To ensure, however, that our longitudinal Medicare data accurately reflect aggregate Medicare spending, we conducted a series of analyses to validate Medicare enrollment and spending calculated from claims against published statistics reported by the CMS in the annual statistical supplement of the Health Care Financing Review. Although the statistics published by the CMS are also derived from claims data, our claims data may potentially yield different results due to errors in data processing during the copying and conversion process from the CMS mainframe files or due to a lag in the claims reconciliation process when the claims data that we (p.85) use were extracted. There are also methodological issues to replicate, such as the determination of the appropriate denominator for calculation of participation rates and average spending. Here we present our results replicating CMS published statistics on enrollment and total spending.

The Balanced Budget Act of 1997 (BBA97) had a significant impact on aggregate Medicare spending and therefore on the trends in average Medicare spending that we report later in this paper. The major provisions of BBA97 included direct reductions to the PPS inpatient hospital annual operating update, PPS inpatient capital payments, indirect medical education (IME) payments, disproportionate share (DSH) payments, direct medical education (DME) payments, and expansion of the transfer policy (which reduces payments for transfers of short-term acute patients in ten DRGs who were discharged to an SNF, PPS-exempt facilities, or a home health agency). Other provisions included the implementation of prospective payment systems for outpatient hospital, skilled nursing facilities, and home health agencies, and the creation of Medicare+Choice managed care plans and more equitable payments for such plans across geographic areas. As the following results demonstrate, however, these changes appear to have caused a large one-time reduction in aggregate Medicare spending, but growth rates and trends appear largely to have reverted to pre-BBA97 levels.

Enrollment

Figure 3.1 compares Medicare enrollment as reported by the CMS in the statistical supplement through 2003 to the same statistic calculated by us

(p.86) using Medicare enrollment files for calendar years 1989 to 1999. Enrollment is reported in units of one thousand. The CMS reports enrollment as a count of beneficiaries at a point in time, namely July 1 of the calendar year, rather than as a count of any beneficiary enrolled at any time during the calendar year. We follow the CMS methodology and report enrollment separately by entitlement (i.e., aged or disabled). In general, enrollment from claims and from the CMS is very close, differing by about 1–2 percent on average for aged beneficiaries, with the difference less than 1 percent for more recent years (since 1996). That amounts to a difference of about 300,000 per year out of 32 to 34 million aged beneficiaries. Our enrollment is slightly higher than the CMS's, which may reflect an imputation on our part of enrollment for beneficiaries that are Medicare eligible due to reaching age sixty-five but that do not appear in Medicare enrollment files until they use Medicare services at an age greater than sixty-five. For disabled beneficiaries, our enrollment counts are slightly lower than CMS's, differing by about 2–3 percent on average. The absolute magnitude of the difference is only about 100,000 beneficiaries out of three to five million.

The results show that Medicare enrollment for aged beneficiaries has grown by about 1 percent per year from 1989 to 1999, from slightly over 30 million to slightly over 34 million. The growth rate was nearly 2 percent in the early years of the decade and declined to less than 1 percent in more recent years. Medicare enrollment for disabled beneficiaries has grown much more rapidly, averaging more than 5 percent per year over the eleven-year time period. As a consequence, the share of beneficiaries entitled due to disability has grown from 9 percent to 13 percent over the decade. Current CMS estimates show that the share reached nearly 15 percent by 2003. Because disabled beneficiaries use more services on average than aged beneficiaries, the growing share of disabled beneficiaries has potential implications for growth in Medicare spending.

Program Payments

Figure 3.2 shows total Medicare program payments from claims data for calendar years 1989 to 1999 and from the CMS through 2003. Medicare reports program payments, which only include monies dispensed from the Medicare trust funds and do not include third-party or beneficiary payments. They also do not include payments made to health plans for Medicare managed care enrollees. We report spending separately for part A (HI) and part B (SMI). For part A, total spending from claims differs from total spending reported by CMS by about 2 percent each year, with spending from claims slightly lower than spending reported by the CMS. In part, this might be due to a lag in claims for larger expenses that take more time to adjudicate. However, the trend in part A spending from claims tracks very closely with the trend in part A spending reported by (p.87)

CMS. Part A spending grew by more than 6 percent per year, with growth exceeding 9 percent per year until the BBA97, when total part A spending actually dropped by 3 percent per year on average. Part A spending grew from $56 billion in 1989 to a peak of $111 billion in 1996, before falling to $98 billion by 1999. More recent statistics reported by the CMS suggest that part A spending has returned to pre-BBA97 levels and growth rates. Part A spending reached $128 billion in 2003, growing at an annual rate of 6 percent since 1999.

Total part B spending from claims differs from part B spending reported by the CMS by 2.5 percent per year on average, although in more recent years the difference is only around 1 percent. The difference is greatest from 1991 to 1994, when CMS was converting to a different claims process for durable medical equipment. The trend in part B spending from claims tracks very closely to the trend in part B spending reported by CMS. Part B spending grew by more than 6 percent per year, and the rate of growth was not affected by the BBA97 as dramatically as part A spending. Part B spending grew from $36 billion in 1989 to $66 billion in 1999. In recent years, part B spending has continued to grow at 6–7 percent per year, reaching $88 billion in 2003. Before the BBA97, part B spending as a share of total Medicare spending had fallen from 40 percent to 35 percent, primarily because of rapid growth in nonacute part A spending in home health agency and skilled nursing facility services. After the BBA97, part B spending reverted to its earlier 40 percent share of total spending, and this share remained constant through 2003.

(p.88) Participation Rates

Figure 3.3 shows participation rates under part A and part B for calendar years 1989 to 1999. The participation rate is defined as the number of Medicare beneficiaries receiving services paid for with program funds during the calendar year divided by the number of Medicare beneficiaries enrolled during the calendar year. For both part A and part B, participation rates calculated from claims and reported by CMS in published statistics are very close. Validation of participation rates (in addition to aggregate spending and enrollment) is important for accurately reflecting trends in average beneficiary spending. Part A participation rates rose slightly from 1989 to 1994, from 201 to 219 per 1,000 beneficiaries. Starting in 1995, CMS began excluding managed care enrollees from the denominator of participation rates, so the participation rates increased slightly to 230 per 1,000 beneficiaries, where it has remained relatively constant in recent years. Part B participation rates also rose slightly from 1989 to 1994, from 755 to 815 per 1,000 beneficiaries. After managed care enrollees were excluded from the denominator, the part B participation rate rose to 833 per 1,000 in 1996 and has continued to increase slightly, reaching 842 per 1,000 beneficiaries in 1999, an increase of around 1 percent per year from 1995 to 1999. Figure 3.3 also shows managed care enrollment, which increased from 5 percent in 1989 to 18 percent in 1999, an increase of 12 percent per year, with most of the increase occurring after 1994. The BBA97, which included provisions to encourage managed care enrollment, seems to have

(p.89) reduced or reversed the rate of growth. The most recent estimates from CMS place managed care enrollment at 5.3 million of 40.8 million beneficiaries (13 percent).

We characterize the growth in annual Medicare expenditures by identifying separate trends among low- and high-cost users of Medicare services by beneficiary age.

Annual Expenditure Percentiles

Table 3.4 shows the level of annual expenditures by beneficiary percentile by year from 1989 to 1999. Percentiles shown are the 10th, 25th, 50th, 80th, 90th, 95th, and 98th. Expenditures are highly concentrated. In

(p.93) 1989, annual expenditures for the 90th percentile were 125 times greater than the annual expenditures for the 10th percentile. The levels ranged from $117 to $14,779. Annual expenditures for the 98th percentile were 2.7 times greater than the annual expenditures for the 90th percentile, with the level of the 98th percentile reaching $40,578. The median annual expenditure was $960. In any given year there are a few Medicare beneficiaries with extremely high expenditures, while most beneficiaries spend less than $1,000 annually.

In general, the growth rates have been greater for the lower percentiles than for the higher percentiles over the decade. For example, average annual expenditures for the 10th percentile increased from $117 to $196, an average annual growth rate of 5.1 percent. The average annual expenditure for the 50th percentile increased from $960 to $1,501, an average annual growth rate of 4.5 percent. The average annual expenditure for the 98th percentile increased from $40,578 to $49,928, an average annual growth rate of 2.1 percent. Post-BBA97 growth rates for the 80th, 90th, 95th, and 98th percentiles were actually negative, at −2.3 percent, −3.3 percent, −3.2 percent, and −3.0 percent, respectively. The BBA97 had the greatest impact on the high end of the annual expenditure distribution.

Share of Total Expenditures by Percentile

Table 3.5 shows the share of total Medicare expenditures accounted for by beneficiaries in the 0–50, 50–80, 80–90, 90–95, 95–98, and 98+ average annual expenditure percentile categories. In 1989, beneficiaries below the median average annual expenditure accounted for only 3.5 percent of total Medicare expenditures; beneficiaries in the 50–80 percentile category accounted for 16.7 percent; beneficiaries in the 80–90 percentile category

(p.94) accounted for 19.3 percent; beneficiaries in the 90–95 percentile category accounted for 18.3 percent; beneficiaries in the 95–98 percentile category accounted for 18.0 percent; and beneficiaries in the 98 + percentile category accounted for 24.2 percent. Because average annual expenditures grew more slowly for the higher percentiles, the share of total expenditures accounted for by the higher percentiles fell slightly. In 1999, the share of total Medicare expenditures accounted for by the 98th percentile fell from 24.2 percent to 23.6 percent. Overall, however, the share of total Medicare expenditures accounted for the various percentile categories has been quite stable over time.

These descriptive tables provide the foundation for more formal models describing the sources of Medicare expenditure growth, therefore illuminating growth forecasts and more effective simulations of policy reforms. We have already developed and applied such models (Garber, MaCurdy, and McClellan 1997). This section summarizes our cohort analysis methods, which model expenditure growth for groups of Medicare beneficiaries as a function of various characteristics, including birth year, time, and rankings in the expenditure distribution. This approach is particularly useful for assessing whether and how expenditure growth has differed between high-cost and low-cost enrollees, as our descriptive analyses suggested.

Specifications for Characterizing Trends

Considering the types of statistical formulations found in the empirical literature for describing the growth of variables—such as expenditures—in a population where differential rates operate across ages and time, two basic frameworks are well suited for our purposes: (a) one that describes the evolution of the cross-sectional relationships between age and a variable y over a sequence of years, and (b) one that models the movements of the life-cycle profiles of y associated with successive cohorts. Without arbitrary identifying assumptions, these two frameworks are statistically indistinguishable. Whereas existing studies of Medicare expenditures invariably apply some variant of the first framework as a basis for projecting growth, our research exploits frameworks falling into the second category for capturing the underlying features of Medicare growth.

To describe the trends of an aggregate quantity y over time—where y may represent a measure of Medicare expenditures, participation rates, and so on—one can model movements in the quantities y(c, α) measuring the values of y associated with cohort c at age α in the year c + α − 65. The most popular approach for describing the evolution of y is to specify the relationship (p.95)

The deterministic function f measures the systematic variation in y, and the errors u capture the contribution of period effects reflecting either cyclical or transitory phenomena. For fixed t, estimating f via equation (1) with observations on different values of y and α yields an estimate of the mapping between age and y at a point in time of the sort obtained in conventional cross-section analyses. Knowledge of how f behaves as a function of t determines how cross-section profiles shift from one year to the next.

The second approach for describing variation in y focuses on modeling the life-cycle profiles of cohorts by specifying the relationship

The deterministic function g measures trends, and the errors u reflect the deviations from these trends. For fixed c, estimating g via equation (2) with observations on different values of y and α yields an estimate of the life-cycle profile followed by y for cohort c. Knowledge of how g shifts as a function of c determines how the age-y relationships differ across cohorts.

The two approaches for characterizing profiles of y are linked by the equalities

Thus, there is no statistical advantage to using either f or g to model the growth of y because both functions convey the same information. This observation reflects nothing more than the inherent identification problem, which is well known in the literature (e.g., Heckman and Robb, 1985), that prevents one from being able to distinguish among age, period, and cohort effects. Using equation (3), it is straightforward to translate cross-section estimates into cohort estimates and vice versa. Our research strategy focuses on estimating g, and we will construct f from these results.

Cohort Profiles and Expenditure Growth

In the empirical analysis, we develop parameterizations of the cohort profiles g(c, a) to model the growth of two categories of quantities: (a) the fraction of the population using services under part A or part A or B, R(c, α) (i.e., the participation rate); and (b) statistics describing the distributional characteristics of the expenditure variables m_it (the value of real Medicare expenditures in year t incurred for individual i who is a member of cohort c and is age α in year t). These quantities jointly determine how distributions of Medicare outlays evolve across ages and over time. Use of the cohort profile g(c, α) to examine the trends followed by the various determinants of Medicare expenses offers a simple framework for understanding the process underlying the growth of health expenses.

(p.96) To illustrate the concept of a cohort profile, consider the use of such a formulation to characterize the evolution of an expenditure statistic, μ_ca, measuring, for example, average annual expenditures per individual in cohort c at age α. Members of the cohort who are sixty-five years old in year 0 experience a life-cycle profile of Medicare expenditures designated by g(c₂, α). Members of an older cohort who turn sixty-five in year c₁ (〈 0) have lifetime expenditures tracking the profile g (c₃, α). The growth of health expenditures experienced by cohort c in year t corresponds to the derivative

evaluated at the point α = t − c. Letting α* denote any particular age, the function g(t − α*, α*) specifies the level of expenditure in year t. The growth of this level is

Plotting g(t − α*, α*) against t shows how the level of Medicare outlays were expended at age α* or how participation at this age shifted over time. In the figure this plot is designated as “entry expenditure” when α* is set equal to age sixty-five. In the figure the cross-section profile of expenditures in year t* is given by the values of g intersecting the vertical line drawn at t*.

Describing the Evolution of the Distribution of Expenditures

Using this framework, we describe many attributes of the distributions of Medicare expenses by choosing the dependent variables y(c, α, x) as various statistics computed using the individual observations on y making up cell (c, α, x). Choosing y(c, α, x) as the fraction of persons with Medicare service use (R[c, α]) implies that the function g describes the life-cycle profile of Medicare participation rates of a cohort at different ages. Finally, choosing y(c, α, x) as percentiles of the Medicare expenditure distribution describes the profile of different points in the distribution of utilization by a cohort.

To estimate how the distribution of Medicare outlays evolve, we model the behavior of several percentiles by interpreting the quantity y(c, α) in equation (2) as a particular percentile of the distributions of the variables m_it corresponding to a specified cohort for a given age or year. In particular, we interpret y(c, α) = P##_ca(m_kt), which represents the ##th percentile of the distribution of Medicare expenditures in year t per patients who are members of cohort c at age α. With y(c, α) calculated as a specified percentile, g depicts its life-cycle profile. By combining information on several (p.97) such formulations of g, we can infer how the distributions of Medicare expenditures vary across cohorts, within cohorts, across ages, and over time during our five-year horizon. We consider three percentile formulations for y(c, α): the 10th, 50th, and 90th. Knowledge of these three percentiles provides a useful tool with which to examine changes in the shape of the distribution of Medicare utilization since it allows us to describe the entire conditional distribution, including the life-cycle and time-series patterns of the spreads of the distributions.

Parameterization and Estimation of Cohort Specifications

To outline our approach for estimating parameterizations of g, denote y_it as the observation on the ith individual in year t. Let the quantities y(c, α, x) correspond to the statistics computed using the values of y_it associated with the appropriate cohort c, age α, and set of demographic characteristics x. After forming the variables y (c, α, x), an observation on equation (2) may be written as

We assume that the disturbances follow the error components model where the subscripts (c, α) map into the subs (t, α) via the relation c = t − α, the errors ū_t are common time effects, and the errors ū_tax are the idiosyncratic deviations from trends for cohort c at age α after the removal of common year components. The time effects ū_t are estimated as parameters subject to the restrictions that they are orthogonal to g(c, α, x) for any c. Thus, the ū_t s represent deviations from trends; they can be interpreted as (macroeconomic) cyclical variations in Medicare utilization.

We estimate a formulation of g that is a variant of the parameterization

The quantities Ф_j (α) determine the shape of a cohort's lifetime profile with respect to age, and the functions β_j (c) capture cross-cohort variation in life-cycle profiles. One can readily consider transformations of expenditures other than logarithms as a dependent variable when using equation (6) as a specification of g; and one can incorporate individual characteristics in the functions β_j and Ф_j to allow profiles to vary across demographic groups.

The particular parameterization of equation (6) estimated in this analysis is

(p.98) According to this parameterization, each cohort's expenditure function is a quadratic in age, and the coefficients of these cohort profiles vary according to a quadratic polynomial in cohort year.

The analysis estimates the coefficients β_kj and the time effects ū_t by applying ordinary least squares (OLS) to equation (6) with g specified by equation (7). The sample includes all available annual observations, with regressions run separately for demographic groups under consideration. As mentioned, the time effects are constrained to be orthogonal to g(c, α) for all c, thus requiring ū_k to satisfy the restrictions

Our empirical work indicates that specification (7) is sufficiently rich to capture the shifts in Medicare expenditures that occurred during the 1990s. Increasing the degree of either the polynomial in age or the polynomial in cohort year fails either to improve the goodness of fit of the statistical model at conventional significance levels or to change the main empirical findings.

We estimate the previous specifications for annual Medicare participation rates and expenditure percentiles—the 10th, 50th, 90th, 95th, and 98th—using all years covering the 1989 to 1999 period.

Growth in Participation

Figure 3.4 shows the rate of growth in participation rates for part A and part A or B. The cross-sectional curves (marked with Xs) show the 1995 participation rates by age. For both part A and part A or B, participation rates increase with age until around age ninety, when participation rates begin to decrease. Part A participation rates increase with age more rapidly than participation rates for part A and B combined until more than 40 percent of beneficiaries in their late eighties or early nineties receive some part A service (i.e., inpatient hospitalization, skilled nursing facility, home health agency, or hospice). The figure also shows three cohort profiles for beneficiaries that turned sixty-five in 1991 (marked with squares), 1981 (marked with diamonds), or 1971 (marked with triangles), respectively. Each cohort profile intersects the cross-sectional curve in 1995 (that is, at age sixty-nine, seventy-nine, and eighty-nine, respectively). The fact that the cohort profiles are steeper than the cross-sectional curve indicates that Medicare participation rates for each cohort increase more rapidly than can be attributed to pure aging. For the 1991 cohort, the percentage change in the part A participation rate was 44 percent (of the 1995 level) and the percentage change in the part A or B participation rate was 27 percent. For the 1981 cohort, the percentage change in the part A participation rate was (p.99)

(p.100) 44 percent, and the percentage change in the part A or B participation rate was 11 percent. For the 1971 cohort, the percentage change in the part A participation rate was 26 percent, and the percentage change in the part A or B participation rate was only 2 percent. Service use was increasing more rapidly for the younger cohorts, at least until the later years of the decade.

The rapid increase in the part A participation rate was due primarily to an increase in service use for skilled nursing facilities, home health agencies, and hospices, although inpatient hospital service use increased as well. Our estimates indicate that the participation rate for skilled nursing facility services increased from 1.9 percent to 4.3 percent from 1989 to 1997; the participation rate for home health agency services increased from 4.9 percent to 9.8 percent; and the participation rate for inpatient hospital services increased from 18.7 percent to 19.5 percent. After BBA97, part A and part A or B participation rates actually decreased, as shown in table 3.2.The participation rate for home health agency services decreased from 9.8 percent to 7.1 percent from 1997 to 1998. Participation rates for inpatient hospital and skilled nursing facility services remained constant. In figure 3.4, the part A participation rate cohort profiles increase until 1997, and then either flatten or decrease slightly, especially in the youngest (1991) cohort. The fact that the cohort profiles are less steep than the cross-sectional curve after 1997 indicates that part A participation rates for the 1991 and 1981 cohorts decreased more rapidly during this time period than can be attributed to pure age effects. Part A participation rates for the oldest cohort (1971) continued to increase despite the policy changes, but at a much lower rate.

Growth Across Different Intensities of Use

Figure 3.5 shows the rate of growth across the distribution of annual Medicare expenditures. The cross-sectional curves show a snapshot of average annual Medicare expenditures in 1995 for the 10th, 50th, and 90th percentiles by age. Intensity of service use increases with age for all three percentile categories, at least until the late eighties or early nineties, when intensity of service use declines. Intensity of service use by age actually increases the most for the median percentile, as can be seen more clearly in figure 3.6, which shows only the 10th and median percentile categories. From age sixty-five to age ninety, the percent change in average annual expenditures is 162 percent for the 10th percentile, 266 percent for the median, and 133 percent for the 90th percentile.

Three cohort profiles are shown for the same percentiles for beneficiaries that turned sixty-five in 1991, 1981, or 1971. The cohort profiles show the actual life cycle growth in Medicare expenditures for the specified percentiles of each cohort. The cohort profiles intersect the cross-section profile at the ages of the cohorts in 1995 (that is, at age sixty-nine, seventy-nine, and eighty-nine). The fact that the cohort profiles are steeper than the (p.101)

(p.102) (p.103) cross-sectional curve indicates that average annual Medicare expenditures grew much more rapidly for each cohort than can be attributed to pure aging, especially for the 90th percentile. However, the cohort profiles also show that annual Medicare expenditures after 1997 declined rapidly in the 90th percentile, remained constant in the median percentile (except for the oldest cohort, where there was a decline; see figure 3.6), and increased for the 10th percentile. The fact that the cohort profiles for the 10th percentile are steeper than those for the 50th or 90th percentiles indicates that Medicare expenditures grew more rapidly at the lower end of the expenditure distribution. The growth at the higher end was mitigated by the impact of BBA97, as indicated by the downward slope of the curves after 1997.

For the youngest (1991) cohort, the percentage change in total expenditures for the period from 1991 to 1999 was 48 percent (of the 1995 level) for the 90th percentile, 65 percent for the 50th percentile, and 80 percent for the 10th percentile. The pre-BBA97 percentage changes were 58 percent, 71 percent, and 57 percent, respectively. For the 1981 cohort, which was seventy-five years old in 1991, the rates of growth were 36 percent, 62 percent, and 57 percent, respectively. The pre-BBA97 percentage changes were 43 percent, 66 percent, and 44 percent, respectively. For the 1971 cohort, which was eighty-five years old in 1991, the rates of growth were 22 percent, 35 percent, and 32 percent, respectively. The pre-BBA97 percentage changes were 30 percent, 41 percent, and 27 percent, respectively. Collectively, our results indicate that expenditures grew more rapidly in the 10th and median percentile expenditure categories than for the 90th percentile category. Expenditures also grew more rapidly in the youngest cohort and less rapidly in the oldest cohort, with the middle cohort in between. Finally, the BBA97 reduced the rate of growth across all cohorts for the median and 90th percentile categories, but especially for the oldest cohort. The BBA97 actually increased the rate of growth for the 10th percentile, especially for the younger cohorts.

Figure 3.7 shows the rate of growth of average annual Medicare expenditures in 1995 for “high-cost” beneficiaries in the 90th, 95th, and 98th percentiles by age. Intensity of service use increases with age for all three percentile categories, at least until the late eighties or early nineties, when the intensity of service use declines. For the 98th percentile, the intensity of service use begins to decline around age eighty. Intensity of service use by age increases the most for the 90th percentile. From age sixty-five to age ninety, the percent change in average annual expenditures is 133 percent for the 90th percentile, 67 percent for the 95th percentile, and 29 percent for the 98th percentile.

Three cohort profiles are shown for the same percentiles for beneficiaries that turned sixty-five in 1991, 1981, or 1971. Similar to the results in figure 3.5, the cohort profiles are much steeper than the cross-sectional curve, indicating that average annual Medicare expenditures grew much more

Three cohort profiles are shown for the same percentiles for beneficiaries that turned sixty-five in 1991, 1981, or 1971. Similar to the results in figure 3.5, the cohort profiles are much steeper than the cross-sectional curve, indicating that average annual Medicare expenditures grew much more (p.104)

(p.105) rapidly for each cohort that can be attributed to pure aging, especially in the 98th percentile. The cohort profiles show that average annual Medicare expenditures for high-cost beneficiaries declined rapidly after 1997 in the 90th, 95th, and 98th percentiles. The fact that the cohort profiles for the 90th percentile are steeper than for the 95th or 98th percentiles indicates that Medicare expenditures grew less rapidly for the highest-cost users of Medicare services.

For example, for the youngest (1991) cohort, the percentage change in total expenditures for the period from 1991 to 1999 was 28 percent (of the 1995 level) for the 98th percentile, 40 percent for the 95th percentile, and 48 percent for the 90th percentile. The pre-BBA97 percentage changes were 36 percent, 48 percent, and 58 percent for the 98th, 95th, and 90th percentiles, respectively. For the 1981 cohort, seventy-five years of age in 1991, the rates of growth were 21 percent, 28 percent, and 36 percent, respectively. The pre-BBA97 percentage changes were 28 percent, 35 percent, and 43 percent, respectively. For the 1971 cohort, those eighty-five years of age in 1991, the rates of growth were 14 percent, 18 percent, and 22 percent, respectively. The pre-BBA97 percentage changes were 24 percent, 27 percent, and 30 percent, respectively. Collectively, our results indicate that expenditures grew more rapidly in the 90th percentile category, less rapidly in the 95th percentile category, and the least rapidly in the 98th percentile category. Expenditures also grew the most rapidly in the youngest cohort across each percentile category and less rapidly in the oldest cohort, with the middle cohort in between. Finally, the BBA97 reduced the rate of growth across all cohorts and all percentile categories, but especially for the oldest cohort and the 98th percentile category.

Trends Have Changed

Figures 3.8, 3.9, and 3.10 show the growth rate of expenditures for an earlier time period than the previous figures. Figures 3.8 and 3.9 show the cross-sectional curves and the cohort profiles using the 1991 cross section as reference for the 10th, median, and 90th percentiles by age. The cohort profiles intersect the cross-section profile at the ages of the 1967, 1977, and 1987 cohorts in 1991 (that is, at age sixty-nine, seventy-nine, and eighty-nine, respectively). Comparison of figures 3.8 and 3.5 demonstrates that the trends forecasted by previous studies were not realized. Using the earlier cohort suggests growth rates of 49 percent, 62 percent, and 58 percent for the 90th, median, and 10th percentiles, respectively. For the 1977 cohort, the growth rates were 41 percent, 67 percent, and 50 percent. For the 1967 cohort, the growth rates were 36 percent, 43 percent, and 30 percent. The policy changes of the late 1990s mitigated these growth rates, especially for the older cohorts and the higher percentiles.

Similarly, figure 3.10 shows the cross-sectional curves and the cohort profiles using the 1991 cross section as reference for the 90th, 95th, and (p.106)

(p.107) (p.108) (p.109) 98th percentiles by age. Using the earlier cohort suggested growth rates of 29 percent, 41 percent, and 49 percent for the 98th, 95th, and 90th percentiles, respectively. For the 1977 cohort, the growth rates were 24 percent, 33 percent, and 41 percent. For the 1967 cohort, the growth rates were 21 percent, 27 percent, and 30 percent. Again, the policy changes of the late 1990s mitigated these growth rates, especially for the older cohorts and the higher percentiles. However, the results from section 3.2 suggest that this reduction in the growth rate represents only a temporary change as the result of policy changes implemented concurrently with BBA97. Administrative data from recent years through 2003 indicate that Medicare expenditures have reverted back to the trends suggested by the 1991 cross section. Therefore, the growth profiles suggested by figures 3.5 and 3.7 are unlikely to be realized as well. The trends in administrative data presented in section 3.2 indicate that the growth in Medicare expenditures has reverted to previous rates and may have even accelerated.

Tables 3.8 through 3.11 present an array of statistics summarizing the incidence of high-cost months for various populations of Medicare beneficiaries. Each table describes the distributions of experiences for months classified as high cost by the table, with definitions of “high cost” different across tables. In addition to months of experience, these tables report shares indicating the size of the beneficiary population and the fraction of total Medicare expenditures accounted for by each group identified in rows of the tables. (p.112)

(p.113) (p.114) (p.115)

(p.116) Characteristics of High-Cost Users and Structure of Tables

The first column of tables 3.8 through 3.11 lists the groups of Medicare beneficiaries whose experiences are summarized in the remaining columns. The rows designated in the first column specify the classifications of beneficiaries considered in the analysis. The table groups rows into five categories:

• • Patients w/experience: This row describes population shares and attributes of the distributions of high-cost months for all beneficiaries who experience at least one month classified as high cost by the table any time during the 1989–99 period. All of the remaining rows report results for segments of this population.

• • Gender: The top set of rows reports experiences for males and females separately.

• • Age: The next set of rows presents distributions of high-cost months by age groups. Groups consist of beneficiaries who were the age indicated at the start of the bracket, with observations included for years covered by the bracket, or until the person died if this event occurred before reaching the age at the end of the age bracket. Beneficiaries can, of course, be members of more than one age group if they experience multiple high-cost episodes that occur at different ages.

• • Cost percentile, any year: This set of rows summarizes the incidence of high-cost months for groups of beneficiaries classified by whether their calendar-year expenditures reach particular thresholds. The “GE 80th percentile” refers to all beneficiaries who had expenditures above the 80th calendar-year percentile in any year during the 1989–99 period, the “GE 90th percentile” group includes all beneficiaries with expenditures above the 90th annual percentile in any year of the period, and so on through the 98 percent level of annual expenditures.

• • Cost percentile, any 2 year: The final set of rows categorizes beneficiaries according to whether their calendar-year expenditures attain the “GE 80th percentile,” “GE 90th percentile,” “GE 95th percentile,” and “GE 98th percentile” categories in two or more years.

The second through the last columns of tables 3.8 through 3.11 present five categories of statistics for each of the Medicare beneficiary groups:

• • Share of Medicare patients: These columns report the share of Medicare patients in the group (defined by the row) for two different Medicare populations. The first column (“All”) shows the share of all patients, where “patients” refers to all beneficiaries who received any Medicare services during the 1989–99 period. The second column (“High-cost month group”) gives the percentage of the group of patients with at least one month classified as high cost by the table.

  (p.117) For example, table 3.8 shows that female patients who experienced at least one of month falling into the 80th expenditure percentile category comprise 50.6 percent of all the beneficiaries who received at least one Medicare service between 1989 and 1999. Further, females account for 58.3 percent of all the patients who make up the 80th percentile group.

  Referring to the rows designating age brackets, table 3.8 shows that 26.5 percent of elderly receiving at least one Medicare service between 1989 and 1999 were ages sixty-five to sixty-nine when they experienced a high-cost month falling into the 80th expenditure percentile category, and 30.5 percent of those in the 80th expenditure percentile category were sixty-five to sixty-nine years old when they experienced at least one high-cost month. The corresponding figures for the seventy to seventy-four age group are 28.3 percent and 32.7 percent, implying that more of the 1989–99 Medicare beneficiary population experienced a high-cost month during ages seventy to seventy-four. Because members of the 1989–99 Medicare beneficiary population are typically represented in several age groups and may experience multiple high-cost episodes at different ages, percentages in these rows do not add up to 100 percent or to the overall population.

• • Total no. of high-cost months: This group of columns summarizes properties of the distribution of the total number of high-cost months experienced during the 1989–99 period. The column labeled “Mean” presents the average number of months experienced by beneficiaries in the specified row, and the next three columns show the percentiles of the distribution of the number of high-cost months.

  Referring again to female patients, this group of columns in table 3.8 shows that women with at least one high-cost month (i.e., in top 20 percent of months) had an average of 9.4 high-cost months during the 1989–99 decade. Twenty percent had two or fewer high-cost months during this period, half had six or fewer such months, and 80 percent experienced up to fourteen high-cost months.

• • No. of high-cost spells: The next set of columns reports statistics describing the number of high-cost month spells experienced by the prescribed population during the 1989–99 period, with a spell defined as a continuous series of high-cost months. These columns show the average number of spells, along with percentiles associated with the distribution.

  Continuing the above example for table 3.8, women with at least one high-cost month (i.e., in top 20 percent of months) had an average of five high-cost spells during the 1989–99 decade. The median such woman had four high-cost spells. Twenty percent had two or fewer high-cost spells, and 80 percent had eight or fewer of these spells.

• • Length of high-cost spells: This set of columns presents statistics characterizing the lengths of all high-cost spells experienced by beneficiaries (p.118) in the specified group of beneficiaries, giving the average and percentiles describing the lengths (in months) of completed spells.

  According to table 3.8, the average length of high-cost spells for women with at least one high-cost month (i.e., in top 20 percent of months) was two months. The 20th, 50th, and 80th percentiles were 1, 1.4, and 2 months, respectively.

• • Share of Medicare expenditures: The last set of columns reports two shares of total Medicare expenditures spent during the 1989–99 period attributable to the group of beneficiaries in a given row, with all quantities measured in nondiscounted 2000$. The left column in this set (“All months”) shows the share of total expenditures accounted for by the sum of expenditures for all months for the population of beneficiaries specified in the row. The right column (“High-cost months”) lists the share of total expenditures due to the sum of expenditures in only over those months classified as high cost by the table.

  Inspection of table 3.8 reveals that female patients who experienced at least one of month classified in the 80th expenditure percentile category account for 57.7 percent of total Medicare expenditures during the 1989–99 period. Counting expenditures for this group that occur only in their high-cost months accounts for 51.4 percent of total Medicare expenditures.

Portrait of High-Cost Month Experiences

Tables 3.8 through 3.11 present results considering only the elderly segment of the Medicare population—beneficiaries 65 years and older. Results differ only marginally considering Medicare beneficiaries of all ages (which also includes the younger disabled population). Table 3.8 reports the distributions of experiences for months classified as having expenditures above the 80th percentile. Table 3.9 gives findings for months with expenditures above the 90th percentile. Table 3.10 lists distributions for the 95th percentile of monthly expenditures. Finally, table 3.11 presents findings for months in the 98th percentile of expenditures.

Inspecting the first row of table 3.8, we see in the first set of columns that 86.8 percent of elderly Medicare beneficiaries experience at least one month during the 1989–99 decade in which they receive medical services costing above the 80th percentile for months during the year of expenditures. So practically everyone can be considered a high-cost user at some time in their lifetime over a short enough period.

Moving to the farthest left set of columns, we see that this group of beneficiaries accounts for 99.6 percent of all Medicare expenditures totaled over the 1989–99 decade. Counting only expenses incurred during their high-cost months covers 89.7 percent of total expenditures. Thus, knowledge of the expenses incurred during only high-cost months for this group essentially accounts for all but a minor portion of Medicare spending.

According to the other columns in table 3.8 for the “Patients w/ experience” (p.119) row, the average number of high-cost months experienced was 9.1, with 20 percent of the group having two or fewer months, 50 percent having six or fewer months, and 20 percent having fourteen or more months classified as high cost. This group experienced an average of 4.9 spells, with 20 percent having eight or more. Spells are short, lasting two months on average, with less than 20 percent of beneficiaries experiencing spells lasting longer. On average, women had slightly more high-cost months than men (9.4 versus 8.8).

Examining table 3.9 for the group of elderly beneficiaries who received health services during a month costing above the 90th percentile level for months during the year, we see that 75.3 percent of all beneficiaries make up this group. Total spending on this group accounts for 98.1 percent of all Medicare expenditures for the 1989–99 decade, and expenses incurred during high-cost months for this group cover 79.1 percent of total expenditures. Thus, considering only this classification of high-cost months covers all but about 20 percent of the decade's expenditure on Medicare. The average number of months defined as high cost by this definition equals 5.3, occurring in an average of 3.3 spells per recipient. Only 20 percent of the group had eight or more months classified as high cost, and 20 percent had five or more spells. Once again, spells tend to be short. Further, women had slightly more high-cost months than men (5.3 versus 5.2 months).

Turning to tables 3.10 and 3.11, which classify high-cost months as achieving expenses reaching the 95th and 98th percentiles for monthly expenditures, reference to the top row indicates that the fraction of beneficiaries experiencing these high levels of spending still involves a substantial segment of the elderly population: 60.7 percent have at least one 95th percentile month, and 39.7 percent have one or more 98th percentile months. Spending on the 98th percentile group consumes 78.6 percent of all Medicare expenditures accumulated over the 1989–99 decade, and the sum of those expenses incurred during only high-cost months covers 40.8 percent of total expenditures. The spending on the 95th percentile group accounts for 62.9 percent of total expenditures for the decade according to the last column of table 3.10. Thus, knowledge of the incidence of 95th percentile months alone explains nearly two-thirds of Medicare spending over the decade.

Inspecting the rows in tables 3.8 through 3.11 describing experiences by age groups reveals that average high-cost months are nondecreasing in age, consistent with the view that costs rise with age given the onset of chronic illness. A notable increase in averages can be seen for the 80th percentile months; however, the difference decreases at higher percentiles. In particular, for 98th percentile months, averages and distributions remain constant across age groups. The shares of total Medicare expenditures accounted for by high-cost-month beneficiaries and by high-cost months do not exhibit a monotonic relationship with age. This reflects the fact that although average Medicare experiences generally rise with age, the overall (p.120) size of age groups declines the older the group. Consequently, the contribution of age groups to the high-cost month population depends on how mortality rates balance against the occurrence of illnesses leading to the incidence of intense use of medical services.

Across all demographic groups considered in tables 3.8 through 3.11, spells of high-cost months are short. Higher months of experience come about due to the occurrence of more spells rather than the length of spells.