The Implications of Regional Variations in Medicare Spending. Part 2: Health Outcomes and Satisfaction with CareFREE

As described in greater detail in Part 1, we carried out a cohort study in four parallel populations using a natural randomization approach (11). In this approach, one or more exposure variables allow assignment of patients into treatment groups (different levels of average spending), as would a randomized trial. Because some of the regional differences in Medicare spending are due to differences in illness levels (enrollees in Louisiana are sicker than those in Colorado) and price (Medicare pays more for the same service in New York than in Iowa), we could not use Medicare spending itself as the exposure. We therefore assigned U.S. hospital referral regions (HRRs), and thus the cohort members residing within them, to different exposure levels using a measure that reflects the component of regional variation in Medicare spending due to physician practice rather than regional differences in illness or pricethe End-of-Life Expenditure Index (EOL-EI). Because regional differences in end-of-life spending are unrelated to underlying illness levels, it is reasonable to consider residence in HRRs with differing levels of end-of-life spending as a random event. The index was calculated as spending on hospital and physician services provided to a reference cohort distinct from the study cohorts: Medicare enrollees in their last 6 months of life. In the current paper, we also present several analyses with an alternative exposure measure, the Acute Care Expenditure Index (AC-EI), to decrease concern about possible residual confounding.

We confirmed that the exposures used to assign the HRRs achieved the goals of natural randomization: 1) Study samples assigned to different levels of the exposure [the EOL-EI] were similar in baseline health status, and 2) the actual quantity of services delivered to the individuals within the study samples nevertheless differed substantially across exposure levels and was highly correlated with average per capita Medicare spending in the HRRs. We followed the cohorts for up to 5 years after their initial hospitalizations and compared the processes of care (Part 1) and health outcomes (Part 2) across HRRs assigned to different exposure levels.

The four study cohorts are described in detail in Part 1. Briefly, we studied fee-for-service Medicare enrollees, ages 65 to 99 years, who were eligible for Medicare Parts A and B. The acute myocardial infarction (MI) cohort was drawn from patients included in the Cooperative Cardiovascular Project, who had index hospitalizations between February 1994 and November 1995. The hip fracture and colorectal cancer cohorts were identified based on an incident hospitalization between 1993 and 1995. The general population sample included participants in the Medicare Current Beneficiary Survey (MCBS) who had initial interviews between 1991 and 1996 (for the survival analysis) or between 1992 and 1995 (for the other analyses) (see Appendix, Section C).

Each cohort member was placed in a spending group according to the EOL-EI (as defined in detail in Part 1) in their HRR of residence at the time of the index hospitalization (chronic disease cohorts), or initial interview (MCBS cohort). Characteristics of the study cohorts were ascertained from a variety of sources, as described in detail in Part 1, including Medicare administrative files and claims (all four cohorts), chart reviews (acute MI cohort), in-person interview (MCBS cohort), U.S Census data (attributes of ZIP code of residence, such as income, for the three chronic disease cohorts), and American Hospital Association data (to characterize hospitals).

As we summarized here and described in detail in Part 1, we used two approaches to determine cohort members' exposure to different levels of Medicare spending in their HRR of residence. Previous research has shown that the dramatic differences in end-of-life treatment across U.S. regions are highly predictive of differences in total spending (8, 12) but are not due to differences in case mix or patient preferences (13). Our primary measure of exposure was the EOL-EI, which was calculated as age-sex-raceadjusted spending (measured with standardized national prices) on hospital and physician services provided to Medicare enrollees who were in their last 6 months of life in each of the 306 U.S. HRRs in mid-1994 to 1997, excluding any members of the study cohorts (Appendix, Section E) We also repeated the major analyses with an alternative exposure measure, the AC-EI, which was based on differences across HRRs in risk-adjusted spending during an acute illness episode (Appendix, Section F). Both measures were highly predictive of average age-sex-raceadjusted Medicare spending at the HRR level (r = 0.81 for the EOL-EI and 0.79 for the AC-EI in the acute MI cohort) and, as was shown in Part 1, of the regional differences in utilization experienced by the study cohorts. For many analyses, we grouped HRRs into quintiles of increasing exposure to the expenditure indices.

To assess the aggregate impact of any differences in individual attributes on average baseline risk for death across regions of increasing EOL-EI, we used logistic regression to determine each individual's predicted 1-year risk for death as a function of his or her baseline characteristics. The models had modest to excellent predictive ability (c-statistics were 0.61 for the colorectal cancer cohort, 0.68 for the hip fracture cohort, 0.77 for the acute MI cohort, and 0.82 for the MCBS cohort). We used these models to determine the average predicted risk for death across quintiles of Medicare expenditure indices.

Tables 1 to 4 in Part 1 present selected characteristics of each study cohort grouped into quintiles according to EOL-EI level in their HRRs of residence. Because the sample sizes are large, many small differences for the chronic disease cohorts were statistically significant. Notable differences were found in racial composition (more black persons in higher-expenditure HRRs) and income (higher-expenditure HRRs had more beneficiaries in the highest and lowest income categories). Smaller differences across quintiles were apparent in age, sex, comorbid conditions, and cancer stage. For the acute MI cohort, patients in the highest quintiles had a higher prevalence of nonQ-wave infarctions and congestive heart failure but a lower prevalence of creatine kinase levels greater than 1000 IU/L. For the MCBS cohort, residents of HRRs in the quintiles with higher EOL-EIs were more likely to report being in fair or poor health but were less likely to live in a facility.

Crude 30-day and 1-year mortality rates and average predicted 1-year mortality rates for each cohort are shown in Table 1. For the hip fracture cohort, average predicted mortality rates at 1 year were slightly but significantly lower in HRRs with a higher EOL-EI. In the acute MI cohort, however, average predicted mortality rates at 1 year were higher in HRRs with a higher expenditure index. No significant differences were found in predicted mortality across HRRs with differing expenditure indices for the colorectal cancer or MCBS cohorts. These findings reveal no consistent trend toward greater illness burden in HRRs with a higher expenditure index. Observed mortality tended to be lower than predicted in the lowest quintile and equal to or higher than predicted in the highest quintile.

Figure 1 presents the relative risk for death over 5 years for residents of HRRs in EOL-EI quintiles 2, 3, 4, and 5 (the higher quintiles) compared with residents of HRRs in the lowest quintile. In each cohort, an increase in EOL-EI was associated with a small increase in the risk for death. We repeated these analyses using the HRR-specific EOL-EI as a continuous variable both overall and in specific subgroups (Figures 2, 3, and 4). A relative risk greater than 1 indicated that residence in an HRR with a higher EOL-EI (higher expenditures) was associated with increased mortality. For every 10% increase in the EOL-EI, the relative risk for death over 5 years was as follows: hip fracture cohort, 1.003 (CI, 0.999 to 1.006); colorectal cancer cohort, 1.012 (CI, 1.004 to 1.019); acute MI cohort, 1.007 (CI, 1.001 to 1.014); and MCBS cohort, 1.01 (CI, 0.99 to 1.03). In none of the subgroups examined was a higher expenditure index associated with a statistically significantly lower mortality rate.

We repeated the mortality analyses using the alternate approach: assigning HRRs to different exposure levels based on the AC-EI. Residents of higher-spending HRRs, according to the AC-EI, had relatively similar baseline health status (Appendix Table 17) and yet received substantially more care (Appendix Table 18). The results of the mortality analyses are summarized in Table 2. For the hip fracture cohort, higher AC-EIs were associated with a small decrease in mortality rates. For all of the other cohorts, mortality rates did not differ or increased slightly in regions with a higher AC-EI.

The average decline in functional status, as measured by using the HALex score, was about 2 points per year (on a 100-point scale) but did not differ across HRRs grouped according to quintiles of the EOL-EI (Table 3). In none of the models examined was an increased expenditure index associated with a statistically significant difference in the average rate of decline in health status (Appendix Table 10).

Figure 5 presents average change in adjusted satisfaction scores across quintiles (compared with quintile 1) for the five summary scales. Each scale ranges from 0 to 100, with higher scores implying greater satisfaction. We found substantial variation in satisfaction with care across the nine major U.S. regions (for example, Northeast and Mid-Atlantic), with satisfaction on each scale averaging over five points higher in the Northeast than in the South, controlling for other factors (data not shown). The differences in satisfaction across EOL-EI quintiles, however, were smaller than these regional differences and did not reveal a consistent pattern of greater satisfaction in HRRs with a higher expenditure index. The overall test for trend across HRRs indicated less global satisfaction with care and more satisfaction with interpersonal aspects of care in higher-spending HRRs. No differences were found across HRRs of differing expenditure indices for the other three measures of satisfaction with care.

The Appendix was developed to provide interested readers with additional detail on the methods of the study as well as supplementary findings referred to in the body of the papers that could not be included there because of space constraints. Section B provides an expanded discussion of the rationale for our study design and its relationship to instrumental variables analysis. Section C describes in greater detail our study populations, exclusions applied, and data quality. Section D describes in detail the rationale behind the approach and the methods used to calculate spending and utilization rates using measures free of bias that could be introduced because of differences in wages, prices, or policy payments to physicians or hospitals. Section E describes in greater detail the End-of-Life Expenditure Index (EOL-EI), the primary exposure used in the analysis, including the study population within which it was calculated and how members of each study cohort were excluded from the sample used to calculate the index used as the exposure for that cohort. Section F describes the motivation, methodology, and results of our sensitivity analysis using the Acute Care Expenditure Index.

In addition, the Appendix also includes supplementary tables that present additional detail on individual patient attributes (Appendix Tables 1, 2, 3, and 4), a table that lists specifically which variables are included in each of the major models used in the analyses (Appendix Table 5), the main models examining survival (Appendix Tables 6, 7, 8, and 9) and change in functional status (Appendix Table 10), a table presenting specific procedure rates for each chronic disease cohort and for all three cohorts combined (Appendix Table 11), and tables summarizing overall health care utilization rates across quintiles for each chronic disease cohort (Appendix Tables 12, 13, and 14).

As is discussed in the overview of the study design in Parts 1 and 2, the ideal approach to addressing the study questionwhether the increased spending observed in some regions of the United States leads to better care or outcomeswould be to carry out a randomized trial. However, such a trial would be difficult and would probably end up answering a slightly different question (depending on the intervention under study).

The field of economic research has addressed this problem through approaches that attempt to create a natural randomization through what is termed instrumental variables analysis. The key notion is that an exposure is identified that allows the study sample to be assigned to different treatment groups in a way that assures that those in different treatment groups are similar in terms of attributes that might affect the outcome (that is, that case mix is similar in the groups). They are nonetheless treated differently.

A good example of this type of natural randomization comes from a study of how serving in the Vietnam War affected the probability of suicides and vehicular deaths (26). Clearly, comparing suicide rates for Vietnam veterans and nonveterans would be statistically suspect, since the underlying characteristics of the two groups would be expected to differ by so much. Draft lottery numbers, chosen randomly on the basis of ones birthday, were used as a natural randomization to place men into the treatment group, those most likely to be sent to Vietnam, and the control group, those least likely to be sent. This method qualified as an instrument because it fulfilled the two [intuitive] requirements of an instrumental variable: 1) It was highly correlated with the exposure variable, which was serving in Vietnam, and 2) it was plausibly uncorrelated with the underlying mental health of the population (or, more formally, with any unmeasured differences in the populations). In other words, any differences in suicide and accident rates between the two groups were very likely to have been the result of serving (or not serving) in Vietnam, and not individual risks for suicide or poor driving. The article by Hearst and colleagues, like our articles, took a reduced-form approach to the problem. In other words, they compared what they called draft eligible (the treatment group) with draft ineligible (the control group).

By the same token, in our papers, we compared outcomes of people living in areas where the health system displays a more aggressive approach to end-of-life care with those of people living in areas where the health system displays a less aggressive approach. We have no a priori reason for believing that these populations in these regions should differ in their underlying health status, but they are treated differently.

Why didnt we use the formal instrumental variables approach, in which we would predict how much an additional $1000 in Medicare spending affects survival? There are three main reasons. First, we are interested primarily in the direction and general magnitude of effect, rather than in the cost of achieving that effect. We recognize that if increased expenditures across regions result in improved health outcomes, knowing the magnitude of the effect of an additional 10% increase in regional spending on survival and functional status for Medicare patients would be important for policy research. If we find no association or that higher spending is associated with lower survival, however, the precise estimate of the coefficient (in terms of dollars) is relatively unimportant. Second, instrumental variables analysis is able to provide unbiased estimates only in certain settings, one of which is a linear model. Our need to use Cox proportional-hazards regression for our mortality analyses precluded a formal instrumental variables analysis using currently developed statistical tools. Finally, it is important to recognize that the fundamental limitation of instrumental variables analysis would remain. One cannot prove that one has a perfect instrument.

We therefore presented our analysis as an observational study. We recognize that unmeasured confounding remains a possibility, but we nevertheless believe that our findings represent a major advance over previous research and that our conclusions that residence in higher-spending regions does not cause improved quality, access to care, or survival (and may cause worse survival) are sound.

For all three study cohorts, we restricted the eligible population to Medicare enrollees between the ages of 65 and 99 years who, at the time of their index admission, were eligible for both Medicare Parts A and B and were not enrolled in a health maintenance organization (HMO).

All of our utilization analyses in which dollar amounts are reported were based on measures of expenditures that have been purged of regional differences in prices or policy payments because the use of actual payments would introduce a bias. Actual reimbursements for hospital and physician services vary substantially according to geographic region due to wage, price, and policy differences (such as subsidies for the costs of medical education). To develop a measure of Medicare spending that was free of regional differences in price and policy payments, we followed the general approach developed by the Medicare Prospective Payment Commission in an earlier report (43) to calculate spending as follows. For inpatient hospital services, we based our measure on the diagnosis-related group (DRG) weight. All DRGs are assigned a relative weight proportional to the average national cost for Medicare patients within that DRG compared to the average cost for all Medicare patients. We converted DRG weights to dollars by multiplying the weight times the national average DRG price for 1996 ($3799). The measure reflects average national resource use for this condition. Hospital spending was defined as the sum of all DRG weights for an individual during a specified period times the DRG price. For physician services, we used the Resource-Based Relative Value Scale that forms the basis of the current Medicare physician fee schedule (44). Relative value units (RVUs) are assigned to each physician service to reflect physician work and the associated practice expense. For services included in the physician fee schedule, we assigned the total RVU value for the specific service from the Medicare fee schedule. For services not included in the fee schedule (primarily laboratory services), we calculated an RVU equivalent by dividing either the standard national price (laboratory services) or the median national allowed charge (for physician services without an RVU in the fee schedule) by the average 1996 factor ($36.14) used to convert RVUs to dollars. When DRG weights and RVUs are used, the measure of spending treats the value of a given service equally regardless of where the service is performed in the country. The measure removes the effect of any geographic differences in prices, wages, and policy payments.

Physician spending was defined as the sum of all RVUs for a given beneficiary during a specified period times the conversion factor. Aggregate spending for an individual is calculated in dollars and equals the sum of hospital spending and physician spending.

Because of concern that our primary exposure (the EOL-EI) may not have fully accounted for differences in population characteristics in different regions, we developed an alternative measure and repeated the analyses using this measure. Although the ideal measure would be risk-adjusted differences in total Medicare spending, we know of no way to calculate such a measure using currently available data. An alternative was to define study populations in which we were reasonably confident in our case-mix measures. Given the probable similarity of the cohorts at baseline across regions, and the high quality of the risk-adjustment data for short-term mortality (for example, 6 months), we decided to use as our alternative exposure measure the regional differences in risk-adjusted 6-month utilization in the complementary cohorts as our measure of the exposure. We describe this approach, and our findings, in the sections that follow.