Current Models of Health Care Cost Projections
This chapter summarizes the first workshop session, which was designed to provide background on the relative merits of current models for projecting health care costs for the Medicare population. The session opened with an overview of a paper prepared for the workshop that describes the major modeling approaches currently used for policy analysis and research, their capabilities and weaknesses, and their uses (see Appendix A).
It was followed by presentations on the policy models and underlying data in use by three federal agencies:
Medicare cost modeling for health care spending at the Congressional Budget Office (CBO);
long-range health care expenditure projections by the Office of the Actuary (OACT) at the Centers for Medicare & Medicaid Services (CMS); and
the role of the Medical Expenditure Panel Survey (MEPS) of the Agency for Healthcare Research and Quality (AHRQ) as a resource for the government’s economic models and projections of health care expenditures.
The CBO and CMS models are important to understand and assess, given the prominent role they play in policy analysis and formulation. By law, CBO must prepare 10-year estimates of the costs of health care reform proposals introduced in Congress, including changes to the Medicare and other government-supported health care programs (CBO also prepares
longer term projections). In turn, Congress must use the CBO estimates in assessing the estimated cost of a specific legislative proposal and its effects on the deficit. CMS OACT estimates are used by the Medicare Trustees to describe the projected financial condition of the program over the short, medium, and long terms out to 75 years. In turn, the Trustees’ reports affect the policy debate on possible changes to the program that could affect benefits and costs. MEPS is a key source of data on many aspects of health care cost modeling, including estimates of people lacking health insurance coverage, provisions of employer-provided health care plans, and estimates of health care coverage and expenditures for the most populous states.
PREDICTING MEDICARE COST GROWTH
John Friedman (Harvard University) began by noting that over the past 45 years Medicare spending has grown faster than the gross domestic product (GDP). If allowed to continue without some change, by 2080 Medicare health care expenditures alone would reach 99 percent of GDP. Clearly that cannot be allowed to happen. The nation is faced with the questions of how and when cost growth will slow and what the consequences of this slowing will be. To solve this problem, academic and government researchers and policy analysts have been developing models for projecting Medicare cost growth.
Friedman provided a brief review of the three main approaches to projecting Medicare cost growth—extrapolation, microsimulation, and computable general equilibrium. He then briefly explained the assumptions, mechanics, strengths, and weaknesses of each and showed how policy makers use these methods.
Extrapolation is the most direct approach to forecasting future growth. It uses historical patterns in aggregate spending as a guide for projecting future growth, relying entirely on a statistical or actuarial approach rather than an economic approach, and is essentially based on a regression. Its transparency is its strength. The end result is clearly the sum of its parts, and that is a real value in keeping things transparent. Extrapolation is best suited for short- to medium-run projections. Its long-run numbers may be accurate but without detailing what drives such numbers, correct or not, extrapolation leaves the researcher unsatisfied.
The fundamental problem with the extrapolation approach is that it does not address the fact that something has to change. In practice, however, researchers often do not simply extrapolate, but also impose some brakes on the system to limit growth. For example, a constraint used by the CBO as a brake on the system is that nonhealth care consumption cannot
decline. That seems like a reasonable restriction, although nothing in the data suggests that it is the right thing to do or what its costs are. Depending on the types of external assumptions made about how and when Medicare cost growth is going to slow down, one gets very different projections with different implications.
Microsimulation has been used by several modelers. It is a form of extrapolation, but in a much more detailed, nonparametric form. For example, consider dividing the entire population into small groups defined by demographic, economic, and health conditions. The groups are defined to be mutually exclusive. The researcher then estimates the transition probability of moving from one group to another during a given year. Depending on the data available, one can estimate this in a fairly flexible way. Microsimulation models are well suited to study the effects of alternative policy scenarios or posited changes in health conditions or health care technology that affect particular aspects of the health care system. They account for heterogeneity in demographic transitions.
The RAND Future Elderly Model is a well-known health-related microsimulation model (Goldman et al., 2004). It uses the Medicare Current Beneficiary Survey of CMS and the Health and Retirement Study to estimate both demographic and health conditions.
The advantage of the microsimulation approach is its flexibility in modeling distributional impacts in the short or medium run. It is easy to consider various posited changes in policy or health conditions because the entire heterogeneity of the health care system is represented. Another strength is transparency: it is easy to see how the links flow. This is especially valuable in short- to medium-run projections.
Its weakness is conceptually the same as with extrapolation. There is no answer to the fundamental problem that something has to change, and one does not know when or how change will occur.
Computable General Equilibrium
The first two approaches above are primarily statistical or actuarial approaches to projecting. There is no sense of what the incentives are, that people have a demand function for health care. The complete opposite of that approach is a computable general equilibrium (CGE) model, which takes very seriously the incentives that drive the demand for and production of health care. It models the economic relationships that drive health care spending. Conceptually, CGE modeling is very rigorous and strong. There is demand for health care consumption or nonhealth care consumption;
there is the health care sector, which employs people and produces health; and the prices in the medical care market equate supply and demand. This approach takes head on the question of what will change and how. For example, if health care prices increase, then demand for it will be lower; if health care consumption increases, then perhaps on the margin the demand for it will be lower. This approach provides a direct answer to the question of what will happen and why.
However, the weakness of CGE models is that they tend to be highly dependent on external assumptions in an opaque way. These models are also too complex to allow consideration of heterogeneity in the population that is of interest for policy purposes. Moreover, without the need to be constrained in some way, these models often cannot be solved or have multiple equilibriums. Conceptually, however, the CGE approach is on the right track; the methodology needs further development so that such models can be used more independently.
Uses of Modeling Approaches
Briefly, different federal agencies combine these three modeling approaches in different ways for projections to guide policy. For example, CMS uses a combination of extrapolation with a CGE model, extrapolating over the first 10 years, interpolating years 11-24, and constraining growth between 25 and 75 years to an average of 1 percentage point in excess of the rate of per capita GDP growth (expressed as the GDP + 1 assumption). CMS assumes the long-run growth rate and then basically uses the CGE model to achieve asymptotic convergence over time. At some stage in the future, when CGE models are more developed, researchers may be able to use them to estimate what the long-term growth rate is, not just how it is going to be distributed over time.
CBO uses more of a constrained extrapolation approach in the long run, with a positive growth constraint on nonhealth care consumption.
AHRQ uses a microsimulation approach with data from the MEPS because the agency is more focused on heterogeneity of the population and because it does not really target the 75-year long-run projection.
The U.S. Department of Veterans Affairs uses microsimulation adapted to the veteran population to project in the short to medium run what veterans are going to need.
These different projection strategies provide very different estimates, especially over the long run. It is important to keep in mind that a tremendous amount of uncertainty accompanies long-term forecasting.
In closing, Friedman observed that technology, which is thought to drive much of the growth in health care costs, is totally absent from all of these models. There is a growing body of research on how technological
development responds to economic factors. Some examples are the development of vaccines and drugs as a response to market size (Acemozğlu and Linn, 2004; Finkelstein, 2004) and hospitals’ response to reimbursement incentives when choosing their labor-capital mix (Acemozğlu and Finkelstein, 2008). Researchers should also try to understand how some technologies lower the cost of existing health care options, whereas others create new, more expensive options. Just asking people what is going to happen with technology is an underrated strategy. RAND researchers (Shekelle et al., 2005) use a Delphi panel approach, as does Weizman (2001).
MEDICARE COST MODELING FOR HEALTH CARE SPENDING AT CBO
Joyce Manchester (Congressional Budget Office) described the framework for CBO’s long-term cost projections for Medicare and other federal health care programs, the outlook for the federal budget, and the assumptions regarding cost growth in Medicare and other health care spending.1 She also identified some of the strengths and limitations of the CBO approach.
Framework for Long-Term Medicare Cost Projections
CBO examines the pressures facing the federal budget over the coming decades in the context of current law.2 Most of Manchester’s presentation was based on current law, which, among other things, assumes that many of the tax reductions passed early in the decade will expire and that Medicare’s reimbursement rates for doctors will be constrained much more than has been true in the past.
Cost projections over the first 10 years are based on detailed program projections that underlie CBO’s baseline. The Medicare projections that go into those 10-year projections are very detailed, looking at specific kinds of Medicare spending.
Beyond 10 years, CBO relies on its long-term model, CBOLT, to analyze the budgetary and distributional effects of the Social Security program and other federal policies and programs, to evaluate potential reforms to federal entitlement programs, and to quantify the nation’s long-term fiscal challenges. CBOLT is primarily a microsimulation model, although an actuarial
framework and an overarching macro model provide targets for certain subgroups of the population as well as aggregate values for some variables. For example, CBO does not have the ability to do detailed 75-year spending projections at the individual level for Medicare and Medicaid at this time. While its long-term detailed projections for Social Security are developed in the microsimulation model, its projections for Medicare, Medicaid, and other health care spending are developed at a more aggregated level in the actuarial framework. All other federal spending is assumed to grow with GDP. The CBO Tax Analysis Division calculates effective rates of major types of taxes for the first 10 years of the projection period, and those rates are used to project tax revenues over time at the aggregate level.
The value of longer term projections is to highlight trends; they also provide a baseline for policy changes. Limitations of the longer term projections include uncertainty, especially surrounding the health care programs. CBO cannot precisely quantify that uncertainty with statistical modeling in the Medicare and Medicaid program projections, although an attempt is made to do so for the Social Security projections.
Interactions with macroeconomic conditions present challenges to the CBO approach as well. The most prominent is very high ratios of debt to GDP projected in the future. However, CBOLT does not currently account for the effects of rising debt to GDP ratios on the economy. For example, CBO assumes that real interest rates stay fixed at 3 percent. CBO is aware that this assumption may not be realistic, but the goal is to provide a baseline against which, given a stable backdrop, Congress can examine reform proposals.
CBO is in the midst of ongoing discussions, both internally and with a panel of outside experts, to improve the long-term projections and especially to communicate the macroeconomic consequences of those projections.
Outlook for the Federal Budget
In the absence of significant changes in policy, the rising costs of health care and the aging of the population will cause federal spending to grow much faster than the economy, putting the budget on an unsustainable path. Based on its June 2009 analysis, CBO projects that by 2035 the share of total government spending for health care will more than double to about 13 percent of GDP, up from about 6 percent of GDP in 2008. Medicare alone will account for about 7 percent of GDP in 2035, up from 3.5 percent in 2008. The emphasis here is on the next 25 years, because so much uncertainty exists beyond that.
Assumptions Regarding Health Care Cost Growth
In CBO’s current long-term budget projections, Medicare spending for the first 10 years, 2009-2019, follows the CBO March 2009 baseline. In 2020, CBO assumes that excess cost growth for Medicare is equal to the average historical rate of 2.3 percentage points. Excess cost growth is the amount by which per capita health spending (adjusted for age, sex, and time until death) is growing faster than per capita GDP growth; the historical average is based on the past 30 years.
CBO assumes that excess cost growth in Medicare, Medicaid, and other health care spending will begin to slow in 2021. With all the pressures that will be brought to bear on nonfederal spending, including the states’ share of Medicaid spending, CBO assumes that part of the slowdown in non-Medicare spending will spill over to Medicare, causing excess cost growth in that program’s spending to slow by one-third of the amount in the non-Medicare sectors.
Excess cost growth for other (non-Medicare and non-Medicaid) health care spending is projected to decline from 1.8 percentage points in 2020, which is the historical rate of growth, to 0.1 percentage point in 2083. That outcome is the result of an assumption that households will be unwilling to spend so much on health care that their real nonhealth care spending per capita will decline during the 75-year projection period. CBO has been using that assumption for about 3 years now and continues to evaluate its validity.
Strengths and Limitations of the CBO Approach
CBO connects its long-term projections to the detailed 10-year forecast from its Medicare analysis. As stated above, the long-term approach is based on historical excess cost growth in health care spending adjusted for age, sex, and time until death. It relies on a simple rule regarding patterns of household consumption—that is, households will not be willing to reduce real nonhealth care per capita spending at any time during the 75-year projection period. And it is designed to be consistent with CBO’s overall long-term budget projections, providing a baseline for policy changes.
CBO’s approach to modeling health care costs, including Medicare, is fraught with uncertainty. A tremendous amount of uncertainty surrounds health care spending growth over 25 years and even more so over a 75-year horizon. Currently the approach is implemented at the aggregate level only and not in the microsimulation model.
The current approach reflects no epidemiological or technological trends other than those reflected in history. Most of the growth in excess
cost has come from technological advances over time, but it is difficult to know how to model those changes going forward.
Finally, CBO’s projections do not take into account the consequences of health care spending being equal to one-half of GDP by 2083, the end of the 75-year projection period. CBO has no forecasts about how labor supply would have to change to provide those services, for example, or what might happen to health care technology along the way.
LONG-RANGE MEDICARE HEALTH EXPENDITURE PROJECTIONS BY THE CMS OFFICE OF THE ACTUARY
Richard Foster (Centers for Medicare and Medicaid Services) described the long-range 75-year health care expenditure projections for the Medicare program developed by the CMS OACT, which are included in the annual report of the Medicare Trustees to Congress.3 He focused on the long-range component of the projections, although CMS, like CBO, produces short-range (10-year) projections that vary by type of service, such as hospital, physician, or durable medical equipment, and involve far more detail than is the case with the long-range Medicare projections.
Medicare projections are required by statute and must be made in the context of current law—that is, premised on the indefinite continuation of existing statutory provisions pertaining to the Medicare program.4 So the Medicare Trustees report is premised on payment over 75 years of projected benefits as specified under current law and projection of program revenues also as scheduled under current law. In particular, regarding Medicare expenditures, CMS seeks to project the state of the world if benefits now promised under current law were maintained indefinitely. For Part A of the Medicare program, projections of full hospital insurance benefits are compared with revenues already available in the Part A trust fund and projected tax revenues yet to be deposited in the trust fund for payment of Part A benefits. The revenues likely to be available for payment of Part A Medicare benefits do not begin to keep up with the likely level of expenditures, but projected benefits are not reduced after
the projected trust fund is exhausted, leading to a projection of a large Part A funding deficit. For Parts B and D, the annual long-range projection assumes that statutory provisions will remain in force that ensure the availability of revenues no matter how high the expenditures for those parts of the program.
The essential issue in projecting something as volatile as health care expenditures in a program as vast as Medicare over a 75-year horizon is what to assume about future expenditure growth rates. Historically, growth in the health care sector has been much faster than the growth of the overall economy—an almost uninterrupted trend. Cost growth at historic rates clearly cannot go on forever. Economies devoted solely to health care cannot exist. The implication therefore is that there will have to be a slowdown in health care cost growth rates compared with the past. But the question is when, how, and at what rate the excess cost growth will slow down.
Although OACT’s approach for making long-range Medicare cost projections has evolved over a lengthy period going back to the late 1970s, most of the more interesting work started in the 1990s.5 The long-range projection uses a core assumption about the average per beneficiary rate of health care expenditure growth (exclusive of adjustments for age and gender effects) in excess of the rate of growth of per capita GDP for the last 51 years of the 75-year projection horizon. A constant differential or excess cost growth rate of 1 percent above economy-wide per capita GDP growth is assumed on the basis of recommendations received from periodic Medicare technical advisory panels. In producing the final expenditure projection, the core excess cost growth assumption is refined using more complicated modeling methods based on a CGE model that allows the average rate of excess cost growth to be allocated along a more plausible path for the 51 years of the projection horizon to which it applies.
In other words, the idea is that Medicare costs per beneficiary, leaving aside demographic effects, will grow 1 percent faster than the per capita rate of GDP growth. For example, if the nominal GDP growth per capita is 5 percent, then the age- and gender-adjusted Medicare expenditure growth per beneficiary would be 6 percent. This projection method can be implemented with either nominal or real GDP. In practice, CMS does it with nominal dollar projections.
Implementing the Method for the 75-Year Projections
CMS implements this method for the last 51 years of the 75-year projection horizon.6 As noted earlier, for the first 10 years of the projection period, CMS uses far more detailed short-run projections broken out by types of service and other factors. For the first 10 years, therefore, there are projections for Part A, Part B, and Part D, with a distinct growth rate projection for each Medicare subpart. For years 11-25 of the projection horizon, the expenditure growth projections are based on a straight-line transition from A, B, and D excess cost ratios for year 10 to consolidated, program-wide excess cost ratios that begin in year 25. Projections for years 25-75 are based on excess cost ratios from the CGE model.
The OACT CGE model, a Ramsey-style general equilibrium macroeconomic model, allocates consumption through time for a representative agent. The model incorporates assumptions about technological change and cost effects for the health sector. The model is simple in the sense that there are only about three factors. One factor measures historically the impact of change in medical care technology on cost growth and assumes that the same historical rate of technology change continues in the future. The second factor has to do with substitution for new technology—to what extent does new medical technology substitute for existing technology?—in which case there is often a hope that it may be cost reducing, but not always. Or does technology provide something new in the way of treatments or drugs or devices?—in which case it typically adds to the costs. The third factor has to do with the representative agent’s preference for health care—more health care (and, presumably, better health) versus other consumption opportunities—that is, everything that is not health care-related consumption.
One cannot use the CGE model to produce an unconstrained solution, such as a unique best forecast for the future. Based on the range of exogenous parameter estimates deemed to be reasonable, there is an infinite number of potentially reasonable solutions to the CGE model. Therefore, without some identifying constraint, one cannot directly get a usable long-range cost growth projection from the CGE model. A solution is identified by constraining the CGE model to produce a Part A actuarial balance that is financially equivalent in present value to the cumulative 75-year actuarial balance under a simple GDP + 1 excess growth rate assumption.7
For more detail about the CMS projection methods, see the Projections Methodology memorandum referenced in the 2009 annual report of the Medicare Trustees available at http://www.cms.hhs.gov/ReportsTrust Funds/downloads/projectionmethodology.pdf.
For a detailed description of the structure of the model, see Borger, Rutherford, and Won (2008).
History of the GDP + 1 Excess Cost Growth Assumption
In the late 1970s, Medicare projections were not made for more than 25 years for Part A and not more than 3 years for Part B. Part D did not exist at that time. It then occurred to OACT that it would be useful to illustrate the impact of demographic factors on Medicare costs in the longer term. At that time there was no intention of depicting any type of excess cost growth, and the age-gender modeling initiative effectively assumed a rate of per beneficiary cost growth exclusive of age-gender effects equal to the rate of per capita economy-wide GDP growth. The idea was just to build in the demographic factors, prepare a longer range projection, and calculate the cost growth in such way that it was neutral, in the sense that cost growth was assumed to be consistent with economic growth or wage growth. The result was a projection method in which costs were increasing or decreasing at a rate different from the GDP growth rate only because of demographics, highlighting the impact of the demographic factors. Regarding the possibility of excess health care cost growth, this modeling approach was equivalent to a GDP + 0 assumption and was used in ad hoc projections for a number of years.
Through 2000, projections in the Medicare Trustees reports were implicitly based on a GDP + 0 assumption of no excess cost growth for projection year 26 forward. The 1991 Medicare Technical Review Panel suggested that this assumption was reasonable, but, as time went on, public and private enthusiasm for the approach diminished.
CMS charged a technical review panel convened in the year 2000 to make a recommendation concerning a long-range cost growth assumption, and that 2000 Medicare Technical Review Panel eventually recommended the GDP + 1 assumption, which was accepted by the Board of Trustees. On this basis, a pure GDP + 1 assumption for projection year 26 forward was implemented in 2001 and has been used in some form since then. The 2004 Medicare Technical Review Panel reaffirmed the GDP + 1 long-range cost growth assumption as “within the range of reasonable assumptions” (Technical Review Panel on the Medicare Trustees Report, 2000 and 2004). More recently, with the help of an informal panel of advisers, a method for refining the GDP + 1 assumption that uses the OACT’s CGE model to “redistribute” average excess cost growth across the last 51 years of the projection horizon was adopted and implemented first in the 2006 Medicare Trustees report.
The OACT long-range expenditure projection method reflects an expectation of a substantial slowdown in the historical rates of excess cost growth. It also reflects the belief that technological change will remain an important driver of excess cost growth compared with the overall economy for most of the 75-year projection horizon. One other aspect of this core
assumption is that it is relatively easy to describe and to understand what causes what.
Rationale for the GDP + 1 Excess Cost Growth Assumption
CMS argues that three natural brakes on cost growth reconcile the idea of a spending slowdown with the idea of current law. (Note that CBO also assumes a spending slowdown but one that follows a somewhat different path, as described in the previous section.)
The first natural brake on excess cost growth is the cost-sharing and premium provisions in the current Medicare law. In the long run, it is expected that cost-sharing burdens in current law will make Medicare costs grow faster than the income and resources available to Medicare beneficiaries. For all parts of the Medicare program, out-of-pocket costs are growing at about the same rate as program costs, which is faster than people’s incomes. Over time, cost-sharing and premiums for Medicare have grown and become an increasing share of burden on beneficiaries. If nothing happens to change these trends, medical care will become less and less affordable for many categories of beneficiaries. When that happens, and if no legislation has occurred in the meantime to address the issue, beneficiaries will begin to reduce their consumption of medical care, and people may even drop coverage under Parts B and D because they cannot afford the premiums, or they may simply drop their Medigap coverage because premiums have become unaffordable, leading to further scaling back in the level of medical care consumption. The contemplated situation has nasty implications: that the nation’s primary social insurance program for health care could become ineffective because people cannot afford even their share of the cost for it. If that day comes—and it can—a slowdown in health care cost growth would be expected.
Another natural brake on excess cost growth is the spillover or diffusion of cost-saving practice patterns. For example, some innovation happens in the ways in which physicians treat patients who are insured by the private sector, and those innovations then spillover into the treatment of all patients, including Medicare patients. Spillovers have gone in both directions, from private insurance sector to Medicare and from Medicare to private insurance, and they can be helpful to both. For example, Medicare introduced the prospective payment system for hospitals, and soon almost all of the private health insurance plans adopted the same mechanism.
The third possible natural brake on excess health care cost growth involves regulatory changes implementable without statutory changes. One example might be the more selective adoption of new technology. This could be controversial, because there is a provision in the Social Security Act that suggests a little ambiguously that CMS ought not to be making decisions on coverage or approval of payments for new techniques of care on
the basis of whether money is saved or not. However, there have been some preliminary efforts toward greater restraint in the extension of Medicare coverage for new technology and treatment methods, and as time passes and the cost problem becomes more urgent, it is possible that social tolerance for more aggressive restraint on technology approvals will increase even without changes in current law.
OACT Research Efforts
OACT is engaged in research initiatives aimed at informing and improving the choice for the long-range cost growth assumption:
OACT is working on development and refinement of a simple Medicare cost-sharing model. It is also following and advising an effort funded by the National Institute on Aging to incorporate features of Medicare cost-sharing provisions into the Urban Institute’s microsimulation model.
A contract to evaluate evidence for cost-saving spillovers among health care subsectors was recently concluded.
A contract that examines the usefulness of time-series methods for long-term health care cost projections is in its final stages.
OACT is also working on an interface to synthesize evidence pertaining to the excess cost growth assumption—all the literature, different models, different approaches, different perspectives—and using that set of information to inform the choice of a long-range growth assumption.
There are constraints on development of these long-range methodologies:
The requirement to stay within the context of current law in producing projections. Current law necessarily involves scenarios in which the existing program is sustained into the indefinite future. The sustainability of such long-term scenarios is necessarily open to question, a point acknowledged in the annual report of the Medicare Trustees.
The Code of Professional Standards (Actuarial Standard 32 pertaining to social insurance). For example, actuaries are required to model current law; they cannot make assumptions about what they think the law might look like or what it should look like.
Stability in projection methods is desirable; erratic swings in long-term projections due solely to methodological brainstorms would send a confusing message to Congress and the public regarding the financial condition of the Medicare program.
Ongoing tension between complexity and transparency. The CGE model was a big step forward, but it is nearly impossible for non-economists or nonactuaries to understand. It is easier to explain GDP + 1.
In concluding, Foster sounded a cautionary note. Projections in the Medicare Trustees report warn policy makers of the financing crisis for long-term social entitlements. One can discuss for a long time whether CBO or CMS produces a long-range Medicare cost growth projection that is closer to the true long-run magnitudes of the program. But looking at the Medicare Trustees report, the massive deficit for Part A, and the level of expenditures and revenues that would be required to pay for current law benefits under Parts B and D expenditures, both CBO and CMS projections make clear that there is a major financing problem. Although the CMS GDP + 1 assumption envisions a larger spending slowdown than some other projections, it still raises an unambiguous sustainability issue for policy makers. OACT continues research aimed at improving its projections.
MEPS AS A RESOURCE FOR ECONOMIC MODELS AND PROJECTIONS OF HEALTH CARE EXPENDITURES
Steven Cohen (Agency for Healthcare Research and Quality) began his presentation by observing that there are growing demands on data resources in support of health care policy formulation. His presentation covered an overview of a sentinel data resource—MEPS—and how it has been used to inform microsimulation models and public policy questions regarding health care. He also addressed the data capacity and statistical quality of modeling efforts and the underlying requirements for the validity and accuracy of health care cost projections.
The significance of health care expenditure trends is clear when one considers current estimates as well as future projections. One-sixth of U.S. GDP is going into health care spending at present, and the rate of growth exceeds other sectors of the economy. Even after recent cost moderation, the projected rate of expenditure growth will be increasing to 1 of every 5 dollars in the next couple of years. Health care expenditures are among the largest components not only of federal and state budgets, but also of consumer outlays. Cost containment is of continuing concern to both private and public payers.
The most recent information on national health care expenditures shows that in 2008 total expenditures were $2.3 trillion, amounting to 16.2 percent of GDP. This 4.4 percent increase over 2007 is the slowest
growth in 48 years. However, health care expenditures are projected to be $4.4 trillion in 2018 or 20.3 percent of GDP (Hartman et al., 2010; http://www/cms.hhs.gov/nationalhealthexpendData/).
Some of the important current issues in formulating public policy for which data are required include the acquisition of health insurance coverage by the uninsured and its implications in terms of expenditures; the structure of the insurance market; the tax treatment of insurance and the federal subsidy for employer-sponsored coverage; the cost of chronic diseases and prevention activities and how that factors into the long-term projections; and prescription drug costs.
MEPS, sponsored by AHRQ, is an ongoing family of surveys. Cohen focused mostly on the household component, which is an annual survey of approximately 14,000 households covering about 30,000 individuals. The survey provides national information on health care use, expenditures, insurance coverage, sources of payment, access to care, and health care quality. In addition to aggregate estimates, MEPS permits studies of the distribution of expenditures and sources of payment, such as the concentration of expenditures among population groups; the role of demographics, family structure, and insurance coverage in health care costs; expenditures for specific conditions; trends over time, such as the persistence of the concentration of expenditures; and impacts of changes in employment and changes in insurance coverage on health care use and expenditures.
Key Features of the Household Component
The household component of MEPS is a survey of the civilian non-institutionalized population. It is a subsample of respondents to the National Health Interview Survey, which is conducted by the National Center for Health Statistics. The survey oversamples minorities and other policy-relevant groups. The fact that it has an overlapping panel design allows for analysis over a 2-year window. A new panel is introduced each year and carried over into a second year; thus there are two representations of the population each year. Continuous data collection over a 2 1/2-year period includes 5 computer-assisted personal interviews. Data from the first year of a new panel are combined with data from the second year of the previous panel for estimation. This design is very helpful for short-run microsimulation modeling, in which one can use one panel and then validate the model and the predictions using the second.
The household component has a number of capabilities for projections and simulation:
It provides estimates of annual health care use and expenditures.
It provides distributional estimates.
It supports both person- and family-level analysis.
It tracks changes in insurance coverage and employment.
The longitudinal design allows linkage to a prior year from the National Health Interview Survey.
Having obtained baseline information on health status, roughly a quarter of the sample is interviewed each year to obtain detailed demographic information. Particular attention is given to the sample of individuals with high health care expenditures or those who are likely to incur high levels of expenditures, both in terms of optimizing response rates and obtaining additional information on expenditures from their medical providers. That is critical, considering that the top 1 percent of users accounts for 27 percent of total health care expenditures and has a significant impact on the precision of overall survey estimates. These individuals include decedents and people who are in or likely to enter long-term care facilities and lengthy hospitalizations.
In addition, to correct for sampling error, the estimates of decedents are adjusted to national estimates for mortality, and the estimates of people admitted to nursing homes are adjusted to more precise survey estimates. One limitation, in terms of making national estimates, is that MEPS covers the civilian, noninstitutionalized population; it does not cover the nursing home or other institutionalized populations.
MEPS has been useful in estimating costs for chronic diseases in a given year and over time. This information is important for high-prevalence conditions for which there could be interventions and to calculate, at least in the short run, the impact in terms of health outcomes and expenditures. Some of the highest cost conditions in 2007 included cancer, trauma, heart disease, mental disorder, pulmonary conditions, diabetes, hypertension, osteoarthritis, hyperlipidemia, back problems, upper gastrointestinal disorders, cerebrovascular disease, kidney disease, skin disorders, and other circulatory conditions. The costs of these conditions for 2007 ranged from about $20 to $98 billion.
Medical Care Provider Component
MEPS does not rely solely on household data. It also includes a medical care provider component to obtain greater accuracy and detail on household expenditures provided by households, to compensate for household
item nonresponse, and to serve as a source for imputation for the remaining missing items. The medical care provider component supports methodological studies.
A targeted sample is drawn to reach all associated hospitals and associated hospital-based physicians, all associated office-based physicians, all associated home health agencies, and all associated pharmacies. (Associated hospitals and other medical care providers are those used by respondents to the household survey.) Data are collected on dates of visits, diagnosis and procedure codes, and charges and payments.
Another part of the MEPS family of surveys is the insurance component, data from which are valuable for cost projections. This component is an annual survey of 40,000 establishments to obtain national and state-level estimates of employer-sponsored coverage, including availability, access, cost of health insurance, and benefit and payment provisions of private health insurance.
A number of questions can be answered by the health insurance component of MEPS:
How does the cost and availability of coverage for workers vary in different economic and employment circumstances, and what are the implications of Medicare Part B coverage on the retiree benefits structure?
How do payment policies affect employee decisions about the purchase and selection of health care services and health insurance?
What are the implications of Medicare Part D prescription drug coverage on consumers, employers, and employees?
Uses of MEPS Data to Inform Health Policy
AHRQ has been able to provide Congress and others with research findings to inform health care policy on coverage trends and costs, such as national estimates of the long-term uninsured in terms of what the cost provisions would be of covering the uninsured, not just at a point in time but over a 2-year period; estimates of the number of uninsured children eligible for the State Children’s Health Insurance Program; state estimates of the availability and cost of employer-sponsored coverage; concentration of health care expenditures; and premium percentiles of high-cost plans.
Some of the areas of research using MEPS data include access, use, and quality of health care services; levels and trends in expenditures; private and public health insurance; and health conditions and health behaviors. MEPS
is also used for microsimulation modeling and for research on survey and estimation methods.
Modeling and Simulation Efforts
In the prior decade, the National Medical Expenditure Survey, the predecessor of MEPS, was used in models of the impacts of proposed health care reforms, including the costs of reform to households, the costs to the nation, changes in coverage, and tax impacts. Today using MEPS data, these capabilities remain the survey’s strength, with the addition of a Medicaid/Children’s Health Insurance Program eligibility simulation model; data on expenditures by service, including prescription drug expenditures; estimates of coverage and expenditures for most populous states; improved tax simulation models; and data from the employer health insurance survey by state.
Attributes of Modeling
Cohen next addressed some of the statistical dimensions in health care modeling that are important to consider in deciding on a database and model specification and in determining the credence to give to the model results for short-term and longer term projections by policy analysts.
Selection of host analytical database/data capacity for a particular underlying projection or microsimulation model—issues of content, national and subnational representativeness, sample size, data quality, timeliness, and accessibility, all would enter into the decision. For example, if one is looking for national estimates, such as a change in coverage and how that affects use, expenditures, and access to care, a survey like MEPS would be relevant, particularly for its strength of expenditure data. But if one is looking for state-specific differentials, one might turn to a survey like the Current Population Survey (CPS), which has state-level capability on insurance coverage. Because CPS does not have expenditure data, many modelers use CPS, with its strengths for insurance coverage, but impute all of the expenditure data.
Model specifications—the decisions on specifications depend on whether the model is to address a distinct set of highly related health care policy questions (specificity) or whether it needs flexibility and utility for addressing a more expansive set of policy questions.
Analytical and statistical oversight—the more a model is based on sound statistical theory and practice, the more the specifications
for the model are tested, and the more the products are subjected to rigorous statistical and substantive review, the more trustworthy its results.
Methodology—documentation of models would include a description of the underlying approach, the survey methodology, the final model specifications, and the results of statistical tests for model fit and error. Static or dynamic approaches to aging would also be clearly described to facilitate understanding and replication.
Replication—the more a model is subject to sensitivity testing and replication, the more credence can be given to its results and the more its limitations can be understood.
Precision—error estimates associated with sampling, imputation processes, and nonsampling errors, which include errors associated with model specification, nonresponse measurement, coverage, and population projections, all need to be provided and documented.
Transparency and good documentation—a summary of uncertainty of estimates, an evaluation of performance, and release of code and audit trails are essential.
Reconciling MEPS and the National Health Expenditure Accounts
The National Health Expenditure Accounts (NHEA), developed by CMS and MEPS, provide the two most comprehensive estimates of health care spending in the United States. Reconciling estimates from both sources serves as an important quality assurance exercise for each. This exercise is critical to development of an adjusted MEPS data set, consistent with NHEA.
The adjusted MEPS data yield a consistent baseline for policy simulation studies. The baseline reconciles MEPS and NHEA by service categories and sources of payment for the MEPS population; poststratifies to up-weight the Medicaid population and high-expenditure cases; closes the remaining gap by scaling expenditures by service categories and payment source; and adds back in selected NHEA components that were removed in the reconciliation.
Cohen summarized his presentation by observing that to complement assessments of the current state of health and health care, policy makers depend on model-based estimates of the future state under alternative demographic, economic, and technological assumptions. These modeling efforts are major benefits of the existing investments in health and health care data collection, as well as initiatives to ensure that such collection yields efficient,
well-coordinated, integrated policy-relevant data sets. However, they also place additional demands on data capacity, research, model development, and statistical standards and rigor to better assess the impacts of revisions to existing health care policies.
He noted the importance of aligning projection modeling efforts with more conventional statistical analyses by providing metrics that convey levels of uncertainty in model outputs. The attributes he presented of the modeling process emphasize the need for standards of data quality and statistical integrity in support of modeling and microsimulation efforts that are comparable to those developed for “current state” analyses. This is essential to ensure that policy makers have a sound understanding of model assumptions, data limitations, and the level of uncertainty associated with model-based estimates, prior to the implementation of a new initiative.
Cohen also observed that, in recent years, AHRQ has been getting a number of calls not only for cost projections but also for analyses of health insurance coverage and access. In an attempt to be transparent, the agency posts on its website the requests and uses of AHRQ data, whether they are from Congress, the U.S. Department of Health and Human Services, or the White House.
Participants had comments and questions on CBO’s assumptions on excess cost growth in the private and public sector, the issue of level of enrollment in Part B, the requirement for CBO and CMS to stay within current law, and the role of taxes as a constraint on the growth of health care spending.
Joseph Newhouse (Harvard University) asked what the basis was for CBO’s assumption that excess cost growth rates in the private sector would slow down at three times the rate of excess cost growth for Medicare. Joyce Manchester responded that without a slowdown in excess cost growth rates, health care spending would amount to 100 percent of GDP by the end of the projection period, an untenable result. Both the private sector and the states would exert tremendous pressures to slow the growth rate of spending on nonfederal health care. Recall that CBO is constrained to look at current law or current policy for Medicare and cannot assume any major reforms. Under that assumption, only spillover effects from medical practice patterns in the private sector and Medicaid would reduce the rate of excess cost growth in Medicare.
Justin Trogdon (RTI International) questioned assumptions about the willingness of households to spend on health care versus other consumption. He asked what kind of utility maximization problem would lead to that kind of decision and suggested that some sort of multistage budgeting
would be another way to motivate the assumption. Manchester responded that CBO had adopted a simple rule but could spend more time motivating that rule if the agency wanted to justify it. An alternative approach would be to take a big step back and develop a different way to go about the problem.
Michael Chernew (Harvard University) commented on the requirement that CBO stay within current law. In a current law framework, how disastrous would it be if the forecast ultimately ends up being something that is essentially not sensible? Is it the case that the agency simply cannot go forward with that or is that in and of itself information?
Manchester observed that the current law framework affects all of CBO’s long-term projections. For example, current law leads to sharply rising ratios of debt to GDP that could have disastrous consequences for the economy if left unchecked. To produce a baseline for policy reform, however, CBO makes the simplifying assumptions that the rise in the debt-to-GDP ratio will not have an effect on how the economy operates and, in particular, that the real interest rate will remain constant at 3 percent for 75 years. In addition, without the arbitrary rule on nonhealth care consumption that brings health care spending down to one-half of GDP, the health care sector by itself under current law would account for 100 percent of the economy. Again, current law produces an untenable situation, so something has to give. CBO’s approach is one way of illustrating to Congress how the current situation is unsustainable and, at the same time, providing a baseline against which to measure reform. CBO is trying to develop better ways to illustrate the unsustainable nature of the current situation to Congress. Concentrating on the next 25 years and showing the consequences of the current path for 25 years may be an alternative way to present the information.
Richard Foster commented that, in contrast to CBO, which has an assumed long-range growth rate for Medicare that is greater than for both Medicaid and private health insurance, CMS assumes that all parts of the U.S. health care sector will grow at about the same rate before demographic effects come into play. The primary reason for this assumption is that while over the short-run health care costs have grown at different rates for different parts of the U.S. health sector, it is difficult to discern long-run differences in cost growth rates across the health care sector. Looking to the future, much of the future health care cost growth, other than demographics, relates to technology—that is, new technology. If that is the case and if, in the long-term future, for example, Medicare costs were to grow faster than private health insurance costs, then that would tend to suggest that Medicare beneficiaries would get all of the new technology that comes along and privately insured persons would not. That scenario is simply not plausible. That is a primary reason that CMS assumes that all parts of the health care sector grow at about the same rate.
In response to a question from Richard Suzman (National Institute on Aging) as to whether CBO has anything in its model on the macro implications of a growing fraction of GDP going into health care, Manchester stated that CBO does not explicitly model how the economy could allocate 50 percent of GDP to health care. That is an issue that warrants further attention.
Dana Goldman (University of Southern California) had two questions. First, he noted that increasing medical care spending presumably leads to better health and longevity. Is this information incorporated into any of incorporated into any of the models? Second, with these projections of rising costs, there is concern that there may not be universal enrollment in Medicare Part B. At the present time, it is at about 97 percent. However, as the cost of premiums gets higher, people may opt out, and therefore spending may be lower. Has there been any effort to model that?
In response to the first question, Foster said that CMS has not explicitly taken into account improved health status and its effect on health care costs in the future, although it does that somewhat implicitly. Clearly, health status is improving generally. The question of what happens to health care costs with better health status is often posed. Does CMS sufficiently take account of possible improvements in the overall population health status in its projections? Under currently available methods, direct feedback of such effects into the projection models is not realistic because they have not been able to answer adequately the question: Is improved health status the cause of lower expenditures or is it the result of higher expenditures?
With regard to the question about Part B take-up—that is, the percentage of eligible people actually enrolled in Part B—Foster said that CMS has considered whether the gradual reduction of the take-up rate that has been observed is related to increasing costs. One would expect the reduction to continue and at some point become critical, but at present CMS does not have a good answer as to why less than 100 percent are enrolled in Part B.
Marilyn Moon (American Institutes for Research) suggested that some of the lack of take-up of Part B may be related to the fact that federal employees who are enrolled in health maintenance organizations do not need Part B. She asked Foster if he has a sense of the problem at this point. Foster responded that CMS did not have information at this time, but the question would be part of any study to figure out why people are not enrolled in Part B. Another aspect of this problem is that, with the introduction of the income-related Part B premium, some numbers of people are expected to drop out—if a beneficiary faces paying as much as 80 percent of the cost of premiums, then it may not be a good deal for such a person to continue Part B coverage.
Moon was struck by the emphasis Foster put on the assumption that over time some slowdown in Medicare spending would occur because of
higher cost-sharing. To some degree that seems to be fully consistent with CBO’s assumption that people do not want to drop all spending on everything else. But CBO does not envision under current law a substantial slowdown in Medicare spending. Foster maintained that significant changes in health care expenditure growth may reasonably be assumed, even with no change in current law.
Chernew asked if MEPS or other data would make it feasible to track the availability of, and project going forward, the prevalence of retiree benefits that cover a lot of the cost-sharing gaps in Medicare.
Foster stated that it is not an area on which CMS focuses directly. He referred the question to Steven Heffler (Centers for Medicare & Medicaid Services) regarding the extent that OACT, in its private health insurance data, looks separately at retiree health care benefits. Heffler explained that most of the projections of coverage levels are relatively aggregated for, say, total private health insurance enrollment or employer-sponsored insurance. But in each category there is a mix of things occurring. He remarked that one lesson learned in doing the health care reform estimates has been a deeper appreciation of the potential impacts on different groups and a greater need to understand them. He expects there will be more efforts to disaggregate categories to better understand what is happening to different coverage groups.
Cohen pointed out that perhaps the best data resource to inform this issue would be an actual linkage from the MEPS household survey with its establishment survey; currently, they are separate entities. AHRQ staff internally have gone through an exercise of statistical matching. In the past they used the household survey, went to the employers, and got the benefit information. However, with all the problems in obtaining permission forms, they had concerns about the accuracy of the data. So there is quite a bit of capability with the data resources with statistical matching, but it does introduce another source of error. Still, it is the best resource available and certainly viable for answering some of the questions.
Miron Straf (National Research Council) asked how sensitive the cost projections are to the different ways of developing population projections, including those that look at trends in lifestyle, diet, and the like, and the later onset of disability and some diseases. Foster responded that one of the key sensitivities has to do with the different assumptions that have been made about improving life expectancy. If one looks at work by Lee and Tuljapurkar (Lee, 2004) compared with what is done by the Office of the Actuary at the Social Security Administration or by the Census Bureau, one does see some sensitivities to different assumptions, particularly for health care costs. It is one thing to look at the cost of a social insurance program relative to taxable payroll and GDP, in which the number of beneficiaries versus the number of workers is very important, so that faster or slower
declines in mortality affect the ratio of workers to beneficiaries. It is another thing entirely to look beyond that at the age pattern of health care costs. The costs are far higher at older ages, and that raises an interesting question: With longer life expectancy and more people living to older ages, will the future elderly have the same pattern of health care costs as today’s elderly, or will their costs be more similar to today’s younger beneficiaries of the Medicare program? The OACT has started to explore this issue based on a suggestion from David Cutler (Harvard University) to look at expenditures for survivors in a year versus decedents in a year as a gross approximation of health status. Progress was made in this effort, but the project eventually had to be put aside because of resource constraints.
Jonathan Skinner (Dartmouth College) questioned the role of taxes as a constraint on growth in health care spending. One of the things he and his colleagues found is that countries seem to bump up against tax constraints at about 40 percent of GDP. They do not like to tax more than that. Denmark and Sweden, which have very high tax rates to begin with, have held the line on health care spending in terms of keeping their growth in spending during the last 30 years to 1 or 2 percentage points of GDP increase, unlike the United States, where health care costs are growing at much faster rates. Have there been any thoughts in this country about constraining health care cost growth by holding the line for collecting no more than 40 percent of GDP in taxes?
Foster remarked that the OACT has had some interesting discussions along those lines. What is a tolerable or sustainable level of revenue collection? A few years ago, there was a rash of models developed by others about the long-term growth of government spending, and many of these ended up projecting unrealistic high levels.
Some of the questions raised in this session about the constraints imposed by assuming current law in cost projections and what brakes could be put on health care spending were also discussed in the next session (see Chapter 3).