The Latest on Geographic Variation in Medicare Spending: A Demographic Divide Persists But Variation Has Narrowed

The primary data source for this analysis is the February 2015 update of the Medicare Geographic Variation Public Use File (GV PUF) from the Centers for Medicare & Medicaid Services (CMS).¹ The GV PUF includes data for the U.S. overall and for three different geographic levels: state, county, and hospital referral region (HRR). We analyzed data at the county level because that is the most-granular level available. Our analysis focuses on the years 2007 and 2013, the earliest and latest years, respectively, in the 2015 GV PUF.

The GV PUF reports spending and beneficiary characteristics only among Medicare beneficiaries enrolled in both Parts A and B (i.e., not just one or the other), and in the traditional Medicare program (i.e., excluding beneficiaries enrolled in a private Medicare Advantage plan). Because of increases in Medicare Advantage enrollment, the share of Medicare beneficiaries included in the GV PUF study population declined from 71 percent in 2007 (33.0 million out of 46.7 million) to 62 percent in 2013 (34.3 million out of 55.2 million).

The GV PUF include three types of spending measures: “actual costs” (unadjusted Medicare payments, excluding beneficiary cost sharing and third-party payments), “standardized costs” (that is, a simulated measure of costs calculated by CMS using a single national price schedule that reflects the quantity and intensity of services but does not include market- or provider-level price adjustments), and “standardized risk-adjusted costs” (that is, standardized costs adjusted for differences in health risk by dividing by the HCC (hierarchical condition categories) score). (In this paper, we refer to “standardized costs” as price-adjusted spending, and “standardized risk-adjusted costs” as price- and risk-adjusted spending.) In addition, for each combination of county, year, and service category we created a set of Medicare price indexes, equal to the ratio of actual costs over standardized costs. In a county with prices equal to the national average, the price index equals 1.00. These price indexes reflect the market-level and provider-specific adjustments that are applied in each county in Medicare’s price-setting formulas, such as differences in local wages and certain provider characteristics such as add-on payments for teaching hospitals. In our discussion of spending by service category, we focus on spending per capita since this measure is the product of the percent of beneficiaries using each service, the Medicare price index for that type of service, and the standardized (i.e., price-adjusted) costs per user.

The county-level GV PUF include 3,136 counties, but many of those counties contain relatively few Medicare beneficiaries. Medicare per capita spending in those small counties can vary widely from county to county and from year to year due to random beneficiary-level variation. To avoid mistakenly identifying counties as high- or low-spending based on a few individual outliers, our analysis included only the 736 counties with an average of 10,000 or more traditional Medicare beneficiaries from 2007 through 2013. These 736 counties included 25.9 million traditional Medicare beneficiaries in 2013, which represents 76 percent of the GV PUF population in that year. The national averages discussed in our analysis include all counties, regardless of the number of beneficiaries in any given year.

We merged the county-level GV PUF with county-level measures of population: the poverty rate among county residents of all ages and the supply of health care providers from the Area Health Resources Files (AHRF) produced by the Health Resources and Services Administration (HRSA).² For some measures of population demographics and supply, the most recent data available in the AHRF was prior to 2013. To calculate county-level supply measures, we divided the number of providers (e.g., active physicians) by the total number of county residents (i.e., not just Medicare beneficiaries). To calculate national average measures of supply, we first summed supply and population nationwide, and then calculated supply per capita.

We analyzed county-level data on the share of traditional Medicare beneficiaries with multiple chronic conditions, using a 5 percent sample of claims from the 2013 CMS Chronic Conditions Data Warehouse (CCW). We calculated the percent of beneficiaries living in metropolitan areas, based on our analysis of the AHRF. We also analyzed county-level data from the American Community Survey (ACS) 2008-2012 five-year pooled data from the 2013-2014 AHRF on the share of county residents ages 65 and older living in poverty. We matched variables from these separate data files to the GV PUF using the five-digit Federal Information Processing Standard (FIPS) codes, which uniquely identify counties and county-level equivalent areas in the U.S.

We ranked the 736 counties in our analysis using three different spending measures: 1) unadjusted Medicare per capita spending in 2013, to show how counties rank on their actual Medicare per spending levels, regardless of price and health risk differences; 2) price- and health-risk adjusted Medicare per capita spending in 2013; and 3) the annual rate of growth from 2007 to 2013 in unadjusted Medicare per capita spending.³ Our measure of county-level price- and health-risk adjusted spending equals price-adjusted spending divided by the mean HCC score.

For each of these three spending measures, we identified the 20 counties at the top and bottom of the rankings, which yielded six sets of 20 counties. We tested whether our results were sensitive to the selection of 20 counties in each group by replicating our analysis for groups of 50 counties. The results were not appreciably different. For each of those six sets of counties, we calculated beneficiary-weighted averages of spending and utilization measures and demographics, which gives greater weight to counties with larger numbers of traditional Medicare beneficiaries.

Our analysis compares these groups of counties in terms of Medicare beneficiary characteristics, health care provider supply measures, spending and utilization measures, and county-level poverty:

• Beneficiary characteristics include percent black, percent Hispanic, percent eligible for both Medicare and Medicaid, and health risk scores (HCCs) from the GV PUF; the percent of traditional Medicare beneficiaries with five or more chronic conditions, based on our analysis of a five percent sample of Medicare claims from the 2013 CCW; and the percent of county residents ages 65 and over living in poverty, based on our analysis of ACS 2008-2012 five-year pooled data from the 2013-2014 AHRF, and the percent of beneficiaries living in metropolitan areas, based on our analysis of the AHRF.
• Measures of health care provider supply are reported per 10,000 county residents and include the number of physicians, primary care physicians as a percent of all physicians, hospital beds, skilled nursing facility beds, home health agencies, ambulatory surgical centers, and hospices. Supply measures are from our analysis of the AHRF for various years.
• Measures of spending and utilization from the GV PUF include the percent of beneficiaries using specific Medicare-covered services (hospital inpatient, outpatient, evaluation & management, procedures, skilled nursing facility, home health, durable medical equipment, and hospice), spending per capita and per user, and event counts (days, visits, procedures) per 1,000 beneficiaries.
• County-level poverty from the AHRF is the poverty rate among county residents of all ages.

We used the “coefficient of variation” (COV) as a summary measure of the amount of geographic variation in county-level Medicare per capita spending to examine whether geographic variation increased or decreased between 2007 and 2013. The COV equals the beneficiary-weighted standard deviation in Medicare per capita spending divided by the beneficiary-weighted mean. We included all 3,136 counties in the COV calculation. The COV is useful for measuring trends in spending variation, while adjusting for inflation.

Our analyses use the entire population of U.S. counties and the entire population of Medicare fee-for-service beneficiaries within those counties, rather than a random sample. Although conventional tests of statistical significance are not necessary in this situation, we present results from statistical significance testing for the comparisons of county-level averages in the appendix tables. For these tests, we used the TTEST procedure in SAS, weighted by the number of beneficiaries and assuming unequal variance.