indian and Alaska Native, Asian, Native Hawaiian and Other Pacific Islander, and other races), % who are noncitizen, social associations per 1,000 population, % with severe housing problems, % with access to exercise opportunities, and % of housing units in rural areas. eAppendix

eFigure 1 .
Contribution to variation in price-adjusted per beneficiary Medicare spending between quintiles 2-5 and quintile 1, excluding clinical risk score

eFigure 3 .
Contribution to variation in CMS HCC score between quintiles 2-5 and quintile 1

eTable 1 List of social determinants of health variables considered in the study ID Categories and Subcategories
Department of Agriculture; USPS: United States Postal Service; HUD: United States Department of Housing and Urban Development.SNF: skilled nursing facility; HHA: home health agency.* Definition from https://seer.cancer.gov/seerstat/variables/countyattribs/static.html#14-18

Correlations between Social determinants of health measures included in the study
© 2021 Zhang Y et al.JAMA Network Open.

eTable 8. Full regression output of sensitivity analysis: coefficients and robust standard errors
PCP: primary care physicians; SNF: skilled nursing facility; HHA: home health agency; RN: registered nurses; ASC: ambulatory surgery center.* P<0.05, **P<0.01,***P<0.001.Results are from the linear regressions using CMS price-adjusted per beneficiary Medicare spending or Dartmouth price-, age-, gender-, and race-adjusted spending as outcome, controlling for variables in each column in the regression models.We reported coefficients and robust standard errors for each variable.
© 2021 Zhang Y et al.JAMA Network Open.

eTable 9. Full regression output of the association of CMS-HCC score with SDoH, demographics, and supply of healthcare sources: coefficients and robust standard errors
PCP: primary care physicians; SNF: skilled nursing facility; HHA: home health agency; RN: registered nurses; ASC: ambulatory surgery center.* P<0.05, **P<0.01,***P<0.001.Results are from the linear regressions using CMS-HCC scores as the outcome.We reported coefficients and robust standard errors for each variable.