Diabetes and multiple chronic conditions increase overall Medicare spending, but spending increases even more in minority beneficiaries compared with White beneficiaries with similar comorbidity combinations.
Objectives: To determine which combinations of type 2 diabetes (T2D) and multiple chronic conditions (MCC) contribute to total spending and differences in spending between groups based on sex, race/ethnicity, and rural residency.
Study Design: Retrospective cohort study using 2012 Medicare claims data from beneficiaries in Michigan with T2D.
Methods: Zero-inflated Poisson regression models to estimate relationships of demographic characteristics and MCC combinations on hospital outpatient, acute inpatient, skilled nursing, hospice, and Part D drug spending.
Results: Across most MCC combinations, there are lower odds of no spending, with a concurrent increase in the expected mean of actual spending when payments are made, except for hospital outpatient costs. For hospital outpatient services, we observed lower spending across all MCC combinations. When controlling for MCC, we generally found that compared with White beneficiaries, Black, Asian/Pacific Islander, and Hispanic beneficiaries experience increased odds of no spending, but when payments were made, payments generally increased. American Indian/Alaska Native beneficiaries are the exception; they experience decreased odds of no payments for hospital outpatient and acute inpatient services, with a concurrent decrease in mean expected payments.
Conclusions: When considering a range of MCC combinations, we observed differences in total payments between racial/ethnic minority groups and White beneficiaries. Our results highlight the ongoing need to make changes in the health care system to make the system more accessible to racial/ethnic minority groups.
Am J Manag Care. 2020;26(11):e362-e368. https://doi.org/10.37765/ajmc.2020.88531
Medicare spending among beneficiaries 65 years and older with type 2 diabetes is greatly affected by the presence of multiple chronic conditions. The presence of multiple chronic conditions increases the odds of any payments being made for services, as well as the mean spending in multiple service categories. However, patient characteristics, especially race, are also associated with variation in total spending for services.
In the United States, direct medical costs associated with type 2 diabetes (T2D) are estimated at $237 billion; 61% of the cost was associated with patients 65 years and older.1 Lifetime medical costs are higher when T2D is diagnosed at earlier ages and for women.2,3 Non-Hispanic Black patients and Hispanic patients generally have increased costs related to T2D.4
Many factors drive costs associated with T2D diagnosis. Medication adherence is associated with lower service costs and higher drug costs.5 Nonadherence among lower-income patients increases hospitalizations and emergency department costs.6
Comorbidity profiles also increase costs associated with T2D.7 Brandle et al report that increased body mass index and presence of comorbidities are associated with higher medical costs.8 Individuals with T2D experience costs 2.8 times higher than what is spent by age-matched and sex-matched individuals without diabetes, and 40% of their costs come from diabetes-related vascular comorbidities.9 Racial and ethnic minority groups have significantly lower out-of-pocket costs than non-Hispanic White patients, whereas White and Black patients have higher rates of insurance to cover costs.10,11 Lee et al report no differences based on race/ethnicity in diabetes management or total related costs.12
Most research considers effects of comorbidities by simply counting numbers of comorbidities or by considering 1 or 2 additional comorbidities to T2D. We take this a step further by including 6 additional chronic conditions that are leading causes of death (LCD) in Michigan to determine which combinations of multiple chronic conditions (MCC) influence hospital outpatient, acute inpatient, skilled nursing, hospice, and Part D drug payments. We hypothesize that specific combinations of MCC contribute more to costs than diabetes alone and that disparities in costs exist among groups based on sex, race/ethnicity, and rural residency, even after controlling for MCC combinations.
We used 2012 claims data from CMS for 1,851,328 Medicare beneficiaries in Michigan. We eliminated 351,002 beneficiaries (remaining n = 1,500,326) younger than 65 years because someone becomes eligible for Medicare before age 65 years only if they are eligible for Social Security disability, and we wanted to eliminate early disability (for disparate reasons such as end-stage renal disease, amyotrophic lateral sclerosis, or disability from railroad employment that we could not ascertain from the claims data) as a confounder in our analyses. The Chronic Conditions Warehouse (CCW) data identify people with diabetes (n = 511,120; 34.1% of beneficiaries 65 years and older), but they include both type 1 diabetes (T1D) and T2D. Because our goal is to study beneficiaries with T2D, we eliminated beneficiaries with T1D diagnosis codes (International Classification of Diseases, Ninth Revision [ICD-9] codes 250.x1, 250.x3) or secondary diabetes codes (ICD-9 codes 249.xx) for a sample size of 448,407 (29.9% of beneficiaries 65 years and older). Finally, we eliminated 4975 cases with missing race codes or that were coded as other or unknown (considered missing data because of the uncertain nature of these codes), for a final sample size of 443,932.
Chronic conditions that are LCD in Michigan are heart disease, cancer, chronic lower respiratory disease, stroke, Alzheimer disease, diabetes, and nephritis/chronic kidney disease (CKD).13 We used variables from the CCW identifying beneficiaries with congestive heart failure (CHF) as a proxy measure for heart disease; chronic obstructive pulmonary disease (COPD) as a proxy measure for chronic lower respiratory disease; and stroke, Alzheimer disease, and CKD as measures of those conditions. We used a combined measure of ever being diagnosed with 1 of the following cancers for our cancer variable: breast, colorectal, endometrial, lung, and prostate. The CCW uses a specific algorithm to determine whether a beneficiary has a condition. For instance, for COPD, the CCW tags a beneficiary as having COPD if they meet the following criteria: at least 1 inpatient, skilled nursing facility, or home health agency or 2 hospital outpatient or Part B Carrier claims with COPD-related diagnosis, diagnosis-related group, or procedure codes in the previous year. Similar algorithms are used to identify beneficiaries with any of the LCD we studied. The possibility exists that diagnoses are undercoded for any given year. However, the CCW also provides an indicator of the first year (prior to the data we used from 2012) when services were billed for related to diabetes or any of the LCD we studied. We also used this information to identify beneficiaries with diabetes and each MCC. Although this does not completely alleviate the possibility of undercoding in 2012, it provides some measure of cumulative disease by that year.
Because our aim was to establish that differences in cost exist based on MCC combinations, we created a new variable that identified beneficiaries as having diabetes alone, as well as any of the 63 combinations of T2D and chronic conditions. The frequency and percentage of each combination is reported in eAppendix Table 1 (eAppendix available at ajmc.com). Instead of including all 64 combinations of disease, we chose a cross-section of combinations representing diabetes plus 1, 2, 3, and 4 additional MCC (generally based on highest prevalence) and created a categorical measure of the top 10 T2D plus MCC combinations, which include T2D alone (reference category), diabetes plus CHF, diabetes plus cancer, diabetes plus COPD, diabetes plus stroke, diabetes plus CKD, diabetes plus CHF/COPD, diabetes plus CHF/CKD, diabetes plus CHF/COPD/CKD, diabetes plus CHF/COPD/stroke/CKD, and diabetes plus all other MCC combinations.
From the Cost and Use Database, we selected variables that reflected spending for each of the selected service categories: hospital outpatient payments (CMS variable names: HOP_MDCR and HOP_BENE), acute inpatient payments (CMS variable names: ACUTE_MD and ACUTE_BE), skilled nursing payments (CMS variable names: SNF_MDCR and SNF_BENE), hospice payments (CMS variable name: HOS_MDCR), and Part D drug payments (CMS variable names: PTD_MDCR and PTD_BENE), in US$. In each case, the MD or MDCR variables reflect total spending by Medicare, whereas the BE or BENE variables reflect coinsurance or deductibles that were the patient’s responsibility. In each case, amounts from these variables were added together to get the total spending variable we used. There is no beneficiary payment for hospice, as this benefit is paid at 100% by Medicare.
Because we are interested in determining if spending differences exist based on demographic characteristics, we included the following independent variables in our analyses: age (years); male sex (reference: female); race: Black, Asian/Pacific Islander, Hispanic, and American Indian/Alaska Native (AI/AN) (reference: White); and rural residence (reference: urban). Because analyzing relationships for 63 combinations of disease would create small subgroup sizes, we included the aforementioned variable that categorized the top 10 MCC combinations, as well as another category that included all other MCC combinations (reference: diabetes alone).
We used zero-inflated Poisson (ZIP) regression models to estimate the relationships of demographic characteristics and the aforementioned categorical measure of 10 combinations of T2D plus MCC with our 5 dependent spending variables, using SAS software version 9.4 (SAS Institute). The typical nature of spending data is that they contain a significant number of cases with zero spending (ie, zero-inflated data), reflecting the notion that some beneficiaries receive no services in a specific service category. Zero-inflated data have a larger proportion of zeros than might be expected from a pure count (Poisson) distribution.14 For our dependent variables, “zeros” are those beneficiaries with no payments in a service category. The percentage of zeros ranges from 20% of hospital outpatient services to 90% for both skilled nursing and hospice services. ZIP models assess the relationships of selected covariates with the probability of membership in the “true zero” group and simultaneously assess the relationships of selected covariates with the dependent counts (ie, spending) for those not in the “true zero” group. Two separate models are used in traditional ZIP regression modeling.15 First, a logit model predicts membership in the zero-coded group (ie, those who did not have any payments). Second, a Poisson model predicts the actual counts (ie, the total payments) for those who are not in the zero-coded group. We included all independent variables as covariates in each model. Using this technique, we first determined which beneficiary characteristics predict not having any spending and then examined how these characteristics influence total spending in each category. We report unstandardized regression coefficients, and for ease in understanding the results, we converted unstandardized coefficients to odds ratios by exponentiating the coefficients to report percentage increase or decrease in odds of being in a “no spending” group or in the mean expected payments.
Table 1 includes descriptive statistics for each of our dependent and independent variables. Mean (SD) annual costs range from a low of $3030 ($6970) for hospital outpatient costs to a high of $20,434 ($23,485) for acute inpatient costs. Table 2 presents the results of the ZIP regression models. For each dependent variable, the first column presents unstandardized logit regression coefficients (predicting membership in the no spending group) and the second column presents unstandardized Poisson regression coefficients (predicting total spending for those who had any payments made).
The results of the logit models reveal consistent patterns of differences in total spending based on race across hospital outpatient, acute inpatient, skilled nursing, and hospice cost categories. In general, Black, Asian/Pacific Islander, and Hispanic beneficiaries have higher odds of no spending in each category, respectively, compared with White beneficiaries. Increased odds range from 2% higher odds of no acute inpatient spending for Black beneficiaries to 78% increased odds of no hospice spending for Asian/Pacific Islander beneficiaries. The results for AI/AN beneficiaries are not as consistent across these 4 categories. AI/AN beneficiaries have 27% and 17% lower odds of no hospital outpatient spending and no acute inpatient spending, respectively. However, the odds of no hospice spending increase by 39% for AI/AN beneficiaries. Results for Part D drug spending are not as consistent across race categories. The logit model predicting no Part D drug spending indicates that Black and AI/AN beneficiaries have 27% and 73% higher odds of no Part D drug spending, respectively, compared with White beneficiaries. Conversely, Asian/Pacific Islander and Hispanic beneficiaries have 64% and 17% lower odds of no Part D drug spending, respectively.
The second part of the ZIP regression model is the Poisson model that predicts the number of counts (ie, total spending) for those identified as the nonzero group. Again, the results are generally consistent across hospital outpatient, acute inpatient, skilled nursing, and hospice cost categories. With few exceptions, mean expected spending increases for Black, Asian/Pacific Islander, and Hispanic beneficiaries across hospital outpatient, acute inpatient, and skilled nursing categories, ranging from 2% for Hispanic beneficiaries to 22% for Black beneficiaries in the acute inpatient category. However, mean expected spending for hospice decreases for Black, Asian/Pacific Islander, and Hispanic beneficiaries, ranging from 1% lower for Black beneficiaries to 12% lower for Asian/Pacific Islander beneficiaries. Again, the results for AI/AN beneficiaries are not as consistent across categories. AI/AN beneficiaries have 2%, 10%, and 22% lower mean expected spending for hospital outpatient, acute inpatient, and hospice services, respectively, and less than 1% increased spending for skilled nursing. The Poisson model predicting Part D drug spending reveals that Asian/Pacific Islander and AI/AN beneficiaries have 14% and 19% greater expected mean spending for Part D drug services than White beneficiaries, respectively. Conversely, Black and Hispanic beneficiaries have 1% and 4% lower expected mean spending for Part D drug services than White beneficiaries, respectively.
Diabetes and MCC Combinations
Regarding T2D and MCC combinations, the logit model revealed that the odds of no spending for each category were significantly lower for every MCC combination group compared with having T2D alone (ie, T2D plus MCC increases the odds of having any payments). These decreased odds of no spending range from 4% lower for T2D plus cancer for Part D drug services to 95% lower for T2D plus CHF/COPD/stroke/CKD for skilled nursing.
When considering the MCC combinations, we generally observed that with every combination, mean acute inpatient, skilled nursing, hospice, and Part D drug spending increased. The increases ranged from less than 1% in hospice spending for T2D plus CHF/COPD/stroke/CKD to 142% in Part D drug spending for T2D plus CHF/COPD/stroke/CKD. Unexpectedly, for every MCC combination, hospital outpatient spending decreased compared with having diabetes alone, from 9% lower for T2D plus COPD to 27% lower for T2D plus CHF/COPD/stroke/CHD. On the face, these results seem surprising, and we discuss these in the context of the other dependent variables in the Discussion section.
Sex and Rural Patient Location
Logit models indicate that men have increased odds of no spending across all categories, except hospice. Increased odds of no spending for men range from 5% for acute inpatient services to 62% for Part D drug services. Male sex increases hospital outpatient and acute inpatient spending by 5% and 11%, respectively. Male sex was associated with 2%, 22%, and 8% reduced mean expected skilled nursing, hospice, and Part D drug spending, respectively.
Beneficiaries living in rural counties have 47% and 7% lower odds of no hospital outpatient and Part D drug spending, respectively. Rural residents also have 5% and 9% higher odds of no acute inpatient and skilled nursing spending, respectively. Rural patient location increases mean expected hospice spending by 18% but reduces mean expected spending across all other service categories.
The results of the ZIP analyses show several general trends related to T2D and combinations of chronic conditions, race, and sex and their influence on the odds of having no spending, as well as the mean spending in different service categories.
Combinations of T2D and MCC decrease the odds of having no spending. In other words, compared with having T2D alone, having T2D plus MCC increases the odds of having spending in the categories we studied. This result was not unexpected. Increased numbers of chronic conditions and specific combinations of conditions make it more likely that a patient would either seek care or need increased levels of care, thereby increasing spending. The pattern of lower odds of no spending with an increase in the expected mean of actual spending was consistent across most of the T2D plus MCC combinations.
The one notable exception to increased expected mean spending is for hospital outpatient services, for which we observed lower payments across all MCC combinations. Hospital outpatient services include physician services, outpatient care, laboratory testing, and preventive services. Our results indicate that for MCC combinations, expected mean hospital outpatient spending decreases compared with having T2D alone. The presence of T2D and MCC can reduce the number of face-to-face physician visits while increasing utilization of other services.16 Our results provide additional evidence that in Medicare beneficiaries with T2D, additional MCC reduced the odds of no payments but also reduced expected mean hospital outpatient spending. Although we have no evidence from this study to indicate that the reduction in mean hospital outpatient spending indicates that patients receive care elsewhere, possibly reflecting the need for higher-acuity services based on MCC profile, it may be a fruitful hypothesis for future research.
When controlling for T2D plus MCC combinations, we observed differences between White beneficiaries and racial/ethnic minority groups. In general, compared with White beneficiaries, racial and ethnic minority groups experience increased odds of no spending, but when costs were incurred, expected mean payments increased for hospital outpatient, acute inpatient, and skilled nursing services. One reason for this may be that racial and ethnic minorities with diabetes do not use primary care services at the same rates as White patients with diabetes, which may result in the eventual need for higher-acuity care.17 In addition, although we did not have information about beneficiary income, the correlation of low income with minority status may prevent some beneficiaries from seeking care due to uncertainty over how to pay for deductibles and co-payments, even with Medicare.18
We observed increased odds of no hospice payments across all racial and ethnic minority groups, and contrary to the results for outpatient, inpatient, and skilled nursing services, we observed a reduction in expected mean spending compared with White beneficiaries. These results support those found in the wide range of literature describing barriers to hospice care and general distrust of end-of-life care by racial and ethnic minority groups.19,20
Finally, except for hospice spending, male sex was associated with increased odds of no spending. This is unsurprising, as men are less likely to go to the doctor, to take their medications, and to need or want skilled nursing or hospice care.21 In addition, regarding rural residency, we observed lower odds of no hospital outpatient spending, perhaps indicating that these beneficiaries get their care from a primary care physician, but we also observed lower mean hospital outpatient spending compared with urban residents, perhaps reflecting lower service utilization and access compared with urban residents.22
One limitation of this research is that we did not analyze specific patterns of service use. However, our group is conducting other research that considers utilization levels for inpatient, outpatient, and emergency department services and disparities among racial/ethnic groups and among groups based on T2D and MCC combinations. In the interest of space, we did not include interactions (which may differentially affect outcomes for groups based on combinations of characteristics) between MCC and our demographic variables, but additional interaction analyses using hospital outpatient spending as an example, shown in eAppendix Table 2, indicate very few significant interactions as predictors of no spending or mean expected spending. Another limitation is the large sample size. For each of the ZIP analyses, we had a sample size of 443,932. With such large sample sizes, almost any effect size will show significance. Statistical significance does not equate to clinical significance, and small effect sizes are suspect. However, the increased (or decreased) odds of no spending range much higher than 10% to 20%, and the increased (or decreased) expected mean total spending also are much higher than 10% to 20% and in some cases extend to 100% to 200% greater. These effects would have been statistically significant even at much lower sample sizes. In addition, the same general trends for relationships hold across the service categories we included in this study. That said, we caution the reader about making conclusions related to small effect sizes. Next, the focus of this study is on Michigan, which could limit the ability to make conclusions about national trends. However, the LCD studied here mirror national trends. In addition, just like other areas of the South and Midwest, where diabetes prevalence and incidence are greater than in other parts of the country, Michigan has higher rates of diabetes risk factors such as smoking, overweight and obesity, physical inactivity, and high blood pressure, among others, which allows for comparison between those areas and Michigan. We also recognize that the removal of some data that lacked race data could affect our results. The race variable for some beneficiaries in the CMS database is generated using an algorithm that takes information from the Social Security Administration (SSA) and is based on a first or last name that was determined to likely be Hispanic or Asian in origin. In other words, people who were categorized as other or unknown by the SSA could be reclassified by the CMS algorithm solely on the basis of their first or last name, but only for first or last names that are likely Hispanic or Asian in origin and only for people who had unknown or missing race data to begin with. We did conduct a limited sensitivity analysis for hospital outpatient spending as an example, which included these beneficiaries to determine their influence on the dependent variables, as well as how they might influence the impact of other race and ethnicity groups. The inclusion of this group with missing data did not materially affect the magnitude or direction of the influence of the other variables (eAppendix Table 3). However, because of the uncertain meaning of this category, we ultimately decided to report only the analyses that excluded these beneficiaries. Finally, we are unable to control for socioeconomic status (SES), as there is no measure of income in the administrative claims data that we obtained. Although SES includes more than the concept of income, no other demographic variables in the claims database would allow us to develop a composite measure of SES. Finally, although linking to other national data sets might have allowed us to expand the number of individual-level covariates, we did not have access to beneficiary identifiers that would allow us to make such linkages.
We used 2012 Medicare claims data from beneficiaries in Michigan 65 years and older with T2D to determine if there are differences in spending across several service categories. Generally, beneficiaries with T2D plus MCC combinations have lower odds of no spending (ie, increased odds of any spending) and increased mean expected spending compared with beneficiaries with diabetes alone. In addition, minority status significantly predicts no spending and generally predicts increased expected mean costs when these beneficiaries seek services, even when controlling for T2D plus MCC combinations. In other words, considering the range of T2D plus MCC combinations and expecting that more complex cases would require the same level of service utilization, we still observed differences in spending among racial and ethnic minority groups and White beneficiaries. Our results highlight the ongoing need to dedicate attention and resources to make changes in the health care system that remove barriers to racial and ethnic minority groups to seek care.
Mr Killingsworth, Mr Weiss, Mr Husain, Ms Zunnu Rain, Ms Clark, Ms Alla, and Ms Jin contributed equally to this work.
Author Affiliations: College of Medicine, Central Michigan University (JMC, NPR, JK, TGW, SAH, MZR, KNC, SA, MGJ), Mount Pleasant, MI; College of Human Medicine, Michigan State University (JMC), Flint, MI.
Source of Funding: Grant No. 1R15DK104260-01A1 (Disparities in Diabetes Comorbidities and Multiple Chronic Conditions) from the National Institutes of Health/National Institute of Diabetes and Digestive and Kidney Diseases (NIH/NIDDK).
Author Disclosures: Drs Clements and Ragina report receiving NIH/NIDDK grant No. 1R15DK104260-01A1. The remaining authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.
Authorship Information: Concept and design (JMC, NPR, TGW, SAH, KMC, SA); acquisition of data (JMC, TGW); analysis and interpretation of data (JMC, JK, TGW, MZR, KMC, SA, MGJ); drafting of the manuscript (JMC, NPR, JK, MZR, KMC, SA, MGJ); critical revision of the manuscript for important intellectual content (JMC, NPR, JK, SAH, MZR, KMC, MGJ); statistical analysis (JMC, KMC); provision of patients or study materials (TGW, SAH); obtaining funding (JMC); administrative, technical, or logistic support (JMC); and supervision (JMC, NPR).
Address Correspondence to: John M. Clements, PhD, College of Human Medicine, Michigan State University, 130 E 2nd St, Flint, MI 48502. Email: firstname.lastname@example.org.
1. American Diabetes Association. Economic costs of diabetes in the U.S. in 2017. Diabetes Care. 2018;41(5):917-928. doi:10.2337/dci18-0007
2. Zhou X, Zhang P, Horger TJ. Lifetime direct medical costs of treating type 2 diabetes and diabetic complications. Am Prev Med. 2013;45(3):253-261. doi:10.1016/j.amepre.2013.04.017
3. Li R, Bilik D, Brown MB, et al. Medical costs associated with type 2 diabetes complications and comorbidities. Am J Manag Care. 2013;19(5):421-430.
4. Jiang HJ, Andrews R, Stryer D, Friedman B. Racial/ethnic disparities in potentially preventable readmissions: the case of diabetes. Am J Public Health. 2005;95(9):1561-1567. doi:10.2105/AJPH.2004.044222
5. Meng J, Casciano R, Lee YC, et al. Effect of diabetes treatment-related attributes on costs to type 2 diabetes patients in a real-world population. J Manag Care Spec Pharm. 2017;23(4):446-452. doi:10.18553/jmcp.2017.23.4.446
6. Booth GL, Hux JE. Relationship between avoidable hospitalizations for diabetes mellitus and income level. Arch Intern Med. 2003;163(1):101-106. doi:10.1001/archinte.163.1.101
7. Walker RJ, Smalls BL, Campbell JA, Strom Williams JL, Egede LE. Impact of social determinants of health on outcomes for type 2 diabetes: a systematic review. Endocrine. 2014;47(1):29-48. doi:10.1007/s12020-014-0195-0
8. Brandle M, Zhou H, Smith BRK, et al. The direct medical cost of type 2 diabetes. Diabetes Care. 2003;26(8):2300-2304. doi:10.2337/diacare.26.8.2300
9. Cheng JS, Tsai WC, Lin CL, et al. Trend and factors associated with healthcare use and costs in type 2 diabetes mellitus: a decade experience of a universal health insurance program. Med Care. 2015;53(2):116-124. doi:10.1097/MLR.0000000000000288
10. Simmons M, Bishu KG, Williams JS, Walker RJ, Dawson AZ, Egede LE. Racial and ethnic differences in out-of-pocket expenses among adults with diabetes. J Natl Med Assoc. 2019;111(1):28-36. doi:10.1016/j.jnma.2018.04.004
11. Harris MI. Racial and ethnic differences in health care access and health outcomes for adults with type 2 diabetes. Diabetes Care. 2001;24(3):454-459. doi:10.2337/diacare.24.3.454
12. Lee JA, Liu CF, Sales AE. Racial and ethnic differences in diabetes care and health care use and costs. Prev Chronic Dis. 2006;3(3):A85.
13. Number of deaths and age-adjusted mortality rates for the ten leading causes of death, Michigan 2018 and United States residents, 2017. Michigan Department of Health & Human Services. Accessed May 29, 2019. http://www.mdch.state.mi.us/pha/osr/deaths/causrankcnty.asp
14. Lambert D. Zero-inflated Poisson regression with an application to defects in manufacturing. Technometrics. 1992;34(1):1-14. doi:10.2307/1269547
15. Long JS, Freese J. Regression Models for Categorical Dependent Variables Using Stata. 2nd ed. Stata Press; 2005.
16. Lin PJ, Kent DM, Winn A, Cohen JT, Neumann PJ. Multiple chronic conditions in type 2 diabetes mellitus: prevalence and consequences. Am J Manag Care. 2015;21(1):e23-e34.
17. Taylor YJ, Spencer MD, Mahabaleshwarkar R, Ludden T. Racial/ethnic differences in healthcare use among patients with uncontrolled and controlled diabetes. Ethn Health. 2019;24(3):245-256. doi:10.1080/13557858.2017.1315372
18. Keeler EB. Effects of cost sharing on use of medical services and health. J Med Pract Manage. 1992;8:317-321.
19. Washington KT, Bickel-Swenson D, Stephens N. Barriers to hospice use among African Americans: a systematic review. Health Soc Work. 2008;33(4):267-274. doi:10.1093/hsw/33.4.267
20. Cohen LL. Racial/ethnic disparities in hospice care: a systematic review. J Palliat Med. 2008;11(5):763-768. doi:10.1089/jpm.2007.0216
21. Pinkhasov RM, Wong J, Kashanian J, et al. Are men shortchanged on health? perspective on health care utilization and health risk behavior in men and women in the United States. Int J Clin Pract. 2010;64(4):475-487. doi:10.1111/j.1742-1241.2009.02290.x
22. Larson SL, Fleishman JA. Rural-urban differences in usual source of care and ambulatory service use: analyses of national data using urban influence codes. Med Care. 2003;41(7 suppl):III65-III74. doi:10.1097/01.MLR.0000076053.28108.F2