Geographic variation in health care: the role of private markets.
Philipson, Tomas J. ; Seabury, Seth A. ; Lockwood, Lee M. 等
ABSTRACT The Dartmouth Atlas of Health Care has documented
substantial regional variation in health care utilization and spending,
beyond what would be expected from such observable factors as
demographics and disease severity. However, since these data are
specific to Medicare, it is unclear to what extent this finding
generalizes to the private sector. Economic theory suggests that private
insurers have stronger incentives to restrain utilization and costs,
while public insurers have greater monopsony power to restrain prices.
We argue that these two differences alone should lead to greater
regional variation in utilization for the public sector, but either more
or less variation in spending. We provide evidence that variation in
utilization in the public sector is about 2.8 times as great for
outpatient visits (p < 0.01) and 3.9 times as great for hospital days
(p = 0.09) as in the private sector. Variation in spending appears to be
greater in the private sector, consistent with the importance of public
sector price restraints.
**********
There is considerable variation in health care utilization and
spending across geographic areas in the United States, but little
evidence of corresponding differences in health outcomes or satisfaction
with care. (l) This variability is often cited as evidence that current
levels of health care spending reflect "flat-of-the-curve"
medicine, that is, treatment for which the marginal benefit of an
additional unit of care is approximately zero. Interpreted this way,
these findings have dramatic implications for the potential to increase
the productivity of health care spending, and for this reason they have
figured prominently in the policy debate.
However, the evidence on regional variation is almost exclusively
limited to the public sector, because it relies on Medicare data. Less
is known about the corresponding patterns in the private sector. A
venerable literature in economics has argued that private firms and
their managers have stronger incentives to restrain costs and boost
efficiency than their public counterparts. (2) In the health insurance
context, Medicare does not face competition over premiums that might
otherwise restrain its costs, and unlike private sector firms, Medicare
does not have direct residual claimants whose standard of living
improves with the efficiency of the enterprise.
To develop the implications of these incentive differences, this
paper provides a theoretical and empirical analysis of how regional
variation in health care differs across the public and the private
sectors. We first examine conceptually how private efforts to control
costs within a region, through selection of providers, might translate
into differences in care across regions. In particular, our analysis
implies that utilization controls within regions in the private sector
should lead to lower regional variation in the private sector than in
Medicare. However, the implications for variation in spending are less
clear, because Medicare may also be able to better control prices
through its greater monopsony power. If the private sector controls
utilization while the public sector controls prices, the result is an
ambiguous prediction for variation in spending.
We examine these implications empirically using individual-level
data on patients with heart disease, comparing utilization and spending
on patients who have private insurance with that on similar patients
within Medicare. Data on the former come from a large database of
private sector medical claims, and on the latter from the Medicare
Current Beneficiary Survey. Both datasets include patient-level
demographics and co-morbidities, which allow us to identify regional
variation distinct from individual characteristics such as health. The
focus on heart disease helps mitigate the confounding impact of regional
differences in health status on our analysis.
Our main object of interest is the regional variation in
utilization and spending across sectors that cannot be explained by
variation in patient characteristics. Our data suggest greater variation
in utilization in the public sector: our main analysis suggests that
variation in the public sector is about 2.8 times as great for
outpatient visits (p < 0.01) and 3.9 times as great for hospital days
(p = 0.09) as in the private sector. There is some evidence of greater
variation for the number of hospitalizations in the public sector, but
this evidence is less robust. Prescription drug utilization serves as
our "placebo" case of insurance that was privately provided in
both samples during the period investigated. Significantly, and unlike
other types of medical care, drug utilization exhibits less variation
among Medicare patients. On the other hand, there is greater spending
variation in the private sector, suggesting the potential importance of
monopsony power in the public sector.
The paper proceeds as follows. Section I provides the conceptual
analysis of how differing cost-control measures within a region might
lead to differences in regional variation in utilization and spending.
Section II reports our empirical analysis comparing regional variation
in the public and the private sectors. Section III discusses how our
findings relate to the existing literature on health care variation and
the resulting policy implications. Section IV discusses some limitations
of our analysis and presents several robustness tests. Section V
concludes.
I. A Simple Analysis of Regional Variation in Utilization and
Spending
This section presents a simple analysis of how private and public
incentives interact to create different degrees of regional variation in
health care utilization and spending between the public and the private
sector. (3) A key assumption is that private insurers have stronger
incentives to restrain costs and utilization than a public insurer such
as Medicare. This assumption is based on the literature demonstrating
that, unlike public enterprises, private firms have to restrain costs in
order to compete on price, and private firms' inefficiencies have
direct impacts on the welfare of their owners and employees (Alchian and
Demsetz 1972, De Alessi 1974a). For example, private payers may
explicitly manage care and exert pressure on providers through
utilization review and case management. They can also selectively
contract with lower-cost providers, steer patients to preferred
providers, and exclude inefficient doctors or hospitals. In addition,
prior authorization of large expenditures is prevalent in the private
health insurance sector, a practice that allocates major spending
decisions to the payer rather than the provider. Finally, private payers
can steer patients toward efficient care through benefits
management--for example, by not covering certain services unless certain
clinical criteria are met. In what follows, we use the shorthand of
"utilization restrictions" (UR) to refer to all these
practices.
We interpret UR as a limit on the provision of treatments whose
costs exceed their benefits. This may still lead to regional variation
in utilization, because there is substantial heterogeneity among
apparently similar patients in the efficacy of different treatments.
Excessive care for one patient may be cost-effective for another.
I.A. Causes of Sectoral Differences within Regions
We first consider the level of utilization in both the private and
the public sectors. Define [y.sup.*] as the efficient utilization level,
that at which marginal benefit equals marginal cost. Following the
earlier literature, we assume that private insurers have stronger
incentives to limit utilization that rises above this level. They do
this through UR, which we assume places an upper bound on utilization,
[y.sub.UR] [greater than or equal to] [y.sup.*], and perfectly
eliminates inefficient utilization above that level. (4) The assumption
of full efficiency is an analytical simplification; the positive
predictions do not depend on it, and we do not emphasize the normative
predictions.
Within any region there is a distribution of providers, who vary in
the level of care they would provide to an identical patient. We
characterize this distribution using the cumulative distribution
function F(y) for the random utilization variable E Private payers'
UR procedures limit utilization and thereby truncate the support of the
providers participating in their plans. This results in the private mean
utilization level, [mu] = E(Y/Y [less than or equal to] [y.sub.UR]).
This constrained private sector mean is thus lower than the
unconstrained public sector mean, [[mu].sup.p] = E(Y).
[FIGURE 1 OMITTED]
Now consider a pure increase in utilization, holding health status
fixed. This can be represented as a rightward shift in the function
F(y). Assuming the efficient level of utilization remains fixed, the
result is a greater difference in mean utilization across the two
sectors, [[mu].sup.p] - [mu]. In other words, in regions with providers
who have greater tendencies toward inefficiency, the difference in
utilization between sectors will be larger.
The second key assumption is the presence of greater monopsony
power in the public sector. The result is greater restraint of prices,
as opposed to utilization, in the public sector. This affects the
analysis of variation in spending, which combines the utilization effect
and the price effect. If the government pays below-market prices through
the exercise of either monopsony power or direct price regulation, the
cost curves will differ across sectors. The result is depicted in figure
1. Average spending per patient in the private sector may exceed that in
the public sector, if equilibrium marginal cost in the public sector,
[MC.sup.p*], is less than equilibrium marginal cost in the private
sector, [MC.sup.*].
I.B. Causes of Sectoral Differences across Regions
Next consider how mean utilization for each sector might vary
across regions. Define the joint distribution G([[mu].sup.p], [mu]) of
mean utilization levels across regions. Specifically, suppose that the
underlying distribution F(y) differs across regions. Figure 2
illustrates how one might then characterize the relationship between
changes in the public mean and the mean difference between sectors:
d[[mu].sup.p] - [[mu].sup.p]/d[[[mu].sup.p] = 1 -
d[mu]/d[[mu].sup.p].
[FIGURE 2 OMITTED]
For example, consider the case of normally distributed public
sector utilization, Y ~ N([[mu].sup.p], [[sigma].sup.2]). In this case,
mean utilization in the private sector follows from the formula for the
mean of a truncated normal random variable, [mu] = [[mu].sup.p] +
[sigma][lambda]([alpha]), where [lambda]([alpha]) [equivalent to]
[phi]([alpha])/[PHI]([alpha]) is the inverse Mills ratio and [alpha] =
[y.sub.UR] - [[mu].sup.p]/sigma. This implies that the slope of the
private mean as a function of the public mean is less than unity or,
equivalently, that the between-sector difference rises with the public
mean: (5)
0 [less than or equal to] d[[mu]/d[[mu].sup.p] [less than or equal
to] 1
d([[mu].sup.p] - [mu])/d[[mu].sup.p] [greater than or equal to] 0.
When the public sector provides more care above the efficient
level, this raises the between-sector difference. This in turn implies
that the variance in the regional means in the public sector will exceed
the variance in the regional means in the private sector:
V([[mu].sup.p]) > V([mu]).
This simple framework leads to several testable empirical
predictions: Private provision should lead to lower mean utilization and
less variance in mean utilization across regions, but not necessarily
lower mean spending. In addition, the difference in utilization between
sectors is likely to rise with the mean level of public utilization.
Note that all these predictions hold patient health status constant.
II. Empirical Analysis of Regional Variation across Sectors
In this section we describe our empirical analysis of regional
variation in the public and private sectors aimed at testing the
implications discussed above.
II.A. Data and Empirical Specification
We compare regional variation between a sample of privately insured
patients and a sample of Medicare patients. The private data come from a
large database of health insurance claims. The data capture all health
care claims, including prescription drugs and inpatient, emergency, and
ambulatory services, by employees and retirees while they are enrolled
in the health plans of 35 Fortune 500 firms. The analytical database
integrates component datasets of medical claims, pharmacy claims, and
enrollment records. This allows us to calculate spending and utilization
for all services provided to the patients over our study period. The
enrollment records allow us to identify basic demographics of the
patients, including age, sex, and some information on income. (6)
Importantly for our purposes, the data also include information on area
of residence, coded by metropolitan statistical area (MSA) and 3-digit
zip code. This allows us to analyze health care spending and utilization
patterns at different levels of geographic aggregation.
Our Medicare sample is taken from the Medicare Current Beneficiary
Survey (MCBS), which is administered to a nationally representative
sample of aged, disabled, or institutionalized Medicare beneficiaries.
Respondents, whether living in the community or residing in health care
facilities, are interviewed up to 12 times over a 4-year period.
Institutionalized respondents are interviewed by proxy. There is
oversampling of the disabled under 65 years of age and of the oldest old
(85 years of age or older). The MCBS uses a rotating panel design with
limited periods of participation. Each fall a new panel is introduced
with a target sample size of 12,000 respondents, and each summer a panel
is retired. The MCBS data include detailed information on self-reported
health status, health care use and expenditure, insurance coverage, and
demographic characteristics. Additional Medicare claims data for
beneficiaries enrolled in fee-for-service plans are also incorporated to
provide more accurate information on health care use and expenditure.
The MCBS data do not include actual claims data on prescription drugs;
all information on prescription drug spending and utilization in the
MCBS is self-reported. This leads to a known undercount of drug spending
in the MCBS. (7)
Both datasets include information on medical claims that is used to
compile utilization, spending, and baseline health information. That is,
although the MCBS contains a survey component, all data on spending and
utilization are compared with Medicare's administrative claims data
(Eppig and Chulis 1997). However, since Medicare does not cover
prescription drugs over our sample period, this validation procedure
applies to medical care but not drugs. Finally, for both datasets we use
information from 2000 to 2006. The one exception is prescription drug
utilization and spending: to abstract from the complexities of Medicare
Part D's introduction, we eliminate the 2006 data for these
variables.
To mitigate differences in health status across sectors and
regions, we condition inclusion in the sample on a diagnosis of ischemic
heart disease (IHD). (8) We also use the diagnosis codes on medical
claims to identify whether patients were treated for any of 30 different
conditions in a calendar year. (9) The claims-based measures of the
number of diseases are available in both the MCBS and the private health
insurance data. (10) This is important because unmeasured differences in
severity across regions could lead to spurious positive correlation
between sectors.
The primary geographic unit of analysis for our study is the MSA.
An alternative candidate would be the hospital referral region (HRR),
used by the Dartmouth Atlas. However, HRRs are not reported in either of
our datasets, and the private sector data do not contain 5-digit zip
codes, which are required to construct an individual's HRR. We
restrict our sample to the 99 MSAs where we have the largest samples.
MSAs are somewhat larger than HRRs, and this may compress the variation
for both sectors in our data.
Our final sample contains 240,028 private patients and 24,800
public patients. (11) Since there are many fewer public patients, it is
important to correct for the effects of sample size on our estimates. We
derive and report these corrections in detail below.
Table 1 reports some summary statistics comparing demographic
characteristics in the public and the private samples. As one would
expect, the average age in the private sample is lower than in the
sample of Medicare patients, most of whom are older than 65. The private
sample contains a greater fraction of males, in part because it is
influenced by current or past employment status. (The private sample
contains both active workers and retirees receiving benefits from their
current or past employers.) Average income is also higher in the private
sample. The greater variance in income for the public sample is likely
due to the fact that income is reported individually in the MCBS, but
imputed at the local level in the private sample.
The table also compares the health of individuals in the two
samples. Since both samples are limited to individuals with a history of
heart disease, we include a variable indicating the fraction of
individuals who are diagnosed with heart disease in a particular year.
In all cases, the presence of disease is taken from claims rather than
from self-reported data. The incidence of heart disease is similar in
the two samples: 0.32 in the private sample and 0.37 in the public
sample.
In addition, the table reports the average number of adverse health
conditions (out of the total of 30, including heart disease) per
patient. As with heart disease, the health conditions are determined
using the ICD-9 diagnosis codes from medical claims in both the public
and the private samples. Unsurprisingly, the elderly individuals in the
public sample are much sicker on average, with 2.9 adverse health
conditions in the year compared with 1.4 in the private sample.
As a matter of course, the public and the private samples are drawn
from different populations. We include a number of controls and analyses
designed to mitigate and test for the impact of these differences, but
heterogeneity across samples remains a possibility. Later we discuss the
sources of heterogeneity, the methods we have employed to address them,
and their possible implications for the analysis.
II.B. Descriptive Statistics
Table 2 presents some descriptive statistics for health care
spending and utilization in the public and the private samples
aggregated over all regions and patient characteristics. We present not
only the mean and the standard deviation but also the 25th-percentile,
median, and 75th-percentile values. Our utilization measures (all
measured as yearly averages per patient) include the number of
hospitalizations, total hospital days across all hospitalizations, the
number of outpatient visits, and the number of 30-day-equivalent
prescriptions in both samples. For spending, we record total (inpatient
plus outpatient), inpatient, and outpatient spending, as well as
spending on prescription drugs.
Utilization, in terms of hospitalizations, hospital days, and
outpatient visits, is lower for the private patients. Spending for this
group also tends to be lower. Total medical spending for individuals in
the private plans is $8,401 per year, compared with $10,245 for the
Medicare patients--about a 20 percent difference. The exception to the
pattern is prescription drugs, for which both utilization and spending
are greater among private patients.
Figures 3 and 4 provide a broad sense of the variation present in
our samples. Figure 3 reports for both samples the estimated kernel
densities of MSA-level deviations from the mean for both hospital days
and outpatient visits. Each data point underlying the kernel estimate is
the difference between an MSA-level mean and the overall sample mean.
For both variables, the distributions appear to be tighter for the
private than for the public sample. However, these distributions are
based on raw, unadjusted numbers that do not account for disease or
other covariates.
[FIGURE 3 OMITTED]
[FIGURE 4 OMITTED]
Figure 4 repeats this exercise for inpatient and outpatient
spending. Here the findings are decidedly more mixed. For outpatient
spending the distribution appears to be slightly tighter for the public
sample. The figure for inpatient spending is harder to interpret
visually, as the differences in the densities are small and asymmetric.
In any event, the differences observable visually between the spending
and the utilization distributions suggest the possible importance of
public sector price restraints, which would lower spending variation
even with greater variation in utilization.
[FIGURE 5 OMITTED]
Finally, figure 5 plots the relationship between deviations from
the MSA-level means for public and private hospital days. This is the
empirical analogue to the theoretical relationship in figure 2. The
figure suggests that mean private hospital days increase slightly with
mean public hospital days, but much less than one for one. This is
consistent with there being less regional variation in the private
sector; we test this hypothesis more formally in the following analyses.
II.C Framework for Estimating Regional Variation
We are particularly interested in the between-MSA variance in
spending and utilization for the public and the private samples. We
begin with the simplest possible approach that evaluates the variance
between MSAs in the sample means. We then move to estimating the
variance in regression-adjusted means, which we estimate from
regressions that control for various factors that might also influence
spending and utilization. In both cases we account for the relative bias
that is created by the substantial differences in sample size across
sectors: because the public samples are much smaller than the private
samples, there is greater sampling variance in the public sector
estimates and thus greater variation in the MSA-level means for Medicare
patients. To estimate the true between-sector differences in regional
variation, we estimate and remove the variability that is due to sample
size differences alone.
Formally, the observed regional variation within a sector is due to
the true variation and the sampling variance in estimating that
variation. Denote by [[mu].sub.r], the true mean for region r and by
[[??].sub.r], the corresponding sample estimate, whether unconditional
or regression-adjusted. The sample mean is equal to the true mean plus
sampling error, according to
[[??].sub.r] = [[mu].sub.r], + [z.sub.r].
The sampling error [z.sub.r] has zero mean, and the covariance of
the sampling error across regions is E(z, [z.sub.s],) =
[[sigma].sub.rs]. Define [bar.[mu]][equivalent to] 1/R
[[summation].sup.R.sub.r=1][[mu].sub.r], the "grand mean"
across regions. Similarly, define the corresponding sample analogue,
[bar.[??]][equivalent to] 1/R [[summation].sup.R.sub.r=1][[??].sub.r].
Finally. define the average sampling error across regions,
[bar.z][equivalent to] 1/R [[summation].sup.R.sub.r=1][z.sub.r]. The
object of interest is the degree of regional variation in the true MSA
means, RV [equivalent to] 1/R [[summation].sup.R.sub.r=1][([[mu].sub.r]
- [bar.[[mu]).sup.2] which has the sample analogue [??][equivalent to]
1/R [[summation].sup.R.sub.r=1][([??].sub.r] - [bar.[??]].sup.2].
The observed variation [??] is a biased estimator of RV, as a
result of sampling error in the estimates. Moreover, this bias is likely
to be larger for our public sector estimates because of the smaller
public sector sample size, which yields noisier estimates of public
sector utilization. However, we can recover a consistent estimate of the
bias and correct for it, according to
E([??] = RV + 1/R [R.summation over (r=1)] [E([z.sub.r] -
[??]).sup.2]
In the appendix we show how to estimate this bias from sample
variances and covariances. Our formula works for both the case of
unconditional sample means and the case of regression-adjusted means. In
the simple case without covariance across regions or zero average error
across regions, this expression simply states that the observed
variation is the true variation plus the average squared standard error.
More generally, the more precisely the sample means are estimated, the
smaller is the bias correction.
In sum, the object of interest in our analysis is RV, which we
estimate as [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] for both
the public and the private sector. Using these estimates of regional
variation, we report both the ratio of public to private variation and
the difference between public and private variation. We construct
standard errors around these by means of a bootstrap procedure, which
samples individuals with replacement within each MSA, so that each
bootstrap sample contains exactly the same number of individuals in each
MSA as the original sample. (12) The bootstrap procedure reflects the
nature of our sample design. We regard the set of MSAs as fixed but each
sample within an MSA as a random sample of that MSA's population.
Statistically, our set of MSAs approximates a population, but we have
samples within each MSA.
Our regression-adjusted estimates employ a model with regional
fixed effects that controls for disease severity and demographics. (13)
For each sector s we estimate
[Y.sub.irts] = [[alpha].sub.0s] + [[CHI].sub.its][[beta].sub.s] +
[[delta].sub.ts] + [[delta].sub.rs] + [[member of].sub.irts].
Here [Y.sub.irts] represents some measure of utilization or
spending by patient i in region r at time t and in sector s. The vector
[CHI] includes the following demographic characteristics for each
patient: age, age squared, sex, income, income squared, age and age
squared interacted with sex, as well as dummy variables for each of the
adverse health conditions listed above. The terms [[delta].sub.ts] and
[[delta].sub.rs] are sector-specific fixed effects for year and MSA,
respectively. The sector-specific variance in the fixed effect
[[delta].sub.rs] is the regression-adjusted analogue to the variance in
the MSA-level sample means.
As a general matter, the covariates have relatively little
predictive power within MSAs but a fair amount between MSAs. Across all
specifications, for instance, the MSA means of the covariates explain
about 50 to 70 percent of the between-MSA variation in utilization and
spending, in the sense of [R.sup.2].
II.D. Regional Variance in Utilization and Spending
Table 3 reports the estimated regional variance in four utilization
measures: number of hospitalizations, number of hospital days, number of
outpatient visits, and number of prescription drugs (in terms of
30-day-equivalents). Again, prescription drug coverage is provided by
the private sector in both populations throughout the sample period, and
therefore we do not expect to see similar differences for prescription
drugs as for the other measures.
The table shows overall between-MSA variation in the public and the
private sectors. The observed variation (first two columns) is computed
as the average MSA-level deviation from the overall mean. The top panel
reports the variation based on unconditional means; in the bottom panel,
both the overall mean and each MSA-level mean are regression-adjusted,
as described above. The corrected variation (second two columns) is
computed by subtracting the expected bias due to sampling error, as
described above. The next column shows the absolute difference between
the public and private variances, and the last column the ratio of the
variances. Asterisks indicate statistically significant differences from
zero for the differences, and from unity for the ratios.
Variation in hospital days is about three times, and variation in
outpatient visits about two times, higher in the public sector. These
differences are statistically significant at the l0 percent level or
higher and appear regardless of whether we adjust for covariates
(although the magnitudes differ somewhat). On the other hand,
prescription drug utilization exhibits statistically less variation in
the Medicare population; this is important because, again, even Medicare
patients obtain their prescription drug insurance privately in our
sample. Finally, there is no statistically significant difference in the
variation for hospitalizations. It is likely that more statistical power
is needed to pin down this variance, in one direction or the other.
Overall, these results provide evidence suggesting higher variance in
the public sector, but for a few of the outcomes our statistical tests
lack the power to generate definitive results.
Table 4 reports the estimated regional variance in four spending
measures: total spending, inpatient spending, outpatient spending, and
prescription drug spending. The regression-adjusted estimates indicate
that outpatient spending exhibits only about 35 percent as much
variation in the public sector as in the private sector. Inpatient
spending exhibits roughly equal variation in the two sectors. Finally,
prescription drug spending varies less for Medicare patients. With that
exception, these results are quite different from the utilization
results and suggest that price restraints play a role in the public
sector. In spite of greater variation in utilization, the public sector
exhibits less variation in spending.
III. Comparisons with Existing Literature
Regional variation in spending and utilization in the public sector
has been well documented in a literature that is almost 40 years old and
well accepted by the academic community. In that sense, our contribution
is to compare this with variation in the private sector, rather than to
establish the existence of public sector variation.
Table 5 summarizes a few representative papers from this vast
literature. (14) John Wennberg and Alan Gittelsohn (1973) provide an
early example. Their study analyzed variation across "hospital
service areas," a precursor to the HRRs typically analyzed in the
modern Dartmouth Atlas of Health Care studies. The table also lists a
couple of important studies that use states or MSAs. It is important to
recognize this difference when comparing our MSA-level analysis with
HRR-level analyses elsewhere, and it is important for future work to
assess the potential implications of this difference.
The 2008 Dartmouth Atlas of Health Care reports that average
spending on health care in the last 2 years of life (for deaths
occurring from 2001 to 2005) ranged from a high of $59,379 in New Jersey
to $32,523 in North Dakota (Wennberg and others 2008). This range, from
28 percent above to 30 percent below the national average, is similar to
the range of quantity utilization reported across MSAs by MedPac: from
39 percent above the national average in Miami to 25 percent below in
rural Hawaii (Medicare Payment Advisory Commission 2009).
These variations are not fully explained by factors such as age,
insurance coverage, average income, and rates of illness or disease.
David Cutler and Louise Sheiner (1999) investigate the extent to which
variation in spending across HRRs can be explained by regional
differences in illness, in the demand for health (for example, as
measured by income and race), or in "exogenous differences in the
structure of medical care markets" (for example, in the ratio of
generalists to specialists). They find that regional demographics can
explain about 70 percent of the variation in medical spending across
regions, but the unexplained variation remains large. For example, when
differences in demographics and the illness of the population are
accounted for, bringing Medicare spending down to the level of the
10th-percentile region would reduce total spending by 15 percent.
Perhaps the existing study most closely related to ours is that of
Michael Chernew and others (2010), who compare HRR-level variations in
Medicare against those in a sample of large firms in the Thomson Reuters
(Medstat) MarketScan Commercial Claims and Encounters Database. They
estimate that the geographic variation in private sector spending is
greater than that in Medicare spending (coefficient of variation of 0.21
versus 0.16). This is consistent with our findings for spending. They
focus less on variation in utilization, although they do report a
positive correlation between Medicare and non-Medicare inpatient days.
IV. Limitations of Our Analysis
There are several empirical questions that our data cannot address
but that should be addressed in future work. The populations of
privately insured and publicly insured patients differ, because the
latter have often opted out of private health insurance options. The
empirical implications of this are not clear a priori. Fee-for-service
Medicare patients are likely to be sicker than their counterparts in
private Medicare health maintenance organizations (HMOs), because HMOs
attempt to select healthier patients (Morgan and others 1997). On the
other hand, the privately insured nonelderly may also be healthier than
the nonelderly overall, if private health insurers select against the
sickest patients for similar reasons. The link between health insurance
and employment in the nonelderly population adds further complexity, as
those who are eligible for employment-based health insurance may be
richer or healthier, or both, than their peers. Finally, the fact that
our private sector data are based only on employees of large (Fortune
500) firms adds a further dimension of selection.
We ran several supplementary analyses to investigate some of these
issues, but our data lack the power to reach definitive conclusions
across the board. First, we narrowed the age range of our comparisons,
to mitigate some of the differences in health status. We compared 60- to
64-year-olds in the commercially insured population with 66- to
70-year-olds in the fee-for-service Medicare population. As this
restriction further reduces the sample, we limit our analysis to the 70
MSAs for which we have at least 25 observations in both samples.
Table 6 reports the result for the samples with the narrow age
ranges. Generally, the point estimates based on these restricted age
ranges are similar to those based on the full sample, but the precision
of the estimates declines enough to eliminate statistical significance.
The point estimates indicate that variation in the public sector is
about 5.1, 3.4, and 1.2 times that in the private sector for hospital
days, outpatient visits, and hospitalizations, respectively. As in the
analysis based on the full sample, variation in prescription drug use is
smaller in the public sector, about 53 percent as large as variation in
the private sector.
Next we investigated the issue of selection based on employment by
comparing our privately insured sample with Medicare patients who also
have coverage from an employer. If an individual has such coverage, we
know that he or she was employed and privately insured at one point.
Roughly 35 percent of Medicare enrollees in our sample also have
employer-provided coverage. They are slightly younger (averaging 77
years, compared with 79 years for those without such coverage), richer
(average income is 58 percent higher), and more likely to be male (52
percent versus 40 percent) than the average Medicare enrollee. Having
employer coverage is associated with very small differences in the
fraction of total expenses paid for by Medicare: Medicare pays 39
percent of the expenses of those without employer coverage and 38
percent of those with such coverage. The lack of a disparity is due to
the fact that once an elderly Medicare beneficiary retires, the
employer-provided coverage becomes secondary to Medicare. In our data
just 9 percent of individuals in the Medicare sample with employer
coverage are working, so for the vast majority Medicare is the primary
payer. It thus seems reasonable to assume that Medicare is the primary
driver of resource allocation for these individuals. A number of MSAs
are left with very small samples after this restriction, so we limit our
analysis to the 77 MSAs where we have at least 50 observations in both
samples.
These results are presented in table 7. Again, the point estimates
are similar to those based on the full sample, but the precision of the
estimates declines enough to eliminate much of the statistical
significance. The point estimates indicate that variation in the public
sector is about 4.1, 3.8, and 1.6 times that in the private sector for
hospital days, outpatient visits, and hospitalizations, respectively.
The greater variation in outpatient visits in the public sample is
statistically significant at the 1 percent level. The other differences
are not significant at the 10 percent level. Variation in prescription
drug use is slightly lower in the public sector, about 93 percent as
large as in the private sector.
V. Concluding Remarks
It has long been recognized that public and private enterprises
face different incentives to control costs. This paper has analyzed
these differences in the health insurance context, along with their
implications for variation in care. Public payers are likely able to
restrain prices better than private payers but have weaker incentives to
control costs through utilization controls. As a result, one might
expect greater variation in utilization for the public sector, but the
effects on total spending are ambiguous. Using samples of heart disease
patients, we presented empirical evidence consistent with these
implications.
Further research should focus more closely on the issue of whether
and to what extent variations across sectors are the result of
differences in the baseline health of the publicly and privately insured
populations. Work is also needed to assess whether our basic findings
can be generalized across other disease categories and geographical
classifications. In addition, the analysis of health outcomes must be
integrated into the analyses of utilization and of spending. As a
related point, although we have focused on estimated variations, further
research should be conducted into the sources of the mean differences in
utilization and spending. Finally, and perhaps most important, research
is needed to draw out the normative implications of variations in care
both within and between sectors.
The normative implications of variation in care are not
straightforward, in spite of the conventional wisdom that greater
variation implies inefficiency. On the one hand, the literature has
consistently found that areas exhibiting higher utilization of health
care services do not exhibit demonstrably better outcomes for patients
(Fisher and others 2003a). This has led many to conclude that these
areas are practicing "flat-of-the-curve" medicine, where the
marginal benefit approaches zero. However, Amitabh Chandra and Douglas
Staiger (20071 demonstrate that productivity spillovers and
specialization can explain regional variation in the utilization of
intensive procedures, without resorting to inefficiency. Most notably,
their model can reconcile the seemingly contradictory evidence that
intensive treatments such as most surgery are often highly effective at
the individual level, but that regions using these treatments more
intensively do not have better average health outcomes. Chandra and
Staiger observe that regions specializing in intensive treatment will
find it optimal to provide that treatment to more patients; therefore,
the marginal patient in such regions will be less suited to it than the
marginal patient elsewhere. This mitigates the greater benefits of
intensive treatment.
For this and other reasons, the efficiency implications of
variation in care require a more careful analysis. Even in our simple
framework, the normative impact of variation is unclear. For instance,
if the private sector is pricing and producing efficiently, then the
theory suggests that the public sector is engaging in inefficiently high
utilization and inefficiently low pricing. On the other hand, if private
sector prices are too high or if utilization is too low, then the
effects of public insurance may actually represent second-best
improvements to welfare. Evidently, it is important to investigate the
baseline efficiency properties of the private health insurance market
and to characterize how these are affected by the presence of publicly
financed health insurance.
Regardless of the conclusions, normative analysis of this issue
will likely generate a number of important policy implications. Many
have noted that Medicare has lower administrative costs than the private
sector. This is often interpreted as part of the value generated by
centralized insurance. This is a typical finding when one is comparing a
centralized with a decentralized model, but it could also be explained
by the cost of administering utilization controls in the private sector.
If so, any efficiency benefits of utilization controls would need to be
weighed against these administrative costs. The benefits generated by
administrative costs are often neglected in the policy debate, as are
related issues such as the deadweight costs of the tax revenue required
to fund public enterprise, the efficiency gains of marketing activities
by private firms, and higher rates of fraud in the public Medicare
system. The last of these is directly related to lax utilization
controls. A fuller analysis of the costs and benefits of public versus
private health insurance is needed.
The relative merits of public enterprise have a number of policy
implications. The first concerns the appropriate size of Medicare
Advantage, which operates through publicly provided premium subsidies to
private HMOs. Medicare Advantage plans are not directly comparable to
private payers, because they compete on quality rather than price, as
long as there is no price competition through competitive bidding for
plan members. Thus differences in incentives for utilization control
operate through the need to enhance quality, subject to available
premium resources, or result from residual claims on profits. Future
research needs to investigate more carefully the differences and
similarities in cost-control measures from this type of coverage and
their effects on regional variations and efficiency.
The second implication regards the timely issue of comparative
effectiveness research (CER), which has been offered as a means of
raising health care quality and reducing costs. The rationale for CER is
to generate better evidence, and to disseminate it to patients, payers,
and providers, about what works and does not work in health care.
Indeed, a common motivation for the use of CER is to reduce cost
inefficiencies due to regional differences in care. Awareness of CER has
been heightened recently by its significant public subsidization through
the American Recovery and Reinvestment Act of 2009. (15) An overriding
question raised by our analysis is whether regional variation in care
occurs because of a lack of information or a lack of incentives for
utilization control in the public sector.
Health economists have not yet paid sufficient attention to the
differences in incentives across the public and the private sectors or
to the corresponding implications for health care variation. The
regional variations documented in the Dartmouth Atlas of Health Care
have led several prominent researchers to conclude that high-use regions
ought to model themselves after their low-use peers (Fisher, Bynum, and
Skinner 2009). Our study suggests the importance of research focusing on
another, different question: whether or not public sector health
insurers ought to model themselves after their peers in the private
sector.
APPENDIX
Sampling Error in Estimation of Regional Variation In the text we
outlined our approach for obtaining a consistent estimate of regional
variation, defined as
RV [equivalent to] 1/R [[summation].sup.R.sub.r=1] [([[mu].sub.r] -
[??]).sup.2]
In this appendix we show how we solve for the bias in the sample
analogue, [??] [equivalent to] 1/R [[summation].sup.R.sub.r=1]
[([[mu].sub.r] - [??]).sup.2], and estimate it consistently using the
variance covariance matrix of the estimates. Recall the definitions from
the text: [[mu].sub.r] is the true population fixed-effect parameter for
region r, [[??].sub.r]. is the corresponding sample estimate, and
[z.sub.r] is a mean-zero sampling error with covariance across regions
E([z.sub.r] [z.sub.s]) = [[sigma].sub.rs]. The sample estimate is the
true value plus sampling error,
[??] = [[mu].sub.r] + [z.sub.r].
Define [??] [equivalent to] 1/R [[summation].sup.R.sub.r=1]
[([[mu].sub.r], the mean regional fixed effect across regions;
[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII], its sample
analogue; and [bar.z] [equivalent to] 1/R [[summation].sup.R.sub.r=1]
[z.sub.r], the average sampling error across regions.
Using the definitions above, we can write
[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII],
where we rely on the fact that we are dealing with regional fixed
effects, rather than random effects, to move the expectations operator
inside the summation. Expanding the right-hand side of the expression
results in
E([??]) = [equivalent to] 1/R
[[summation].sup.R.sub.r=1][E[([[mu].sub.r] - [??]).sup.2]].
Since [[mu].sub.r], and [??] are both scalars, this simplifies to
E([??]) = [equivalent to] 1/R
[[summation].sup.R.sub.r=1][([[mu].sub.r] - [??]).sup.2] -
2([[mu].sub.r] - [??])E ([z.sub.r] - [??]) + E[([z.sub.r] -
[??]).sup.2]].
The distributional assumptions on z imply that E([z.sub.r]) =
E([??]) = 0. Therefore, we can write
E([??]) = RV + 1/R [[summation].sup.R.sub.r=1]E[([z.sub.r] -
[??]).sup.2].
To characterize the bias, note that
E[([z.sub.r] - [??]).sup.2] = E([z.sup.2.sub.r]) - 2E([z.sub.r] 1/R
[R.summation over s=1] [z.sub.s]) + E[(1/R [R.summation over s=1]
[z.sub.s]).sup.2],
which we can write in terms of the variance and covariance
parameters as
E[([z.sub.r] - [??]).sup.2] = [[sigma].sub.rr] - 2/R [R.summation
over s=1] [[sigma.sub.rs] + 1/[R.sup.2] [R.summation over s=1]
[R.summation over t=1] [[sigma].sub.st].
The bias due to sampling variance is equal to the above expression,
averaged across all regions. A consistent estimate of the bias can be
calculated by summing up and taking the appropriate averages of
estimated variances of and covariances between the estimated regional
fixed effects. The more precisely the regional fixed effects are
estimated, the smaller is the bias correction.
ACKNOWLEDGMENTS We thank the editors, David Cutler, Mark McClellan,
and seminar participants at the University of Chicago, RAND, and the
Brookings Institution for comments. Financial support from the George
Stigler Center for the Economy and the State at the University of
Chicago, the Leonard Schaeffer Center for Health Economics and Policy at
the University of Southern California, and Pfizer Inc. is gratefully
acknowledged.
References
Alchian, Armen A., and Harold Demsetz. 1972. "Production,
Information Costs, and Economic Organization." American Economic
Review 62, no. 5: 777-95.
Baicker, Katherine, Elliott S. Fisher, and Amitabh Chandra. 2007.
"Malpractice Liability and the Practice of Medicine in the Medicare
Program." Health Affairs 26, no. 3: 841-52.
Chandra, Amitabh, and Douglas O. Staiger. 2007. "Productivity
Spillovers in Health Care: Evidence from the Treatment of Heart
Attacks." Journal of Political Economy 115, no. 1: 103--40.
Chernew, Michael E., and others. 2010. "Geographic Correlation
between LargeFirm Commercial Spending and Medicare Spending."
American Journal of Managed Care 16, no. 2: 131-38.
Christensen, Sandra, and Judith Wagner. 2000. "'The Costs
of a Medicare Prescription Drug Benefit." Health Affairs (Millwood)
19, no. 2: 212-18.
Cutler, David M., and Louise Sheiner. 1999. "The Geography of
Medicare." American Economic Review 89, no. 2: 228-233.
De Alessi, Louis. 1974a. "An Economic Analysis of Government
Ownership and Regulation: Theory and the Evidence from the Electric
Power Industry." Public Choice 19: 1-42.
--. 1974b. "Managerial Tenure under Private and Government
Ownership in the Electric Power Industry." Journal of Political
Economy 82, no. 3: 645-53.
Drummond, Michael, and others. 1993. "Economic Evaluation of
Pharmaceuticals: A European Perspective." Pharmacoeconomics 4, no.
3:173-86.
--. 1999. "Current Trends in the Use of Pharmacoeconomics and
Outcomes Research in Europe." Value in Health 2, no. 5: 323-32.
Eppig, Franklin J., and George S. Chulis. 1997. "Matching MCBS
(Medicare Current Beneficiary Survey) and Medicare Data: The Best of
Both Worlds." Health Care Financing Review 18, no. 3:211-29.
Fisher, Elliott S., Julie P. Bynum, and Jonthan S. Skinner. 2009.
"Slowing the Growth of Health Care Costs--Lessons from Regional
Variation." New England Journal of Medicine 360, no. 9: 849-52.
Fisher, E. S., and others. 2003a. "The Implications of
Regional Variations in Medicare Spending. Part 1: The Content, Quality,
and Accessibility of Care."
Annals of Internal Medicine 138, no. 4: 273-87. --. 2003b.
"The Implications of Regional Variations in Medicare Spending. Part
2: Health Outcomes and Satisfaction with Care." Annals of Internal
Medicine 138, no. 4: 288-99.
Fowler, Floyd J., and others. 2008. "Relationship between
Regional per Capita Medicare Expenditures and Patient Perceptions of
Quality of Care." JAMA 299, no. 20: 2406-12.
Gottlieb, Daniel J., and others. 2010. "Prices Don't
Drive Regional Medicare Spending Variations." Health Affairs 29,
no. 3: 537-43.
Kessler, Daniel, and Mark McClellan. 1996. "Do Doctors
Practice Defensive Medicine?" Quarterly Journal of Economics 111,
no. 2: 353-90.
--. 2002a. "Malpractice Law and Health Care Reform: Optimal
Liability Policy in an Era of Managed Care." Journal of Public
Economics 84, no. 2: 175-97.
--. 2002b. "How Liability Law Affects Medical
Productivity." Journal of Health Economics 21, no. 6:931-55.
Medicare Payment Advisory Commission. 2009. "Measuring
Regional Variation in Service Use." Washington.
Morgan, Robert O., and others. 1997. "The Medicare-HMO
Revolving Door-The Healthy Go In and the Sick Go Out." New England
Journal of Medicine 337, no. 3: 169-75.
Phelps, Charles E. 2000. "Information Diffusion and Best
Practice Adoption." In Hundbook of Health Economics. vol. 1A,
edited by Anthony J. Culyer and Joseph P. Newhouse. Amsterdam: Elsevier
Science.
Rettenmaier, Andrew J., and Thomas R. Saving. 2009.
"Perspectives on the Geographic Variation in Health Care
Spending." Special Publication. Washington: National Center for
Policy Analysis (June).
Sampford, M.R. 1953. "Some Inequalities on Mill's Ratio
and Related Functions." Annals of Mathematical Statistics 24, no.
1: 130-32.
Sirovich, Brenda E., and others. 2006. "Regional Variations in
Health Care Intensity and Physician Perceptions of Quality of
Care." Annals of Internal Medicine 144, no. 9: 641--49.
Spann, Robert M. 1977. "The Macroeconomics of Unbalanced
Growth and the Expanding Public Sector: Some Simple Tests of a Model of
Government Growth." Journal of Public Economics 8, no. 3: 397-404.
Sutherland, Jason M., Elliott S. Fisher, and Jonathan S. Skinner.
2009. "Getting Past Denial--The High Cost of Health Care in the
United States." New England Journal of Medicine 361, no.
13:1227-30.
Wennberg, John, and Alan Gittelsohn. 1973. "Small Area
Variations in Health Care Delivery." Science 182, no. 4117:1102-08.
Wennberg, John E., and Megan McAndrew Cooper, eds. 1998. Dartmouth
Atlas of Health Care, 1998. Chicago: Dartmouth Medical School and
American Hospital Association.
Wennberg, John E., and others. 2008. Tracking the Care of Patients
with Severe Chronic Illness: The Dartmouth Atlas of Health Care 2008.
Lebanon, N.H.: Dartmouth Institute for Health Policy and Clinical
Practice.
TOMAS J. PHILIPSON
University of Chicago
SETH A. SEABURY
RAND Corporation
LEE M. LOCKWOOD
University of Chicago
DANA P. GOLDMAN
University of Southern California
DARIUS N. LAKDAWALLA
university of Southern California
(1.) The main data source used to document regional variations is
the Dartmouth Atlas of Health Care, which can be found at
www.dartmouthatlas.org/(accessed January 15, 2010).
(2.) Alchian and Demsetz (1972) showed a greater incentive for
shirking and inefficiency in public enterprise, where managers' and
employees' own standards of living are unaffected by poor
performance. De Alessi ( 1974a, 1974b) observed that inefficient private
firms disappear, whereas inefficient public firms can last for long
periods. Spann (1977) argued that private firms typically produce
similar goods and services at much lower cost than their public
counterparts.
(3.) This analysis is general enough to include several possible
sources of regional differences, and in particular it allows for such
differences to be efficient. However, differences in liability (Kessler
and McClellan 1996, 2002a, 2002b, Baicker and Chandra 2007) or
productivity (Chandra and Staiger 2007), for example, may imply
differences in efficient levels of care.
(4.) Imperfect UR has qualitatively similar theoretical
implications. The difference is one of degree rather than nature.
5. We use the fact that the derivative of the inverse Mills ratio
with respect to [alpha] is strictly between zero and l,
[lambda]'([alpha]) [member of] (0,1). (Sampford 1953).
(6.) Our proxy for income is median household income at the 3-digit
zip code level; this is taken from the 2000 Census.
(7.) When estimating the cost of Medicare Part D (for example), the
Congressional Budget Office scaled the reported MCBS prescription drug
spending up by 33 percent for the noninstitutionalized population
(Christensen and Wagner 2000).
(8.) Also called myocardial ischemia, IHD is characterized as
reduced blood flow to the heart. In the private data we identify
patients with IHD as those with at least one inpatient or outpatient
claim with a primary diagnosis ICD-9 code of 410.xx, 411.xx, 412.xx,
413.xx, or 414.xx. In the public data we identify patients based on
self-reports of ever diagnosed with heart disease. See the online data
appendix (available on the Brookings Papers webpage at
www.brookings.edu/economics/bpea.aspx, under "Conferences and
Papers") for more information.
(9.) The specific conditions considered are essential hypertension,
congestive heart failure, diabetes, asthma, hypercholesterolemia, ulcer,
depression, chronic obstructive pulmonary disease, allergic rhinitis,
migraine, arthritis, chronic sinusitis, anxiety disorder, cardiac
disease, vascular disease, epilepsy, gastric acid disorder, glaucoma,
gout, hyperlipidemia, irritable bowel syndrome, malignancy, psychotic
illness, thyroid disorder, rheumatoid arthritis, tuberculosis, angina,
human immunodeficiency virus infection, anemia, and stroke. Most
co-morbidities are relatively uncommon, except for the ones involving
heart disease (or risk lactors for heart disease).
(10.) The MCBS also contains self-reports of a number of distinct
health conditions, as well as the individual's self-reported
general health status (coded 1 to 5, with 1 indicating poor and 5
indicating excellent). Out regression analysis relies on the
claims-based, rather than self-reported, disease measures for both the
public and the private samples. More details appear in our online data
appendix.
(11.) See the footnotes to tables 1 and 2 for a few sample size
issues specific to certain variables.
(12.) The alternative block-bootstrap that samples MSAs with
replacement generates nearly identical inferences for statistical
significance in our analysis, and so does a "'flat"
bootstrap.
(13.) A possible alternative is a random-effects model, but the
Hausman test rejected this more efficient model in favor of the
fixed-effects model in the majority of cases we analyzed.
(14.) For useful summaries from both the economic and the clinical
literatures on geographic variation in health care, see Wennberg and
Cooper (1998), Phelps (2000), Fisher and others (2003a, 2003b), Chandra
and Staiger (2007) and Sutherland. Fisher, and Skinner (2009).
(15.) The explicit use of comparative effectiveness assessments is
much more common outside the United States, particularly in the European
Union. However, this is a relatively recent trend: no European countries
formally required economic assessments for pricing and reimbursement
decisions as of 1993. but a majority had such a policy either in place
or in development by 1999 (Drummond and others 1993, 1999).
Comment and Discussion
COMMENT BY
DAVID M. CUTLER The issue of geographic variation in medical care
spending played an important role in the recent health care reform
debate. Supporters of reform cited the wide variation in spending across
U.S. regions as evidence that care is inefficiently provided. According
to a widely cited analysis by Elliott Fisher and others (2003), "If
the United States as a whole could safely achieve spending levels
comparable to those of the lowest-spending regions, annual savings of up
to 30% of Medicare expenditures could be achieved." One-third of
the nation's medical spending amounts to over $700 billion
annually--a huge savings.
Few dispute that health care is characterized by significant
inefficiency, but how best to limit that inefficiency is a subject of
great debate. Broadly speaking, there are two approaches. The first is
to use consumer demand to limit wasteful care. In this approach,
informed, incentivized consumers would shop for efficient care just as
they do for other products. Private insurers responding to this demand
would then squeeze out excessive spending. The main barriers to this
favorable competition, proponents argue, are the tax subsidy favoring
the purchase of more generous insurance by employers in the private
sector, and the lack of good stewardship of Medicare and Medicaid in the
public sector. In the latter category, the biggest issue is the absence
of a profit incentive to manage care for the bulk of Medicare and
Medicaid enrollees.
The second approach focuses on the supply side. Here the key issue
is how physicians and hospitals are paid for treating people. Doctors
are reimbursed more for providing more care, but not for providing
higher-quality care or limiting the need for care in the first place.
Thus surgeons are paid for performing operations, but primary care
physicians are not paid for the management involved in preventing
surgery. Care will exceed the optimum because quantity is favored over
quality.
These two theories have very different implications for health care
reform. In the demand-side view, competition among insurers is the
ideal, and private markets are preferred to public programs. Thus
Medicare would do better to encourage entry by private plans than to
focus on improving the fee-for-service system. In the supply-side view,
the key is changing the financial incentives for providers. This is best
accomplished by reforming the payment system in the traditional Medicare
program, and working to spread that change throughout the health care
system. Thus encouraging more private plans would not be a helpful step
for Medicare, and might be harmful.
This paper by Tomas Philipson and his colleagues is directed at
this debate. The idea behind the paper is to examine whether private
insurers do a better job at limiting health care utilization than
Medicare does. Evidence that private plans are better at rationing care
would support the demand-side view of reform. The test that the authors
propose is to look at the variability of care across a sample of
metropolitan statistical areas. If private plans are better at
eliminating excessive care, they should have lower spending in
metropolitan areas where care is otherwise overprovided--defined in the
paper as areas where Medicare fee-for-service costs are higher.
Analytically, this would show up as lower variance of health care
utilization across areas in the private sector.
Philipson and his coauthors test this hypothesis using data on
patients with ischemic heart disease, a condition that bridges the over-
and under-65 populations. They reach two conclusions. First, they show
that the variation in two measures of utilization--the number of
hospital days and outpatient visits--is indeed greater in Medicare than
in private insurance. For both measures, the standard deviation across
areas in Medicare is three times that across areas in private insurance.
Interestingly, the average number of hospital stays per patient is not
more variable in Medicare; rather, it is the average length of stay
(hospital days) that is more variable. Private insurers seem better at
getting people out of the hospital sooner.
In contrast, however, spending is much less variable across areas
than is utilization. Areas with low private sector utilization have
higher private sector prices than areas with higher utilization. This is
not true in Medicare, where prices are set nationally.
Whether one concludes from these findings that the public or the
private sector is more efficient depends on whether one puts more weight
on utilization or on spending. Since greater utilization has a real
resource cost, whereas higher prices paid for given resources are just a
transfer (if one excludes the deadweight loss of raising tax revenue to
support the public programs), one might imagine that lower utilization
at higher prices is preferable to higher utilization at lower prices.
This logic favors the private sector. On the other hand, the additional
care that people receive in Medicare is not necessarily wasteful. If the
added care in the public sector has high marginal value, more care
received at lower prices would be preferred to less care at higher
prices, and Medicare would be judged superior.
A priori, it is not clear whether the additional care provided in
high-utilization areas is worth the cost. If health care obeyed standard
market theory, the care that well-informed patients facing price
incentives decide not to purchase (either when the care is offered, or
earlier, when they choose a less generous insurance plan) would be the
care that is least valued. But patients are not always sure about what
care they need, and providers have imperfect incentives to fully inform
them. Thus one natural extension of the paper would be to compare health
outcomes in the public and private sector, to see whether patients get
care of real value. The data that Philipson and his coauthors employ do
not include ideal measures of outcomes, but some outcome assessment is
possible: for example, one could examine the rate at which patients in
the sample are later admitted to a hospital with cardiac complications.
There is also an obvious alternative explanation for the greater
geographic variance of utilization in Medicare than in private
insurance: health status may vary more in the Medicare population than
in the privately insured population. To control for this variation,
Philipson and coauthors perform a variety of risk adjustments. Having
already limited the sample to people with ischemic heart disease, they
further control for other conditions that those patients may have. But
these adjustments are imperfect, as any risk adjustment is.
The reason to suspect that greater variation in health status in
the Medicare population is important is that the reduction in variation
in the private sector occurs at both the top and the bottom end of the
utilization distribution. The theory that the authors put forward is
that private insurers should ration care more when underlying
utilization is higher. Thus private sector utilization should be
disproportionately lower in high-use areas. That is the point of figure
2 in the paper. In contrast, however, their figure 3 shows a uniform
reduction in the standard deviation of care in the private sector. That
is, high-utilization areas are less far above the mean in private
insurance, but low-utilization areas are less far below the mean. The
reason why care in low-utilization areas would be greater in the private
sector is not clear. One might imagine that providers are skimping on
the Medicare population in low-utilization areas, and that private
insurers encourage more care than providers would otherwise prescribe.
But the analysis by Fisher and his coauthors does not suggest
significant under-use of care in the low-spending Medicare areas, nor do
patients report themselves less satisfied in those areas. Given the
approximately proportional reduction in variation in the private sector
compared with Medicare, it seems that a theory of variation in
underlying health status garners at least as much support as a theory of
efficient rationing.
[FIGURE 1 OMITTED]
This conclusion is buttressed by a consideration of aggregate
differences between Medicare and private insurance. An analysis of
aggregate spending data does not suggest any superiority of private
insurers in limiting cost increases over time. My figure 1 shows annual
data on growth of real Medicare spending per beneficiary and private
insurance premiums per enrollee. In each case the spending measure is
limited to a common set of benefits, including hospital inpatient and
outpatient care and physician services. The two series generally mirror
each other. Cost growth was high in the 1980s, low in the 1990s, and
resurgent in the 2000s. Overall, cost increases were greater in the
private sector (averaging 4.9 percent per year) than in Medicare (3.7
percent).
A few anomalies in the pattern are of interest. Private insurance
premiums rose more rapidly than Medicare spending in the 1980s, when
prospective payment was introduced in the public sector. The managed
care era of the 1990s saw moderately slower cost growth in the private
sector than in Medicare. Managed care used the type of utilization
restrictions that Philipson and his coauthors argue are effective.
But a good share of these utilization controls were reversed in the
2000s, when a backlash against managed care caused insurers to change
their practices. Requirements for prior authorization were substantially
weakened, and financial incentives for providers to perform less care
were loosened. Further, providers responded to the presence of managed
care by merging, and this led to higher prices in the private sector,
where prices are flexible. The result was a decade of much higher
private sector cost increases than Medicare cost increases.
The data that Philipson and coauthors analyze are from the 2000s.
Figure 1 shows the peculiarity of the case that they make. They argue
for the superiority of private utilization controls using data from a
time when private sector controls on utilization were falling and costs
were rising rapidly. It may be that people were wrong to reject the
managed care strictures of the 1990s, and that a return to those
practices would be beneficial. But few people argue that the version of
managed care that prevailed in the 2000s is a sustainable model.
In the recent health care reform debate, the conclusion that public
and private insurance were relatively similar in cost growth was more
commonly accepted than the argument that private insurers were superior
in reducing excess utilization. Added to this was the sense that private
insurers engaged in significant and costly risk selection, wasting money
and making insurance difficult to obtain. Thus, regulating private
insurers to limit risk selection is a significant element of the
recently enacted Patient Protection and Affordable Care Act (PPACA),
whereas opening up new markets in Medicare for private insurers is not.
Indeed, the act reduces Medicare payments to private insurers. In its
place are a series of pilot programs and demonstration projects aimed at
reforming the incentives in the existing fee-for-service program.
Whether the strategy underlying the PPACA is the right one or not
will be determined in the next few years. If reform fails, it may be
because private insurers were strangled too tightly. If so, this paper
may point the way to a future reform. But for now we are likely to see
tighter restrictions on private insurance before any significant new
utilization controls.
REFERENCE FOR THE CUTLER COMMENT
Fisher, Elliott S., David E. Wennberg, Therese A. Stukel, Daniel J.
Gottlieb, F. L. Lucas, and Etoile L. Pinder. 2003. "The
Implications of Regional Variations in Medicare Spending. Part 2: Health
Outcomes and Satisfaction with Care." Annals of Internal Medicine
138: 288-98.
GENERAL DISCUSSION Christopher Sims raised the question of how
secondary private insurance fits into the classification of public
versus private. Many Medicare beneficiaries have supplementary coverage,
and whether an individual has it or not could affect his or her
utilization choices. If there is regional variation in the share or
beneficiaries using Medicare supplements, it could be affecting the
paper's results.
Daniel Sacks asked whether the privately covered individuals in the
sample were truly representative of people with private coverage
generally. If the sample is drawn from Fortune 500 companies only, their
health insurance plans may be better than the average private plan and
have better, or at least different, cost management. He also wondered
what incentives the executives of both public and private insurance
plans actually face and whether they are as different as the paper
assumes.
David Romer observed that the theory laid out in the paper makes a
clear prediction of where the expected change in the distribution of
fixed effects ought to appear: The incentives facing the private sector
should lead to compression of the upper tail of the distribution in that
sample but not the lower tail, yet no such pattern is visible in the
data.
Table 1. Selected Patient Demographic and Health Characteristics (a)
Private
Standard
Patient characteristic Mean deviation
Age 55.4 7.1
Percent male 65 48
Income (thousands of 2004 dollars) 42.8 10.8
Percent with heart disease in year 32 47
No. of adverse health conditions 1.4 1.4
Public
Standard
Patient characteristic Mean deviation
Age 78.2 7.9
Percent male 46 50
Income (thousands of 2004 dollars) 28.5 46.8
Percent with heart disease in year 37 48
No. of adverse health conditions 2.9 2.4
Sources: Data on private patients come from a modified version
of the Ingenix Touchstone product. Data on public patients come
from the MCBS.
(a.) History of heart disease is self-reported in the public
sample and identified using medical claims in the private
sample. The private sample has 240,028 observations and the
public sample 24,800 observations.
Table 2. Distributions of Spending and Utilization Measures (a)
Standard 25th
Measure Sample Mean deviation percentile
Utilization (number per patient per year)
Hospitalizations Private 0.36 1.15 0
Public 0.57 1.14 0
Hospital days Private 1.23 7.13 0
Public 2.93 8.59 0
Outpatient visits Private -5.56 5.86 1
Public" 8.59 11.05 1
Drug prescriptions 'c) Private 45.78 42.20 13
Public 35.45 29.93 14
Spending (thousands of 2004 dollars)
Total spending Private 8.40 22.98 0.56
Public 10.25 18.8 1.25
Inpatient spending Private 4.02 18.36 0
Public 4.94 13.21 0
Outpatient spending Private 4.38 9.83 0.54
Public 5.30 9.14 1.13
Prescription drug Private 2.80 5.78 0.53
spending 'c) Public 1.92 2.05 0.58
75th
Measure Median percentile
Utilization (number per patient per year)
Hospitalizations 0 0
0 1
Hospital days 0 0
0 1
Outpatient visits 4 8
5 12
Drug prescriptions 'c) 36 66
29 50
Spending (thousands of 2004 dollar
Total spending 2.10 6.88
3.91 11.4
Inpatient spending 0 0
0 4.65
Outpatient spending 1.85 4.86
2.94 6.44
Prescription drug 1.67 3.42
spending 'c) 1.39 2.63
Source: Authors' calculations.
a. Figures are yearly averages during 2000-06 (2000-05 for drug
prescriptions) for patients with a history of heart disease,
which is self-reported in the public sample and identified using
medical claims in the private sample. Except where noted
otherwise, the private sample has 240.028 observations and the
public sample 24.800 observations.
b. Survey responses (used to cross-validate the claims data)
were incomplete in 3,769 cases, so that the public sample has
21,031 observations.
(c.) Because observations from 2006 are omitted, the private sample
has 231,802 observations and the public sample 21,140 observations.
Number of prescriptions is in terms of 30-day-equivalents.
Table 3. Regional Variation in Mean Utilization (a)
Observed Corrected
variation (b) variation (c)
Utilization measure Private Public Private Public
Unconditional means
Hospitalizations 0.013 0.015 0.012 0.009
Hospital days 0.322 1.016 0.230 0.659
Outpatient visits 1.736 5.154 1.676 4.585
Drug prescriptions (d) 72.746 32.758 70.090 28.403
Regression-adjusted means (e)
Hospitalizations 0.006 0.011 0.005 0.006
Hospital days 0.169 0.610 0.080 0.313
Outpatient visits 0.988 3.255 0.942 2.677
Drug prescriptions (d) 29.190 25.856 27.086 21.131
Difference,
public to Ratio of
minus public to
Utilization measure private private
Unconditional means
Hospitalizations -0.003 0.728
(0.003) (0.204)
Hospital days 0.429 ** 2.870 *
(0.199) (1.017)
Outpatient visits 2.909 *** 2.735 ***
(0.502) (0.323)
Drug prescriptions (d) -41.687 *** 0.405 ***
(3.896) (0.043)
Regression-adjusted means (e)
Hospitalizations 0.001 1.266
(0.002) (0.430)
Hospital days 0.233 * 3.907 *
(0.124) (1.684)
Outpatient visits 1.735 *** 2.841 ***
(0.322) (0.379)
Drug prescriptions (d) -5.955 * 0.780 **
(3.130) (0.106)
Source: Authors' calculations.
(a.) Numbers in parentheses are standard errors on the difference
between public and private variation or the ratio of public to private
variation and are bootstrapped within MSAs, separately for public and
private patients, with 200 bootstrap draws. For both sectors, then,
the number of patients in each region in each bootstrapped sample is
the same as the number of patients in the original sample. Asterisks
indicate differences statistically significantly different from zero
or ratios statistically significantly different from 1 at the *** 1
percent. ** 5 percent, and * 10 percent level.
(b.) Variance in the regional means or fixed effects of the
utilization variables.
(c.) Unbiased measure of the true variance in the regional means or
fixed effects corrected for sampling error, as described in the text
and the appendix. All differences and ratios are based on these
numbers.
(d.) 30-day-equivalents.
(e.) Estimates of regional fixed effects on each utilization variable
from a regression that includes as other independent variables year
fixed effects, quadratic specifications of patient age and income,
patient sex, sex interacted with age, and dummy variables for 30
separate types of disease.
Table 4. Regional Variation in Mean Spending (a)
Observed Corrected
variation (b) variation (c)
Spending measure Private Public Private Public
Unconditional means
Total medical spending 4.443 3.907 3.571 2.352
Inpatient spending 1.634 1.587 1.109 0.842
Outpatient spending 1.263 1.048 1.082 0.634
Prescription drug spending 0.418 0.144 0.377 0.124
Regression-adjusted means (d)
Total medical spending
2.890 2.782 2.111 1.463
Inpatient spending
1.186 1.357 0.698 0.695
Outpatient spending
0.924 0.628 0.758 0.265
Prescription drug spending
0.251 0.086 0.214 0.064
Difference,
public Ratio
minus public to
Spending measure private private
Unconditional means
Total medical spending
-1.219 0.659 *
(0.981) (0.207)
Inpatient spending
-0.266 0.760
(0.416) (0.247)
Outpatient spending
-0.447 0.586
(0.355) (0.266)
Prescription drug spending
-0.253 *** 0.329 ***
(0.049) (0.060)
Regression-adjusted means (d)
Total medical spending
-0.647 0.693
Inpatient spending (0.728) (0.258)
-0.004 0.995
Outpatient spending (0.324) (0.298)
-0.494 * 0.349 **
Prescription drug spending (0.261) (0.272)
-0.150 *** 0.300 ***
(0.043) (0.082)
Source: Authors' calculations.
(a.) Spending is measured in 2004 dollars. Numbers in parentheses arc
standard errors on the difference between public and private variation
or the ratio of public to private variation and are bootstrapped
within MSAs, separately for public and private patients, with 2W
bootstrap draws. For both sectors, then, the number of patients in each
region in each bootstrapped sample is the same as the number of
patients in the original sample. Asterisks indicate differences
statistically significantly different from zero or ratios
statistically significantly different from 1 at the *** 1 percent,
** 5 percent, and * 10 percent level.
(b.) Variance in the regional means or fixed effects of the spending
variables.
(c.) Unbiased measure of the true variance in the regional means or
fixed effects corrected for sampling error, as detailed in the text
and the appendix. All differences and ratios are based on these
numbers.
(d.) Estimates of regional fixed effects on each spending variable
from a regression that includes as other independent variables year
fixed effects, quadratic specifications of patient age and income,
patient sex, sex interacted with age, and dummy variables for 30
separate types of disease.
Table 5. Key Findings on Variation in Regional Health Care Spending
Using Medicare Data (a)
Geographic
Study aggregation Summary
Wennberg and Gittelsohn (1973) Hospital Studied geographic
service variation in
area utilization and
spending in Vermont.
Cutler and Sheiner (1999) HRR Calculated share of
regional variation
in Medicare spending
attributable to
regional differences
in health and
demographics.
Fisher and others (2003x) HRR Compared patients
across regions
holding other
characteristics
constant.
Sirovich and others (2006) MSA Compared variation
in intensity of
treatment with
physician
perceptions of
quality of care.
Chandra and Staiger (2007) HRR Specified a model of
patient treatment
choice with
productivity
spillovers and
tested the model
using treatment
choices and health
outcomes of heart
attack patients.
Fowler and others (2008) HRR Used as a patient
survey to compare
local variation in
spending and
utilization with
patient perceptions
of quality.
Wennberg and others (2008) HRR Summarized
Darthmouth Atlas
findings on
geographic
variations in
Medicare spending
and their
implications.
Rettenmaier and Saving (2009) State Studied how state
rankings in medical
spending per capita
change when
different
definitions of
spending are used,
such as Medicare
only or total
spending by all
payers.
Sutherland, Fisher, and Skinner HRR Updated Dartmouth
(2009) Atlas findings on
geographic variation
in Medicare spending
and their
implications.
Chernew and others (2010) HRR Compared HRR-level
variation in medical
spending between
Medicare and large
firms.
Gottlieb and others (2010) HRR Examined role of
Medicare prices in
driving geographic
variations in health
care.
Study Key findings
Wennberg and Gittelsohn (1973) Found wide
variations
apparently due to
differences in
practice style
rather than in
population health.
Hospital days in
highest-use area
were 1.5 times that
in lowest.
Cutler and Sheiner (1999) Regional differences
in demographics can
explain about 70
percent of regional
differences in
Medicare spending.
but significant
differences remain
unexplained.
Fisher and others (2003x) Patients in higher-
spending regions
received
approximately 60
percent more care.
The increased
utilization mostly
arose from more
frequent physician
visits.
Sirovich and others (2006) Medicare spending
per capita in
highest intensity
quintile was 1.58
times that in
lowest.
Chandra and Staiger (2007) Patterns of which
patients benefit and
which lose from
intensive medical
care are consistent
with productivity
spillover model.
Fowler and others (2008) Regional differences
in spending were not
systematically
related to
differences in
patient perceptions
of care quality.
Wennberg and others (2008) Three states spent
more than 20 percent
above the national
average of $46,412.
Conversely, three
states spent 25
percent below the
national average or
less. Inter quartile
ratio (75th
percentile over
25th) is 1.26 for
HRRs.
Rettenmaier and Saving (2009) Found a state-level
correlation between
Medicare spending
and total spending
of 0.21, and that
variation in
Medicare spending
exceeds variation in
private spending.
Sutherland, Fisher, and Skinner Inpatient days per
(2009) beneficiary in
highest cost
quintile were 1.50
times that in
lowest; physician
visits in highest
cost quintile were
1.36 times that in
lowest.
Chernew and others (2010) Found substantial
regional variation
in spending, greater
for large firms than
for Medicare
(coefficient of
variation 0.21 v.
0.16). Correlation
between private and
public inpatient
utilization was
0.59.
Gottlieb and others (2010) Prices explain a
small portion of
variation in
spending. The 80th
percentile of price-
adjusted Medicare
Part B spending was
1.37 times the 20th
percentile.
Sources: Literature cited.
(a.) HRR = hospital referral region; MSA = metropolitan statistical
area.
Table 6. Regional Variation in Regression-Adjusted Mean
Utilization, Patients Aged 60 to 70 (a)
Observed Corrected
variation (b) variation (c)
Utilization
measure Private Public Private Public
Hospitalizations 0.015 0.035 0.011 0.014
Hospital days 0.258 2.038 0.130 0.656
Outpatient visits 1.460 7.207 1.335 4.524
Prescriptions 52.869 47.198 46.741 24.765
Difference, Ratio of
Utilization public minus public to
measure private private
Hospitalizations 0.003 1.244
(0.010) (0.921)
Hospital days 0.526 5.060
(0.886) (4.134)
Outpatient visits 3.190 3.390
(2.316) (1.695)
Prescriptions -21.975 ** 0.530
(11.090) (0.209)
Source: Authors' calculations.
(a.) The private sample is restricted to patients aged 60 to 64 and
the public sample to patients aged 66 to 70. Both samples are
restricted to include only the 70 MSAs with at least 25 observations
in both samples. The private sample has 67,414 observations and the
public sample 3,568 observations. Numbers in parentheses are
standard errors on the difference between public and private
variation or the ratio of public to private variation and are
bootstrapped within MSAs. and separately for public and private
patients, with 200 bootstrap draws. For both sectors, then, the
number of patients in each region in each boot strapped sample is
the same as the number of patients in the original sample. Asterisks
indicate differences statistically significantly different from
zero or ratios statistically significantly different from 1 at the
*** I percent, **5 percent, and * 10 percent level.
(b.) Variance in the regional means or fixed effects of the
utilization variables.
(c.) Unbiased measure of the true variance in the fixed effects
corrected for sampling error, as detailed in the text and the
appendix. All differences and ratios are based on these numbers.
Table 7. Regional Variation in Regression-Adjusted Mean Utilization,
Patients with Some Private, Employer-Provided Coverage (a)
Observed Corrected
variation (b) variation (c)
Utilization
measure Private Public Private Public
Hospitalizations 0.006 0.017 0.005 0.008
Hospital days 0.144 0.709 0.049 0.198
Outpatient visits 3.774 4.462 0.911 3.438
Prescriptions 28.982 34.778 26.885 25.107
Difference,
public Ratio
Utilization minus public to
measure private Private
Hospitalizations 0.003 1.572
(0.004) (0.984)
Hospital days 0.149 4.058
(0.240) (3.861)
Outpatient visits 2.527 *** 3.774 ***
(0.518) (0.614)
Prescriptions -1.778 0.934
(6.032) (0.213)
Source: Authors' calculations.
(a.) The public sample is restricted to patients who report at least
some form of private, employer provided insurance coverage. Both
samples are restricted to include only the 77 MSAs that have at
least 50 observations in both samples. The private sample has
202,202 observations and the public sample 8,416 observations.
Numbers in parentheses are standard errors on the difference between
public and pri vate variation or the ratio of public to private
variation and are bootstrapped within MSAs, and separately for
public and private patients. For both sectors, then, the number of
patients in each region in each bootstrapped sample is the same as
the number of patients in the original sample. Asterisks indicate
differences statistically significantly different from zero or
ratios statistically significantly different from 1 at the ***I
percent, **5 percent, and * 10 percent level.
(b.) Variance in the regional means or fixed effects of the
utilization variables.
(c.) Unbiased measure of the true variance in the fixed effects
corrected for sampling error. as detailed in the text and the
appendix. All differences and ratios are based on these numbers.
