Medicare hospital payment adjustments and nursing wages

Abstract

Despite the importance of the nursing profession for healthcare delivery, costs, and quality, there is relatively little research on how provider payments to hospitals affect the labor market for nurses. This study deals with the hospital wage index (HWI) adjustment to Medicare hospital payments, an area-level adjustment intended to compensate hospitals in high-cost labor markets. Since the HWI adjustment is based on hospital-reported labor costs, some argue that it incentivizes hospitals in concentrated markets to pay higher wages to nurses and other workers (the “circularity” critique). We investigate this critique using market-level data on the relative wages reported by nurses and hospital-level data on the average hourly wage for healthcare workers. For identification, we exploit a 2005 change in the geographic area used to define labor markets, which resulted in exogenous changes in the ability of some hospitals to influence their area’s wage index. We find that worker-reported relative nurse wages and hospital-reported healthcare worker wages are higher in some locations where hospitals experienced increased opportunities to game the circularity of the wage index, but these effects appear to be driven by pre-existing wage growth. Medicare’s HWI adjustment method does not appear to suffer from inefficiency due to circularity.

Appendices

Appendix

See Tables 5, 6, 7 and 8.

Table 5 Descriptive statistics for sample of 419 CBSAs

Table 6 Area-level analysis first stage results dependent variable is change in Actual HHI, 2000 to 2008

Table 7 Descriptive statistics for hospital sample

Table 8 Hospital-level analysis first stage results; all hospitals that did not reclassify, weighted by beds (n=2170)

Data Appendix: Details on variable construction and linkage by geography

1. Relative nurse wages

Area-level relative nurse wages were constructed from respondent-level census data as described in the text. We calculate the hourly wage (W) as the annual wage and salary income from the prior year divided by the product of weeks worked and usual hours per week. The 2008 ACS weeks worked variable is recorded in bins (1–13, 14–26, 27–39, 40–47, 48–49, and 50–52 weeks). We merge in a crosswalk created with the pooled 2005–2007 ACS samples that include this interval-valued variable and also its continuous-valued companion variable. For each weeks worked bin in the ACS, we impute the average number of weeks reported by 2005–2007 ACS respondents in the same bin.

We drop observations with imputed values for income, occupation, or industry and we drop outlier observations with hourly wages in the highest 1% and lowest 1% of the distribution in each sample (2000 and 2007–2008). We control for various determinants of wages in X, including years of potential labor market experience and indicators for sex, race/ethnicity, region, marital status, education, public sector employment, and part-time. We also control for the contemporaneous unemployment rate in the state using the 1999 and 2006–2007 annual average unemployment rates for each state reported by the Bureau of Labor Statistics (BLS, various years).

Areas were defined as CBSAs in both the 2000 Census and the pooled 2007–2008 ACS. Each CBSA is either a metropolitan or a micropolitan area. We follow CMS definitions and treat micropolitan CBSAs in the same state as part of a single “rest of state” area. In addition, eleven large CBSAs are divided into Metropolitan Divisions. CMS treats Metropolitan Divisions within a CBSA as separate areas, so we do the same. For example, there is a single CBSA code for Seattle, Tacoma, Bellevue, WA, but this large area is subdivided into Seattle–Bellevue–Everett, WA and Tacoma-Lakewood, WA. Finally, some parts of the country are not in CBSA-defined areas, and we count each of these as a part of its “rest of state” area.

Our original crosswalk file from MABLE/Geocorr2k (Missouri Census Data Center 2010) defined 940 CBSAs. Some respondents in our sample resided in areas outside these CBSAs, so we assigned them by state to 46 rest-of-state areas. In addition, 574 CBSAs are micropolitan CBSAs that we assign to 47 rest-of-state areas. These two sets of rest-of-state areas contained different states, so the total number of rest-of-state areas created is 48. 11 CBSAs are subdivided into 29 Metropolitan Divisions. Our original crosswalk file uses the CBSA definitions from 2010, but our wage data are from 2007 and 2008. Therefore, we adapt the crosswalk to reverse changes that occurred between 2007 and 2010. This subtracts 5 CBSAs and adds 2 CBSAs to the crosswalk. Thus, we measure relative nurse wages in 429 CBSAs.²⁶ We exclude 8 Maryland CBSAs (where hospitals did not participate in the IPPS) and 2 CBSAs where there were no IPPS hospitals in 2007 (and therefore without an HHI value for that year). Our analysis sample includes the remaining 419 CBSAs.

Individual respondents were assigned to CBSAs based on the PUMA of residence. We used two versions of the MABLE/Geocorr2k crosswalk (Missouri Census Data Center 2010) between PUMAs and CBSAs; one version described overlaps between PUMAs and CBSAs with the population distribution in 2000, and the other used the 2009 population distribution. In cases where the PUMA matched to a single CBSA, this is a simple assignment. However, some PUMAs overlap multiple CBSAs. For respondents in these PUMAs, we duplicated their observations to include one in each overlapping CBSA. When we estimate the relative nurse wage regression, we weight each observation by the population overlap between that observation’s original PUMA and the CBSA to which it was assigned. For example, suppose a PUMA shares 80, 15, and 5% of its population with CBSAs A, B, and C, respectively. Then each respondent in that PUMA becomes three observations in our regression sample, each assigned to a CBSA (A, B, and C). The respondents assigned to CBSA A, B, and C received a regression weight of 0.85, 0.15, and 0.05, respectively. In our samples, 87% of respondents live in a PUMA that overlaps a single CBSA; among the remaining 13% of respondents, the average PUMA population share in the most-overlapping CBSA is 71%.

In Tables 9 and 10, we illustrate the number of observations used in the construction of CBSA-level nurse wages from the 2000 Census and the pooled 2007–2008 ACS. Our concern is the precision with which we estimate relative nurse wages at the CBSA (location) level. The average location-specific relative nurse wage relies on data from 157.2 RNs and 1312 other workers in the 2000 Census, with somewhat smaller sample sizes for the pooled 2007–2008 ACS. While there are CBSAs with quite small samples of RNs (as few as 5), most CBSAs rely on much larger samples. We believe these are sufficient sample sizes to have confidence in our measures of relative nurse wages across locations. The sample sizes in Table 10 support the pooling of multiple ACS surveys: CBSA-specific sample sizes would be quite small for individual years.

Table 9 Summary statistics for observation counts, by CBSA, 2000 census

Table 10 Summary statistics for observation counts, by CBSA, 2007–2008 ACS

2. Herfindahl-Hirschman Index (HHI)

We construct area-level values of the Herfindahl-Hirshman index using hospital-level data on the number of beds from the CMS Medicare IPPS Impact Files. Each year’s Impact File contains data collected in the summer preceding the start of the federal fiscal year (NBER 2014). For example, the 2000 Impact File includes data collected in the summer of 1999, since fiscal year 2000 began October 1, 1999. Therefore, to align the HHI measures with the relative nurse wage estimates obtained from Census data for 2000 and 2007–2008, we use data from the 2000 and 2007 Impact Files.

In each year, we construct area-level HHI estimates using the geographic area used by Medicare to construct hospital wage indices. In the 2007 Impact File, we define the HHI at the CBSA level. The 2007 Impact File does not contain the CBSA for hospitals in the Indian Health Service, so we assign CBSAs to these hospitals by merging the Impact File observations to a Mable/Geocorr county-to-CBSA crosswalk. Because this crosswalk uses FIPS county codes and the Impact File contains only SSA county codes, the merge requires that we first assign FIPS county codes to the Impact File Hospitals using a crosswalk from SSA county to FIPS county, available from NBER (2012). We then define the CBSA-level HHI by calculating each hospital’s share of beds out of the total beds at Impact File hospitals in the CBSA, then squaring the shares and summing the shares in the CBSA. In this same step, we also define CBSA-level measures of the bed share at the largest four hospitals and the largest hospital’s share of all beds.

In the 2000 Impact File, we define the HHI at the MSA level. Because the 2000 Impact File does not report the MSA for a small number of hospitals in the Indian Health Service, we first obtain missing county identification codes for these hospitals from the 1999 AHA annual survey (using a merge by hospital Medicare Provider Number, or MPN) and we then obtain missing MSA codes with a county-MSA crosswalk. We define the MSA-level HHI by calculating each hospital’s share of beds out of the total beds at Impact File hospitals in the MSA, then squaring the shares and summing the shares in the MSA. In this same step, we also define MSA-level measures of the bed share at the largest four hospitals and the largest hospital’s share of all beds.

This same 2000 MSA-level measure of HHI serves as year 2000 values of the simulated HHI. To obtain year 2007 values of simulated HHI, we use the 2000 Impact File, and assign each hospital to the CBSA it would belong to in 2007 using a county-to-CBSA crosswalk from NBER (2005), but updated with a change that took place in July 2006 (CBSA 46940 was renamed CBSA 42680). Once each hospital was assigned to a CBSA, we define the CBSA-level HHI using each hospital’s share of beds out of the total beds at Impact File hospitals in the CBSA (and by squaring and summing the shares), as well as CBSA-level measures of the bed share at the largest four hospitals and the largest single hospital.

However, prior to assigning hospitals to their 2007 CBSA, we first needed to assign a county code to each hospital in the 2000 Impact File, since no county code is included in the file. For this step, we obtained street addresses of hospitals in the 2000 HCRIS (Medicare Cost Report files) that could then be merged by Medicare Provider Number or MPN back to the 2000 Impact File. Note that the HCRIS data contain a county name, but not a county code. To obtain a county code, we geocoded hospital locations using Texas A&M geocoding services. The geocoding process yields a specific latitude and longitude for each hospital address and an indication of the level of the geocoded match (i.e., whether the hospital address was geocoded to an exact location, a line segment, a zipcode, etc.). We then used a spatial join procedure in which all points (i.e., hospitals) were linked to counties in the U.S. Census 2000 County Boundary files. The output from this spatial join procedure was inspected and a county code was assigned to each hospital according the following steps. First, if the county name obtained from the 2000 County Boundary file matched the county name in the HCRIS data (allowing for spelling errors or abbreviations), we used the county code from the 2000 County Boundary file. Alternatively if the county name was not identical in both sources, or was missing in the HCRIS data, but the geocoding occurred at the exact location or the Tiger/Line segment level, we used the county code from the 2000 County Boundary file. For those cases where the geocoding did not produce an exact match and where county name matches could not be used to confirm the county code, we used a 2000 Mable/Geocorr crosswalk between ZCTA and County to identify ZCTAs that lie entirely within a single county. If a hospital geocoded to the centroid of such a ZCTA, that hospital was assigned to the county code. For the small number of remaining hospitals that were not assigned following the steps above, the county code was confirmed either using address information from the 1998 HCRIS data, or using the hospital’s county code from the American Hospital Association (AHA) annual survey from 1999.

3. Assigning MSA-level HHI measures to 2000 CBSA-level relative nurse wages

Each CBSA-level relative nurse wage estimate from the 2000 Census was assigned an MSA-level HHI measure constructed from the 2000 Impact file with a crosswalk that maps each CBSA to a single MSA. This CBSA-to-single-MSA crosswalk was constructed for this project from a 2005 crosswalk available from NBER (NBER 2012) and based on CMS data. The NBER/CMS crosswalk lists both the CBSA and the MSA to which each U.S. county is assigned. As noted above, the NBER/CMS crosswalk uses the same county borders in New England that CMS uses in the geographic adjustment of hospital payments, and is thus preferred in those states over the OMB-defined MSAs used in Mable/Geocorr, which follow town and city borders in New England. Since the crosswalk reflects assignments as of 2005 and our goal was to assign CBSAs from 2007 to an MSA, the NBER/CMS crosswalk was first updated to reflect a change effective July 1, 2006 in which CBSA 46940 was renamed CBSA 42680.

The NBER/CMS crosswalk was then combined with 2000 county-level population counts from the U.S. Census Bureau (American FactFinder), and the population data were used to define the share of CBSA residents that lived in a given MSA for each possible CBSA-MSA match. The CBSA-to-single-MSA crosswalk was produced by assigning each CBSA to the MSA in which the greatest share of the CBSA’s residents lived. To illustrate this process, Table 11 shows five hypothetical counties that comprise a single CBSA (CBSA C1), four of which fall into MSA M1 and one of which falls into MSA M2. Since 85% of the CBSA residents lived in counties located within MSA M1, and only 15% lived in the county located within MSA M2, CBSA C1 is assigned to MSA M1.

Table 11 Illustration of CBSA-MSA crosswalk construction

Tables 12 and 13 provide additional details on the crosswalk. As shown in Table 12, some CBSAs overlapped with as many as five different MSAs, but a large majority of CBSAs (64%) overlapped with only one MSA, and another large share overlapped with only two MSAs (30%). Similarly, Table 13 shows that in 267 of the 419 CBSAs in our sample, 100% of CBSA residents lived in a single MSA. Another 78 CBSAs were matched to an MSA where more than 90% (but less than 100%) of residents lived, and another 52 CBSAs were matched to an MSA where more than 80% (but less than 90%) of residents lived.

Table 12 MSA-CBSA overlap

Table 13 Percent of the CBSA population residing in its matched MSA

We examined the group of CBSAs that matched to a single MSA more closely. This group consists of two types of CBSAs: (1) CBSAs that were formed from identical boundaries as a single MSA (these may be either urban or rest-of-state areas), and (2) CBSAs that were formed from a sub-area within a single MSA (again these may be either urban or rest-of-state areas). Of the 267 CBSAs that match to a single MSA, 151 are of the first type, and 116 are of the second type. This latter group is interesting for two reasons. First, the adoption of CBSAs (as opposed to MSAs) in defining the hospital market creates exogenous variation in the hospital concentration in the market. Second, because a new CBSA was formed from a subset of an MSA, we can be confident about linking the CBSA-level relative nurse wage in 2000 to the MSA-level market concentration measure in that year, since all the nurses residing in that CBSA were living in areas that were part of the MSA.

We also examined those CBSAs that match to more than one MSA (prior to our selection of a single MSA); this may arise for several reasons. In some cases, one urban CBSA may have formed from two or more urban MSAs; in other cases, one urban CBSA may have formed from one or more urban MSAs and part of a rest-of-state area. Alternatively, a rest-of-state CBSA may have been formed by combining part of a rest-of-state MSA with part or all of one or more urban MSAs. In these cases, the adoption of CBSAs in defining the hospital market also creates exogenous variation in hospital concentration. Because the new CBSA was formed from multiple MSAs, this introduces some noise into the assignment of the CBSA to a single MSA for the purposes of linking the CBSA-level relative nurse wage in 2000 to the MSA-level hospital concentration measure in that year. For this reason, we use the data on population overlap in Table 13 to conduct some sensitivity tests of our results, as described in the text.

4. Area-level bite measures

We use two measures to capture the importance of the Medicare HWI adjustment process to hospitals and workers in an area. The first measure is the share of all inpatient days in a CBSA that occur in IPPS hospitals and are paid by Medicare. The second is an index that reflects the degree to which Medicare days at IPPS hospitals are concentrated in the CBSA. Both measures are constructed from information in the 1999 AHA annual survey (as opposed to Impact File data) to account for the importance of the HWI among all hospitals in the CBSA. To construct each measure, we first identify those hospitals in the 1999 AHA survey that are IPPS hospitals. If the hospital has a valid MPN, we can identify IPPS status from its last four digits (between 0001 and 0879 for IPPS hospitals), per CMS documentation (DHHS 2001, p. 2–164). If the hospital had a missing MPN or an MPN equal to “777777,” we use other fields on the survey to select those hospitals that are “community” and “general medical and surgical” hospitals; for each of these hospitals, we obtained MPNs from an online search and used the last 4 digits of the MPN to designate them as IPPS hospitals or not. In three cases, we could not obtain the MPN through a manual lookup; we designated two of these three as IPPS hospitals because they did not appear on a list of Critical Access Hospitals (which are not paid according to IPPS rules). In addition to identifying IPPS hospitals, the construction of these measures also requires assigning each hospital in the 1999 AHA survey to the CBSA it would belong to in 2007. For this step, we use the same NBER/CMS crosswalk described above (after it was modified for the one CBSA change that took place in July 2006).

Once IPPS hospitals and CBSAs were identified, we constructed the Medicare-IPPS share as the total number of Medicare inpatient days occurring at IPPS hospitals in the CBSA divided by the total number of all inpatient days occurring at all hospitals in the CBSA. We constructed the Medicare-IPPS HHI by calculating for each hospital, the number of its Medicare inpatient days paid by IPPS as a share of all inpatient days occurring at all hospitals in the CBSA (this share is equal to 0 for non-IPPS hospitals), then squaring the shares, and summing the squared shares in each CBSA.