# Physicians' cost shifting behavior: Medicaid versus other patients.

I. INTRODUCTION

Much of the debate about rising health care costs in the U.S. centers on the notion of "cost shifting." Cost shifting is loosely defined as charging one set of patients a higher price to offset losses on another set of patients. For example, to offset losses from treating charity patients, a hospital might charge a relatively high price to privately insured patients, or a physician might increase his standard fee to compensate for lower Medicare reimbursements.

Despite the certainty of the conventional wisdom, empirical evidence of cost shifting is, at best, mixed. In a study of Illinois hospitals in the early 1980s, Dranove (1988) finds evidence of cost shifting from government patients, primarily Medicaid and Medicare recipients, to private patients. Morrisey and Sloan (1989) find evidence of cost shifting in urban hospitals but not in rural hospitals. Several other empirical studies find no evidence of cost shifting.

One aspect of the cost shifting debate that the empirical work has ignored is whether or not doctors - as opposed to hospitals - practice cost shifting. Although standard profit-maximizing models of physicians would rule out cost shifting, the widely used "income targeting" model of physicians (Feldstein, 1993) does allow for such behavior. The probable reason for the dearth of empirical work is the lack of good data on physicians' pricing behavior. The analysis here uses the Physicians' Practice Costs and Income Survey, 1983-1985 (PPCIS, expanded version) to investigate this important question.

The PPCIS includes a set of questions asking physicians their standard fees and Medicaid reimbursement rates for a variety of medical and surgical procedures. The survey also contains detailed information on the physician and medical conditions in the local geographic area.

The basic research methodology is straightforward. Since Medicaid reimbursements are largely controlled by state legislatures, considerable variation exists in reimbursement rates across the United States. If cost shifting is a widespread phenomenon, one would expect that relatively low reimbursement rates would result in higher fees to privately insured patients, controlling for other factors. If cost shifting does not occur, one would expect Medicaid reimbursement rates to have no effect or possibly result in lower fees depending on the market structure facing the physician. The analysis also examines additional evidence that suggests that lower Medicaid reimbursements lead physicians to treat fewer Medicaid patients.

The results suggest that physicians do not practice "cost shifting" but instead act in a way consistent with the hypothesis of profit-maximization for a firm with some market power. That is, lower Medicaid reimbursements tend to lead to lower prices to privately-insured patients. This appears to be true across a wide spectrum of procedures performed by different specialties. Also consistent with the profit-maximization hypothesis, lower Medicaid reimbursements tend to cause physicians to treat relatively fewer Medicaid patients.

II. ALTERNATIVE MODELS OF PHYSICIAN BEHAVIOR

Health economists widely recognize that profit maximization precludes cost shifting. However, since most of the cost shifting literature focuses on non-profit hospitals that may be more likely to engage in cost shifting, reviewing the implications for physicians seems useful.

Profit maximizing models of physician behavior preclude cost shifting because under a maximization hypothesis, physicians extract as much as possible from all their patients under all circumstances. To say that physicians can shift costs from one set of customers to another, or more specifically, get more revenues from one set of customers to offset a decrease in revenues from another, is to say that the physicians were not profit maximizing in the first place.

To fix this principle, consider a stylized model of physician behavior. A profit-maximizing physician sells a homogeneous good, Q, "physician services" (the treatment of one patient). Assume that the physician has two types of patients. Type 1 patients ([Q.sub.1]) pay through a government program at a fixed rate, [P.sub.1]. The physician can treat as many or as few of these patients at that rate as he wishes. Type 2 patients ([Q.sub.2]) have a downward-sloping demand for the physician's services given by [P.sub.2] ([Q.sub.2]). Obviously, if a physician took [P.sub.2] as fixed as implied by a perfectly competitive model, then there would be no opportunity for cost shifting, so some kind of market power (implied by a downward sloping demand curve) is a necessary condition for cost shifting behavior. Also note that the price charged to the Type 2 patients (the "private" patients) in general will differ from the price that the physician receives for the Type 1 patients (the "public" patients). That the prices are different does not in and of itself imply either the absence or the presence of cost shifting behavior.

The cost of providing the good is a function of the total number of patients the doctor serves, C([Q.sub.1] + [Q.sub.2]). Costs are increasing and strictly convex. To ensure an interior solution, marginal revenue from Type 2 patients must be declining in [Q.sub.2] at the optimum.

The physician chooses the number and type of patients to serve. The profit function is then

(1) [Pi] = [P.sub.1][Q.sub.1] + [P.sub.2]([Q.sub.2])[Q.sub.2] - C([Q.sub.1] + [Q.sub.2])

with C[prime] [greater than] 0, C[double prime] [greater than] 0, [P.sub.2][prime] [less than] 0, and [P.sub.2][double prime] [Q.sub.2] + 2[P.sub.2][prime] [less than] 0 (declining marginal revenue condition) where [prime] and [double prime] denote the first and second derivative, respectively.

The optimization problem is to choose [Q.sub.1] and [Q.sub.2] to maximize profits. The first order conditions of this problem are

(2a) [Delta][Pi]/[Delta][Q.sub.1] = [P.sub.1] - C[prime] = 0

(2b) [Delta][Pi]/[Delta][Q.sub.2] = [P.sub.2][prime][Q.sub.2] + [P.sub.2] - C[prime] = 0

Equation (2a) implies that marginal cost will be set equal to the reimbursement rate, [P.sub.1]. Equation (2b) implies that the marginal revenue from Type 2 patients in turn also will be set equal to the reimbursement rate. This is a standard result - i.e., the marginal revenues across different markets will be equalized at the optimum. But these results rule out the shifting of costs to Type 2 patients. Cost shifting in this model is defined as the charging of a higher price to Type 2 patients when the reimbursement rate to Type 1 patients is lowered. But working through the implications of the first order conditions yields exactly the opposite result.

Suppose [P.sub.1] falls. Because marginal revenues are equated at the margin, the marginal revenue of Type 2 patients also should fall. But the assumption for profit maximization implies that marginal revenue is declining in [Q.sub.2] so a decrease in marginal revenue implies an increase in [Q.sub.2] itself. But since demand is downward sloping, an increase in [Q.sub.2] would imply a decrease in [P.sub.2]. So, as [P.sub.1] falls, the price charged to Type 2 patients also would fall; hence, the hypothesis of cost shifting is rejected.

More directly, (2a) and (2b) imply that

(3) [P.sub.2][prime][Q.sub.2] + [P.sub.2] = [P.sub.1]

Taking a total derivative of this equation and recognizing that d[P.sub.2] = [P.sub.2][prime]d[Q.sub.2], one gets

(4) d[P.sub.2]/d[P.sub.1] = [P.sub.2][prime]/[P.sub.2][double prime][Q.sub.2] + 2[P.sub.2][prime]

Since both the numerator and the denominator are negative, this quantity is positive, implying that as [P.sub.2] increases, so does [P.sub.1]; [P.sub.2] and [P.sub.1] move in the same direction in contradiction to the cost shifting story.

An additional result is that as [P.sub.1] falls, [Q.sub.1] also falls, meaning that as the reimbursement rate falls, the number of government-financed patients also decreases. This result also hinges on the second order conditions of the maximization problem, but the intuition is straightforward: as the marginal revenue from the government-financed patients falls, the physician will choose to serve fewer of them.

When the profit-maximization hypothesis is dropped, so is the key theoretical constraint against cost shifting. Dranove (1988) presents one example of an alternative model to strict profit maximization. He posits a model where a firm has a utility function that takes as arguments profits and output. The motivation is that for non-profit organizations, other factors besides profits are likely to influence production decisions. An important result is that for cost shifting to occur in this framework, the firm must exhibit some market power.

McGuire and Pauly (1991) present another optimization model based on demand-inducement behavior of physicians. In this model, physicians can increase revenues from a set of patients but only by incurring disutility for engaging in demand inducement. Such a model allows for cost shifting (gaining more revenues from one group as revenues from a second group exogenously declines) if income effects are sufficiently strong.

Empirical tests for cost shifting focus primarily on hospitals where the profit-maximization hypothesis probably is not the best model. Hadley and Zuckerman (1990); Morrisey, Sloan and Valvona (1988); Zuckerman and Holahan (1988); and Morrisey and Sloan (1989) find little evidence of pervasive cost shifting. (Morrisey, 1994, and Coulam and Gaumer, 1991, review the existing empirical evidence on cost shifting.) Although physician services are an important component of the health service industry, they have received little empirical attention in the cost shifting debate due to the paucity of micro-level data on physician pricing behavior. This study begins to fill this gap by using PPCIS data.

III. DATA DESCRIPTION

The primary data for this study are the 1983-1985 PPCIS conducted by the National Opinion Research Center under contract to the Health Care Financing Administration. This cross-sectional survey was conducted from October 1984 through June 1985 and includes responses from 4,729 physicians (out of 6,847 eligible) drawn from a stratified random sample of physicians from the American Medical Association's 1984 Physician Master File. The physicians were asked numerous detailed questions regarding pricing policies, reimbursement rates, and practice characteristics. The data also contain variables concerning the physicians' personal characteristics (age, sex, specialty, etc.) obtained from the AMA Physician Masterfile. The expanded version of the 1983 PPCIS contains county-level demographic information (age composition, ethnic composition, MDs per capita, etc.).

The strength of the data is their uniqueness. Depending on their specialty, self-employed physicians were asked for standard fees and reimbursement rates (Medicaid and Medicare) for three to five different medical procedures. For example, family practice physicians were asked for several different prices on an intermediate office visit, an intermediate follow-up hospital visit, a vaginal delivery, and an EKG report. Most specialists were asked questions about the intermediate office visit and the hospital follow-up visit.

Having both a Medicaid price that will vary by state of residence and a "usual fee" for a given procedure allows one to test for cost shifting straightforwardly by checking for a statistically significant negative correlation between Medicaid reimbursements and usual fees, controlling for other observable characteristics. The alternative hypothesis is that physicians are profit-maximizers, which would imply by the simple model in the previous section that a positive correlation should exist between Medicaid reimbursements and fees charged to non-Medicaid patients.

Such a test relies on the exogeneity of Medicaid reimbursements at the state level. This assumption seems plausible given the wide latitude individual state legislatures have in setting Medicaid reimbursement rates for physicians. Indeed, the freedom states have in setting the parameters of the Medicaid system and the resulting variation across states in how Medicaid operates has been the subject of some controversy (e.g., Cromwell et al., 1995). However, but for the purposes of this study, such arbitrary variation obviously is beneficial. The analysis here also checks the exogeneity assumption using instrumental variables estimation.

As with most data, the PPCIS also has some weaknesses. First, it is cross-sectional, meaning that unobservable physician characteristics cannot be controlled for as they would be, for example, with a panel data set. Second, the price information is self-reported by the physician or a designated representative. Such self-reporting possibly could be biased although the direction of the bias is unclear and a priori seems unlikely to favor one hypothesis over another (cost shifting versus profit maximization).

Another limitation of the data is that only a small subset of procedures is sampled for each physician instead of the full range of physician activities. This precludes the estimation of a fully specified cost or production function for physician services.

Finally, although the PPCIS is unique in its composition, it was conducted early in the current phase of changes in the health care industry. The logical presumption is that the market for physician services has become more competitive since this survey was taken. This would imply that physicians' ability to "cost shift" has lessened, implying that any findings of cost shifting using these data represent an upper bound estimate for cost shifting in the current environment.

Table 1 gives summary statistics for the data. Prices have been adjusted to account for differences in costs-of-living across states using the American Chamber of Commerce Association's cross-sectional price index. The price index is based on a population-weighted index of city-level price indexes reported in 1985 (most information is from the 1st quarter report, but a few states have no reported cities for the 1 st quarter, so the analysis uses the next closest quarter with a reporting city, either the 3rd quarter or the 4th of 1985). The weights are from the 1980 census. Several variations on the price index do not materially change the results. (Details on this and other aspects of the paper are available from the author.)

The table gives information on 1,915 physicians who report both a standard fee and a Medicaid reimbursement rate. The information on fees and reimbursements includes all procedures reported by a given physician.

IV. EMPIRICAL RESULTS

Intuitively, if cost shifting is a widespread phenomena operating through the fees that physicians charge and if the mechanism is that low Medicaid reimbursements result in higher fees to non-Medicaid patients, one might be able to see differences at the aggregate state level since Medicaid reimbursements vary widely by state.

A simple test is to compute the mean Medicaid reimbursement and mean standard fee for the most frequently answered procedure (intermediate office visit for an established patient) by state and then compute the correlation across states. (Hawaii, Alaska, and Arizona have been excluded. Cost-of-living information was not available for Hawaii. Alaska has too few respondents to be of value. Arizona obtained a waiver from participating in the Medicaid program.) The Pearson Correlation is 0.37, which is consistent with profit-maximization and contradicts cost shifting.

[TABULAR DATA FOR TABLE 1 OMITTED]

Next, the analysis runs a weighted least squares regression of the mean standard fee on a constant and the state mean Medicaid reimbursement for an office visit where the weight is the number of surveyed physicians with a value for their office fee in each state. The coefficient on the Medicaid is 0.43 with a 2.4 t-statistics. The interpretation of this coefficient is that not adjusting for any other factors, a $1 increase in the Medicaid reimbursement rate will raise the average standard fee by 43 cents. The [R.sup.2] of the regression is relatively low at 0.11, but this simple evidence lends no support to the notion of cost shifting.

A. Regression Analysis

Such simple correlations suggest a positive relationship. However, after controlling for other demographic characteristics, the gross positive correlation might disappear. Therefore, a more detailed regression model is necessary:

(5) [fee.sub.ij] = [[Beta].sub.0] + [[Beta].sub.1][medaid.sub.ij] + [x.sub.ij][Beta] + [[Epsilon].sub.ij]

where i denotes the physician, j denotes the state, fee is the standard fee for a given medical procedure, medaid is the Medicaid reimbursement for that procedure, and [x.sub.ij] is vector [TABULAR DATA FOR TABLE 2A OMITTED] of other explanatory variables including the characteristics of the doctor and local demographic (county level) variables. Under the simplest interpretation of cost shifting, a decrease in Medicaid payments should result in an increase in the standard fee, which would imply a negative value for [[Beta].sub.1]. The initial assumption is that the Medicaid rate is set exogenously by state legislatures. Later analysis uses a specification test to check this assumption. (The [x.sub.ij] vector includes the following: the physician's age and age squared, a dummy variable equalling 1 if physician was foreign trained, a race variable [1 if non Caucasian], a gender variable [1 if female], and county-level data [hospitals per 1,000; hospital beds per 1,000; physicians per 1,000; and median family income]. Including a variety of other control variables does not materially affect the estimates.)

Most physicians were asked to respond to a question about an intermediate office visit and a follow-up hospital visit. Table 2A lists the estimated coefficients on the Medicaid rate for each specialty to these two questions. Look first at the office visit result. The estimated coefficient is positive across all specialties and is statistically significant at the 5% level for 11 of the 13. (The regressions for tables 2, 3 and 4 were checked for heteroskedasticity using the White test; all but 2 do not reject the hypothesis of homoskedasticity. Using robust standard errors does not change the inference results on Medicaid for these two regressions, and thus the analysis uses standard OLS estimates for all the reported standard errors.) All the estimates are under 1 with estimates ranging from 0.07 for Pediatricians up to 0.95 for general practitioners. The sample sizes are relatively small, ranging from 44 (other specialties) to 236 (family practice). The [R.sup.2]s tend to be quite good by cross-sectional standards with the low value being 0.17 for OB/GYN and the high being 0.46 for General Practice. This evidence again supports profit-maximization and hence lends no support to the hypothesis of cost shifting. Additionally, the results are consistent with physicians' exercising [TABULAR DATA FOR TABLE 2B OMITTED] some market power as opposed to acting as price-takers.

Table 2B gives the results for a hospital visit, which are quite similar to those from the office visit regressions. The sample sizes are somewhat smaller than for office visits, ranging from 31 up to 183. All the estimated coefficients again are positive with approximately the same range of values as found with office visits, although the direct correlation between the estimates in the two panels appears weak. Eight of the 13 are statistically significant at the 5% level.

Table 2C gives the results for a variety of specialized procedures. The story is the same: largely positive and statistically significant coefficient estimates. Only two of the 25 estimates are negative, and they are insignificant (Diagnostic D&C and Arthrocentesis). Eighteen of the remaining 23 are positive and statistically significant. Sample size ranges from 33 (EKG cardiology) up to 163 (EKG general/family practice).

These results support the hypothesis of profit-maximization and reject the notion of cost shifting. They also support the hypothesis that physicians exercise some market power.

B. Endogeneity

One possible problem is that the variable of interest, the Medicaid reimbursement, is correlated with the error term, thus inducing a bias in the results. Such a situation could arise if the reimbursement for a particular state is correlated with some unmeasured characteristics of the state or of the physician. A related problem would arise if state reimbursement rates are set as a function of the prevailing standard fees, implying a reverse causality from that assumed above.

One possible test for the exogeneity of the Medicaid reimbursement involves constructing an instrument that is the predicted reimbursement rate for physician i in state j using information in all other states except j. For example, to get an instrument for Californian Medicaid reimbursements, one can run a regression of Medicaid reimbursements on a set of physician characteristics, local demographic characteristics, and procedure type using all physicians in the sample except those in California and then use those estimated coefficients to predict the Californian physician's Medicaid reimbursement. Such an instrument by construction is not contaminated with unmeasured state-specific effects. After estimating the reimbursement rate for physicians in the other states in a similar manner, one can use this instrument to construct a Hausman/Wu specification test that Medicaid payments is an exogenous variable.

The results from this test suggest no serious endogeneity problem. Of the 51 tests (for each of the regressions in table 2), [TABULAR DATA FOR TABLE 2C OMITTED] seven would reject exogeneity at the 5% level, which is higher than expected with a 5% probability of a Type I error, although not substantially out of line. The IV results for those seven failures actually give somewhat stronger results than the OLS estimates. Six of the seven IV estimates are larger than their OLS counterparts and remain statistically significant. Over the full 51 regressions, IV tends to give somewhat wider dispersion in the estimates, but the same general results hold as in the OLS models. The instruments themselves work reasonably well. The first-stage results of the instrumental variables regression show that the instrument is significant at the 10% level in all but five of the 51 regressions, and only seven are insignificant at the 5% level. These results indicate that endogeneity of the Medicaid rate is not a serious problem in these data. (Constructing the specification test involves including both the Medicaid rate and the instrument in an OLS regression and evaluating the t-statistic on the instrument [Hausman, 1978]).

C. Including Physicians Who Do Not Report a Medicaid Price

Another possible problem is that those physicians who report a Medicaid rate, and thus are included in the regressions, may be systematically different from those physicians who do not report a Medicaid rate. Checking for this possibility involves following the test for selection bias outlined in Heckman (1979) for each regression. The test is as follows: (i) Estimate a probit regression with the dependent variable being 1 if the physician reports a Medicaid reimbursement rate and 0 otherwise. (ii) Use the predicted inverse Mills ratio from the probit regression as an additional regressor in the linear regression equation (Equation 5). The test for no selection bias is that the coefficient on the inverse Mills ratio is equal to 0. All the tests indicate no selection bias when the same set of regressors are used in both steps (computation of the standard errors used the formulas discussed in Greene, 1981). Including additional variables in the probit regression that were not in the linear regression reveals some cases where the coefficient on the inverse Mills ratio is statistically significant, but the result generally is that the positive coefficient on the Medicaid rate becomes more significant, indicating that even after accounting for possible selection bias, a rise in the Medicaid rate tends to increase the physician's standard fee. (See Greene, 1993, Ch. 22; Maddala, 1983, Chs. 8 & 9). (The tests with additional regressors in the first-stage probit regression include variables such as % population black, % population urban, % population below poverty line, % population over age 65, and % population under age 14. The results of interest on the Medicaid reimbursement coefficient change very little and thus are not reported here.)

An additional check on this potential problem involves imputing a Medicaid reimbursement rate for each non-reporting physician and then rerunning the regressions including all physicians. Estimating a Medicaid reimbursement rate for non-reporting physicians involves running a regression of the log of the Medicaid reimbursement for all reporting physicians on variables included in the previous regressions, dummy variables for states, specialty and procedure, plus a number of other demographic variables describing the county in which the physician lives. Performing a state-by-state regression is impractical due to the low number of observations for some states. Variables not employed in other regressions described in the paper are included because sample size is not a problem. The predicted reimbursement is then computed by exponentiating the predicted values for the non-reporting physicians plus 1/2 the estimated variance from the regression. The results differ little from those given in table 2.

D. Alternative Tests of Profit-Maximization

Another implication of the profit-maximization model is that the number of Medicaid patients should be negatively correlated with the Medicaid reimbursement rate as outlined in section II. This is a fairly intuitive implication: as the marginal revenue from Medicaid patients decreases from a lowering of the reimbursement rate, a physician will choose to treat fewer of these patients. While not perfectly suited to testing this hypothesis, the PPCIS data also include some questions that shed light on this issue.

Part of the survey asked physicians: (i) the percentage of their patients who had state Medicaid as the primary payer (%M_AID), and (ii) whether the physician was not accepting new Medicaid patients (REJECT). Under profit-maximization, the percentage of Medicaid patients should be positively correlated with the Medicaid reimbursement rate - the higher the Medicaid reimbursement, the higher the percentage of Medicaid patients in a physician's caseload mix. The probability of a physician's rejecting new Medicaid patients should be negatively correlated with the reimbursement rate - the higher the reimbursement, the lower the probability of rejecting a new Medicaid patient.

Ideally, one would have information on the caseload mix (Medicaid versus other types of patients) for each procedure the physician performs as well as the Medicaid reimbursement for that procedure. However, the questions noted above were asked only of the physician's entire practice. The response rate is highest on the question for the prices of an intermediate office visit, and this is a question asked of almost all physician specialties. Therefore, the analysis uses the reported Medicaid reimbursement for that procedure as a proxy for the generosity of the Medicaid reimbursement level for a physician's entire practice. Table 3, panel A, gives the parameter estimates from the linear regression of %M_AID on the Medicaid reimbursement and a set of control variables. [TABULAR DATA FOR TABLE 3 OMITTED] The first column includes only those physicians who report a Medicaid reimbursement. The estimated coefficient is positive and statistically significant, having a value of 0.160 with a standard error of 0.043, suggesting that an increase in the reimbursement rate increases the percentage of Medicaid patients. The second column gives the results when a Medicaid reimbursement is imputed for those physicians who do not report one. The estimated coefficient is very similar at 0.156 and also is statistically significant at standard significance levels.

The OLS estimates do not constrain the predicted values to be in the zero to 100 interval required of percentages. Other tests using a log odds ratio yielded similar results. When other permutations were tried, the positive correlation between reimbursement rates and the number of Medicaid patients came through quite strongly.

Panel B gives the estimates from the probit equations where REJECT is the dependent variable. As with the linear regressions, two sets of results are reported. In the first column, which uses physicians who report a Medicaid reimbursement rate, the estimate is -0.014 with a standard error of 0.005, implying that as the reimbursement rate increases, the probability of not accepting a new Medicaid patient decreases. This is consistent with the findings in the linear models reported in panel A of table 3. The results using imputed values in the second column show very similar results, with a coefficient of -0.015 and a standard error of 0.005.

Taken together with the evidence from the previous sections on prices, these results provide strong evidence that physicians act as profit-maximizers and apparently do not practice cost shifting. In some respects, these results are very troubling. According to the popular "cost shifting" story, a lowering in government prices on Medicaid or Medicare reimbursements leads to a higher price of privately insured patients, but implicitly the assumption remains that the poor being served by the government program will be served regardless. These results suggest otherwise. The evidence reported here suggests that physicians act as profit-maximizers, implying that if the government lowers the reimbursement for the poor, physicians will serve fewer of them.

V. ANALYSIS AND CONCLUSION

Higher Medicaid reimbursement rates are positively correlated with higher fees to privately insured patients for a physician's services, contradicting the popular notion that physicians engage in cost shifting. Using instrumental variables, the analysis checks for exogeneity of the Medicaid rates and finds little evidence that the rates are correlated with the error term. This strengthens the argument that causality runs from the Medicaid rates to the setting of fees rather than vice versa.

While useful, the data do not allow for a full specification of the production or cost function. Therefore, one must view these results with some caution since a more complete specification might reveal interesting interdependencies among procedures in the setting of fees and their relationship with costs. However, the theoretical arguments against cost shifting will be similar with multi-product firms. A more complete specification might lead to different magnitudes for the effect of Medicaid on fees than found in this study, but the signs probably would not change.

The direct implication is that physicians' fees move in the direction of government reimbursements: a cut in Medicaid reimbursements tends to lower fees for non-government patients. Although this study uses only Medicaid information due to the cross-sectional nature of the data, the implications for the other large federal health program, Medicare, are similar. Profit-maximization implies that fees ought to move in the direction of Medicare reimbursements rather than in the opposite direction as suggested by proponents of cost shifting.

Additionally, lower Medicaid reimbursements tend to lead physicians to treat fewer Medicaid patients. Given the results of this paper, the pool of Medicaid patients does not appear "fixed" but rather varies with the reimbursement rate being offered by the government. Obviously, these results one must view with some caution due to limitations inherent in the data, but this general topic of the relationship between reimbursement levels and the utilization of Medicaid services by the eligible population (the "take-up" rate) suggests a fruitful avenue for future research.

A somewhat more indirect implication is that the market for physician services might suffer inefficiency if in fact physicians are acting as local monopolies. Monopoly behavior has well-known inefficiency properties. Thus, analysts might focus on how to make physicians services a more competitive industry - perhaps by eliminating barriers to entry.

ABBREVIATION

PPCIS: Physicians' Practice Costs and Income Survey, expanded version

The author gratefully acknowledges comments of Michael Grossman, Val Lambson, Tom McGuire, and an anonymous referee and the financial support of a Brigham Young University research grant.

REFERENCES

Coulam, Robert F., and Gary L. Gaumer, "Medicare's Prospective Payment System: A Critical Appraisal," Health Care Financing Review: 1991 Annual Supplement, 45-77.

Cromwell, Jerry, Killard W. Adamache, Carol Ammering, William Bartosch, and Ann Boulis, "Equity of the Medicaid Program to the Poor Versus Taxpayers" Health Care Financing Review, Spring 1995, 75-104.

Dranove, David, "Pricing by Non-Profit Institutions: The Case of Hospital Cost-Shifting," Journal of Health Economics, 7:1, 1988, 47-57.

Feldstein, Paul J., Health Care Economics, Delmar Publishers Inc. 4th edition, 1995.

Greene, William H., "Sample Selection Bias as a Specification Error: Comment," Econometrica, 49:3, 1981, 795-798.

-----, Econometric Analysis, Macmillan Publishing Company, 1995.

Hadley, Jack, and Stephen Zuckerman, "Changes in Hospitals' Charges to Privately Insured Patients during PPS," Project Report 3429-07, The Urban Institute, Washington, D.C., 1990.

Hausman, Jerry, "Specification Tests in Econometrics," Econometrica, 46:6, 1978, 1251-1271.

Health Care Financing Administration, Office of the Actuary, Health Care Financing Program Statistics: Analysis of State Medicaid Program Characteristics, 1986, Washington, D.C., 1987.

-----, Health Care Financing Program Statistics: The Medicare and Medicaid Data Book, 1983.

Heckman, James J., "Sample Selection Bias as a Specification Error," Econometrica, 47:1, 1979, 153-161.

McGuire, Thomas G., and Mark V. Pauly, "Physicians Response to Fee Changes With Multiple Payers," Journal of Health Economics, 10:4, 1991, 385-410.

Morrisey, Michael A., Cost-Shifting in Health Care, AEI Press, Washington D.C., 1994.

Morrisey, Michael A., and Frank A. Sloan, "Hospital Cost Shifting and the Medicare Prospective Payment System," Working paper, University of Alabama at Birmingham, 1989.

Morrisey, Michael A., Frank A. Sloan, and J. Valvona, "Shifting Medicare Patients Out of the Hospital," Health Affairs, 7:4, 1988, 52-64.

Sloan, Frank A., Janet Mitchell, and Jerry Cromwell, "Physician Participation in State Medicaid Program," Journal of Human Resources, 12(Supp.), 1978, 211-245.

Zuckerman, Stephen, and John Holahan, "PPS Waivers: Implications for Medicare, Medicaid, and Commercial Insurers," Journal of Health Politics, Policy, and Law, 13:4, 1988, 663-681.

Much of the debate about rising health care costs in the U.S. centers on the notion of "cost shifting." Cost shifting is loosely defined as charging one set of patients a higher price to offset losses on another set of patients. For example, to offset losses from treating charity patients, a hospital might charge a relatively high price to privately insured patients, or a physician might increase his standard fee to compensate for lower Medicare reimbursements.

Despite the certainty of the conventional wisdom, empirical evidence of cost shifting is, at best, mixed. In a study of Illinois hospitals in the early 1980s, Dranove (1988) finds evidence of cost shifting from government patients, primarily Medicaid and Medicare recipients, to private patients. Morrisey and Sloan (1989) find evidence of cost shifting in urban hospitals but not in rural hospitals. Several other empirical studies find no evidence of cost shifting.

One aspect of the cost shifting debate that the empirical work has ignored is whether or not doctors - as opposed to hospitals - practice cost shifting. Although standard profit-maximizing models of physicians would rule out cost shifting, the widely used "income targeting" model of physicians (Feldstein, 1993) does allow for such behavior. The probable reason for the dearth of empirical work is the lack of good data on physicians' pricing behavior. The analysis here uses the Physicians' Practice Costs and Income Survey, 1983-1985 (PPCIS, expanded version) to investigate this important question.

The PPCIS includes a set of questions asking physicians their standard fees and Medicaid reimbursement rates for a variety of medical and surgical procedures. The survey also contains detailed information on the physician and medical conditions in the local geographic area.

The basic research methodology is straightforward. Since Medicaid reimbursements are largely controlled by state legislatures, considerable variation exists in reimbursement rates across the United States. If cost shifting is a widespread phenomenon, one would expect that relatively low reimbursement rates would result in higher fees to privately insured patients, controlling for other factors. If cost shifting does not occur, one would expect Medicaid reimbursement rates to have no effect or possibly result in lower fees depending on the market structure facing the physician. The analysis also examines additional evidence that suggests that lower Medicaid reimbursements lead physicians to treat fewer Medicaid patients.

The results suggest that physicians do not practice "cost shifting" but instead act in a way consistent with the hypothesis of profit-maximization for a firm with some market power. That is, lower Medicaid reimbursements tend to lead to lower prices to privately-insured patients. This appears to be true across a wide spectrum of procedures performed by different specialties. Also consistent with the profit-maximization hypothesis, lower Medicaid reimbursements tend to cause physicians to treat relatively fewer Medicaid patients.

II. ALTERNATIVE MODELS OF PHYSICIAN BEHAVIOR

Health economists widely recognize that profit maximization precludes cost shifting. However, since most of the cost shifting literature focuses on non-profit hospitals that may be more likely to engage in cost shifting, reviewing the implications for physicians seems useful.

Profit maximizing models of physician behavior preclude cost shifting because under a maximization hypothesis, physicians extract as much as possible from all their patients under all circumstances. To say that physicians can shift costs from one set of customers to another, or more specifically, get more revenues from one set of customers to offset a decrease in revenues from another, is to say that the physicians were not profit maximizing in the first place.

To fix this principle, consider a stylized model of physician behavior. A profit-maximizing physician sells a homogeneous good, Q, "physician services" (the treatment of one patient). Assume that the physician has two types of patients. Type 1 patients ([Q.sub.1]) pay through a government program at a fixed rate, [P.sub.1]. The physician can treat as many or as few of these patients at that rate as he wishes. Type 2 patients ([Q.sub.2]) have a downward-sloping demand for the physician's services given by [P.sub.2] ([Q.sub.2]). Obviously, if a physician took [P.sub.2] as fixed as implied by a perfectly competitive model, then there would be no opportunity for cost shifting, so some kind of market power (implied by a downward sloping demand curve) is a necessary condition for cost shifting behavior. Also note that the price charged to the Type 2 patients (the "private" patients) in general will differ from the price that the physician receives for the Type 1 patients (the "public" patients). That the prices are different does not in and of itself imply either the absence or the presence of cost shifting behavior.

The cost of providing the good is a function of the total number of patients the doctor serves, C([Q.sub.1] + [Q.sub.2]). Costs are increasing and strictly convex. To ensure an interior solution, marginal revenue from Type 2 patients must be declining in [Q.sub.2] at the optimum.

The physician chooses the number and type of patients to serve. The profit function is then

(1) [Pi] = [P.sub.1][Q.sub.1] + [P.sub.2]([Q.sub.2])[Q.sub.2] - C([Q.sub.1] + [Q.sub.2])

with C[prime] [greater than] 0, C[double prime] [greater than] 0, [P.sub.2][prime] [less than] 0, and [P.sub.2][double prime] [Q.sub.2] + 2[P.sub.2][prime] [less than] 0 (declining marginal revenue condition) where [prime] and [double prime] denote the first and second derivative, respectively.

The optimization problem is to choose [Q.sub.1] and [Q.sub.2] to maximize profits. The first order conditions of this problem are

(2a) [Delta][Pi]/[Delta][Q.sub.1] = [P.sub.1] - C[prime] = 0

(2b) [Delta][Pi]/[Delta][Q.sub.2] = [P.sub.2][prime][Q.sub.2] + [P.sub.2] - C[prime] = 0

Equation (2a) implies that marginal cost will be set equal to the reimbursement rate, [P.sub.1]. Equation (2b) implies that the marginal revenue from Type 2 patients in turn also will be set equal to the reimbursement rate. This is a standard result - i.e., the marginal revenues across different markets will be equalized at the optimum. But these results rule out the shifting of costs to Type 2 patients. Cost shifting in this model is defined as the charging of a higher price to Type 2 patients when the reimbursement rate to Type 1 patients is lowered. But working through the implications of the first order conditions yields exactly the opposite result.

Suppose [P.sub.1] falls. Because marginal revenues are equated at the margin, the marginal revenue of Type 2 patients also should fall. But the assumption for profit maximization implies that marginal revenue is declining in [Q.sub.2] so a decrease in marginal revenue implies an increase in [Q.sub.2] itself. But since demand is downward sloping, an increase in [Q.sub.2] would imply a decrease in [P.sub.2]. So, as [P.sub.1] falls, the price charged to Type 2 patients also would fall; hence, the hypothesis of cost shifting is rejected.

More directly, (2a) and (2b) imply that

(3) [P.sub.2][prime][Q.sub.2] + [P.sub.2] = [P.sub.1]

Taking a total derivative of this equation and recognizing that d[P.sub.2] = [P.sub.2][prime]d[Q.sub.2], one gets

(4) d[P.sub.2]/d[P.sub.1] = [P.sub.2][prime]/[P.sub.2][double prime][Q.sub.2] + 2[P.sub.2][prime]

Since both the numerator and the denominator are negative, this quantity is positive, implying that as [P.sub.2] increases, so does [P.sub.1]; [P.sub.2] and [P.sub.1] move in the same direction in contradiction to the cost shifting story.

An additional result is that as [P.sub.1] falls, [Q.sub.1] also falls, meaning that as the reimbursement rate falls, the number of government-financed patients also decreases. This result also hinges on the second order conditions of the maximization problem, but the intuition is straightforward: as the marginal revenue from the government-financed patients falls, the physician will choose to serve fewer of them.

When the profit-maximization hypothesis is dropped, so is the key theoretical constraint against cost shifting. Dranove (1988) presents one example of an alternative model to strict profit maximization. He posits a model where a firm has a utility function that takes as arguments profits and output. The motivation is that for non-profit organizations, other factors besides profits are likely to influence production decisions. An important result is that for cost shifting to occur in this framework, the firm must exhibit some market power.

McGuire and Pauly (1991) present another optimization model based on demand-inducement behavior of physicians. In this model, physicians can increase revenues from a set of patients but only by incurring disutility for engaging in demand inducement. Such a model allows for cost shifting (gaining more revenues from one group as revenues from a second group exogenously declines) if income effects are sufficiently strong.

Empirical tests for cost shifting focus primarily on hospitals where the profit-maximization hypothesis probably is not the best model. Hadley and Zuckerman (1990); Morrisey, Sloan and Valvona (1988); Zuckerman and Holahan (1988); and Morrisey and Sloan (1989) find little evidence of pervasive cost shifting. (Morrisey, 1994, and Coulam and Gaumer, 1991, review the existing empirical evidence on cost shifting.) Although physician services are an important component of the health service industry, they have received little empirical attention in the cost shifting debate due to the paucity of micro-level data on physician pricing behavior. This study begins to fill this gap by using PPCIS data.

III. DATA DESCRIPTION

The primary data for this study are the 1983-1985 PPCIS conducted by the National Opinion Research Center under contract to the Health Care Financing Administration. This cross-sectional survey was conducted from October 1984 through June 1985 and includes responses from 4,729 physicians (out of 6,847 eligible) drawn from a stratified random sample of physicians from the American Medical Association's 1984 Physician Master File. The physicians were asked numerous detailed questions regarding pricing policies, reimbursement rates, and practice characteristics. The data also contain variables concerning the physicians' personal characteristics (age, sex, specialty, etc.) obtained from the AMA Physician Masterfile. The expanded version of the 1983 PPCIS contains county-level demographic information (age composition, ethnic composition, MDs per capita, etc.).

The strength of the data is their uniqueness. Depending on their specialty, self-employed physicians were asked for standard fees and reimbursement rates (Medicaid and Medicare) for three to five different medical procedures. For example, family practice physicians were asked for several different prices on an intermediate office visit, an intermediate follow-up hospital visit, a vaginal delivery, and an EKG report. Most specialists were asked questions about the intermediate office visit and the hospital follow-up visit.

Having both a Medicaid price that will vary by state of residence and a "usual fee" for a given procedure allows one to test for cost shifting straightforwardly by checking for a statistically significant negative correlation between Medicaid reimbursements and usual fees, controlling for other observable characteristics. The alternative hypothesis is that physicians are profit-maximizers, which would imply by the simple model in the previous section that a positive correlation should exist between Medicaid reimbursements and fees charged to non-Medicaid patients.

Such a test relies on the exogeneity of Medicaid reimbursements at the state level. This assumption seems plausible given the wide latitude individual state legislatures have in setting Medicaid reimbursement rates for physicians. Indeed, the freedom states have in setting the parameters of the Medicaid system and the resulting variation across states in how Medicaid operates has been the subject of some controversy (e.g., Cromwell et al., 1995). However, but for the purposes of this study, such arbitrary variation obviously is beneficial. The analysis here also checks the exogeneity assumption using instrumental variables estimation.

As with most data, the PPCIS also has some weaknesses. First, it is cross-sectional, meaning that unobservable physician characteristics cannot be controlled for as they would be, for example, with a panel data set. Second, the price information is self-reported by the physician or a designated representative. Such self-reporting possibly could be biased although the direction of the bias is unclear and a priori seems unlikely to favor one hypothesis over another (cost shifting versus profit maximization).

Another limitation of the data is that only a small subset of procedures is sampled for each physician instead of the full range of physician activities. This precludes the estimation of a fully specified cost or production function for physician services.

Finally, although the PPCIS is unique in its composition, it was conducted early in the current phase of changes in the health care industry. The logical presumption is that the market for physician services has become more competitive since this survey was taken. This would imply that physicians' ability to "cost shift" has lessened, implying that any findings of cost shifting using these data represent an upper bound estimate for cost shifting in the current environment.

Table 1 gives summary statistics for the data. Prices have been adjusted to account for differences in costs-of-living across states using the American Chamber of Commerce Association's cross-sectional price index. The price index is based on a population-weighted index of city-level price indexes reported in 1985 (most information is from the 1st quarter report, but a few states have no reported cities for the 1 st quarter, so the analysis uses the next closest quarter with a reporting city, either the 3rd quarter or the 4th of 1985). The weights are from the 1980 census. Several variations on the price index do not materially change the results. (Details on this and other aspects of the paper are available from the author.)

The table gives information on 1,915 physicians who report both a standard fee and a Medicaid reimbursement rate. The information on fees and reimbursements includes all procedures reported by a given physician.

IV. EMPIRICAL RESULTS

Intuitively, if cost shifting is a widespread phenomena operating through the fees that physicians charge and if the mechanism is that low Medicaid reimbursements result in higher fees to non-Medicaid patients, one might be able to see differences at the aggregate state level since Medicaid reimbursements vary widely by state.

A simple test is to compute the mean Medicaid reimbursement and mean standard fee for the most frequently answered procedure (intermediate office visit for an established patient) by state and then compute the correlation across states. (Hawaii, Alaska, and Arizona have been excluded. Cost-of-living information was not available for Hawaii. Alaska has too few respondents to be of value. Arizona obtained a waiver from participating in the Medicaid program.) The Pearson Correlation is 0.37, which is consistent with profit-maximization and contradicts cost shifting.

[TABULAR DATA FOR TABLE 1 OMITTED]

Next, the analysis runs a weighted least squares regression of the mean standard fee on a constant and the state mean Medicaid reimbursement for an office visit where the weight is the number of surveyed physicians with a value for their office fee in each state. The coefficient on the Medicaid is 0.43 with a 2.4 t-statistics. The interpretation of this coefficient is that not adjusting for any other factors, a $1 increase in the Medicaid reimbursement rate will raise the average standard fee by 43 cents. The [R.sup.2] of the regression is relatively low at 0.11, but this simple evidence lends no support to the notion of cost shifting.

A. Regression Analysis

Such simple correlations suggest a positive relationship. However, after controlling for other demographic characteristics, the gross positive correlation might disappear. Therefore, a more detailed regression model is necessary:

(5) [fee.sub.ij] = [[Beta].sub.0] + [[Beta].sub.1][medaid.sub.ij] + [x.sub.ij][Beta] + [[Epsilon].sub.ij]

where i denotes the physician, j denotes the state, fee is the standard fee for a given medical procedure, medaid is the Medicaid reimbursement for that procedure, and [x.sub.ij] is vector [TABULAR DATA FOR TABLE 2A OMITTED] of other explanatory variables including the characteristics of the doctor and local demographic (county level) variables. Under the simplest interpretation of cost shifting, a decrease in Medicaid payments should result in an increase in the standard fee, which would imply a negative value for [[Beta].sub.1]. The initial assumption is that the Medicaid rate is set exogenously by state legislatures. Later analysis uses a specification test to check this assumption. (The [x.sub.ij] vector includes the following: the physician's age and age squared, a dummy variable equalling 1 if physician was foreign trained, a race variable [1 if non Caucasian], a gender variable [1 if female], and county-level data [hospitals per 1,000; hospital beds per 1,000; physicians per 1,000; and median family income]. Including a variety of other control variables does not materially affect the estimates.)

Most physicians were asked to respond to a question about an intermediate office visit and a follow-up hospital visit. Table 2A lists the estimated coefficients on the Medicaid rate for each specialty to these two questions. Look first at the office visit result. The estimated coefficient is positive across all specialties and is statistically significant at the 5% level for 11 of the 13. (The regressions for tables 2, 3 and 4 were checked for heteroskedasticity using the White test; all but 2 do not reject the hypothesis of homoskedasticity. Using robust standard errors does not change the inference results on Medicaid for these two regressions, and thus the analysis uses standard OLS estimates for all the reported standard errors.) All the estimates are under 1 with estimates ranging from 0.07 for Pediatricians up to 0.95 for general practitioners. The sample sizes are relatively small, ranging from 44 (other specialties) to 236 (family practice). The [R.sup.2]s tend to be quite good by cross-sectional standards with the low value being 0.17 for OB/GYN and the high being 0.46 for General Practice. This evidence again supports profit-maximization and hence lends no support to the hypothesis of cost shifting. Additionally, the results are consistent with physicians' exercising [TABULAR DATA FOR TABLE 2B OMITTED] some market power as opposed to acting as price-takers.

Table 2B gives the results for a hospital visit, which are quite similar to those from the office visit regressions. The sample sizes are somewhat smaller than for office visits, ranging from 31 up to 183. All the estimated coefficients again are positive with approximately the same range of values as found with office visits, although the direct correlation between the estimates in the two panels appears weak. Eight of the 13 are statistically significant at the 5% level.

Table 2C gives the results for a variety of specialized procedures. The story is the same: largely positive and statistically significant coefficient estimates. Only two of the 25 estimates are negative, and they are insignificant (Diagnostic D&C and Arthrocentesis). Eighteen of the remaining 23 are positive and statistically significant. Sample size ranges from 33 (EKG cardiology) up to 163 (EKG general/family practice).

These results support the hypothesis of profit-maximization and reject the notion of cost shifting. They also support the hypothesis that physicians exercise some market power.

B. Endogeneity

One possible problem is that the variable of interest, the Medicaid reimbursement, is correlated with the error term, thus inducing a bias in the results. Such a situation could arise if the reimbursement for a particular state is correlated with some unmeasured characteristics of the state or of the physician. A related problem would arise if state reimbursement rates are set as a function of the prevailing standard fees, implying a reverse causality from that assumed above.

One possible test for the exogeneity of the Medicaid reimbursement involves constructing an instrument that is the predicted reimbursement rate for physician i in state j using information in all other states except j. For example, to get an instrument for Californian Medicaid reimbursements, one can run a regression of Medicaid reimbursements on a set of physician characteristics, local demographic characteristics, and procedure type using all physicians in the sample except those in California and then use those estimated coefficients to predict the Californian physician's Medicaid reimbursement. Such an instrument by construction is not contaminated with unmeasured state-specific effects. After estimating the reimbursement rate for physicians in the other states in a similar manner, one can use this instrument to construct a Hausman/Wu specification test that Medicaid payments is an exogenous variable.

The results from this test suggest no serious endogeneity problem. Of the 51 tests (for each of the regressions in table 2), [TABULAR DATA FOR TABLE 2C OMITTED] seven would reject exogeneity at the 5% level, which is higher than expected with a 5% probability of a Type I error, although not substantially out of line. The IV results for those seven failures actually give somewhat stronger results than the OLS estimates. Six of the seven IV estimates are larger than their OLS counterparts and remain statistically significant. Over the full 51 regressions, IV tends to give somewhat wider dispersion in the estimates, but the same general results hold as in the OLS models. The instruments themselves work reasonably well. The first-stage results of the instrumental variables regression show that the instrument is significant at the 10% level in all but five of the 51 regressions, and only seven are insignificant at the 5% level. These results indicate that endogeneity of the Medicaid rate is not a serious problem in these data. (Constructing the specification test involves including both the Medicaid rate and the instrument in an OLS regression and evaluating the t-statistic on the instrument [Hausman, 1978]).

C. Including Physicians Who Do Not Report a Medicaid Price

Another possible problem is that those physicians who report a Medicaid rate, and thus are included in the regressions, may be systematically different from those physicians who do not report a Medicaid rate. Checking for this possibility involves following the test for selection bias outlined in Heckman (1979) for each regression. The test is as follows: (i) Estimate a probit regression with the dependent variable being 1 if the physician reports a Medicaid reimbursement rate and 0 otherwise. (ii) Use the predicted inverse Mills ratio from the probit regression as an additional regressor in the linear regression equation (Equation 5). The test for no selection bias is that the coefficient on the inverse Mills ratio is equal to 0. All the tests indicate no selection bias when the same set of regressors are used in both steps (computation of the standard errors used the formulas discussed in Greene, 1981). Including additional variables in the probit regression that were not in the linear regression reveals some cases where the coefficient on the inverse Mills ratio is statistically significant, but the result generally is that the positive coefficient on the Medicaid rate becomes more significant, indicating that even after accounting for possible selection bias, a rise in the Medicaid rate tends to increase the physician's standard fee. (See Greene, 1993, Ch. 22; Maddala, 1983, Chs. 8 & 9). (The tests with additional regressors in the first-stage probit regression include variables such as % population black, % population urban, % population below poverty line, % population over age 65, and % population under age 14. The results of interest on the Medicaid reimbursement coefficient change very little and thus are not reported here.)

An additional check on this potential problem involves imputing a Medicaid reimbursement rate for each non-reporting physician and then rerunning the regressions including all physicians. Estimating a Medicaid reimbursement rate for non-reporting physicians involves running a regression of the log of the Medicaid reimbursement for all reporting physicians on variables included in the previous regressions, dummy variables for states, specialty and procedure, plus a number of other demographic variables describing the county in which the physician lives. Performing a state-by-state regression is impractical due to the low number of observations for some states. Variables not employed in other regressions described in the paper are included because sample size is not a problem. The predicted reimbursement is then computed by exponentiating the predicted values for the non-reporting physicians plus 1/2 the estimated variance from the regression. The results differ little from those given in table 2.

D. Alternative Tests of Profit-Maximization

Another implication of the profit-maximization model is that the number of Medicaid patients should be negatively correlated with the Medicaid reimbursement rate as outlined in section II. This is a fairly intuitive implication: as the marginal revenue from Medicaid patients decreases from a lowering of the reimbursement rate, a physician will choose to treat fewer of these patients. While not perfectly suited to testing this hypothesis, the PPCIS data also include some questions that shed light on this issue.

Part of the survey asked physicians: (i) the percentage of their patients who had state Medicaid as the primary payer (%M_AID), and (ii) whether the physician was not accepting new Medicaid patients (REJECT). Under profit-maximization, the percentage of Medicaid patients should be positively correlated with the Medicaid reimbursement rate - the higher the Medicaid reimbursement, the higher the percentage of Medicaid patients in a physician's caseload mix. The probability of a physician's rejecting new Medicaid patients should be negatively correlated with the reimbursement rate - the higher the reimbursement, the lower the probability of rejecting a new Medicaid patient.

Ideally, one would have information on the caseload mix (Medicaid versus other types of patients) for each procedure the physician performs as well as the Medicaid reimbursement for that procedure. However, the questions noted above were asked only of the physician's entire practice. The response rate is highest on the question for the prices of an intermediate office visit, and this is a question asked of almost all physician specialties. Therefore, the analysis uses the reported Medicaid reimbursement for that procedure as a proxy for the generosity of the Medicaid reimbursement level for a physician's entire practice. Table 3, panel A, gives the parameter estimates from the linear regression of %M_AID on the Medicaid reimbursement and a set of control variables. [TABULAR DATA FOR TABLE 3 OMITTED] The first column includes only those physicians who report a Medicaid reimbursement. The estimated coefficient is positive and statistically significant, having a value of 0.160 with a standard error of 0.043, suggesting that an increase in the reimbursement rate increases the percentage of Medicaid patients. The second column gives the results when a Medicaid reimbursement is imputed for those physicians who do not report one. The estimated coefficient is very similar at 0.156 and also is statistically significant at standard significance levels.

The OLS estimates do not constrain the predicted values to be in the zero to 100 interval required of percentages. Other tests using a log odds ratio yielded similar results. When other permutations were tried, the positive correlation between reimbursement rates and the number of Medicaid patients came through quite strongly.

Panel B gives the estimates from the probit equations where REJECT is the dependent variable. As with the linear regressions, two sets of results are reported. In the first column, which uses physicians who report a Medicaid reimbursement rate, the estimate is -0.014 with a standard error of 0.005, implying that as the reimbursement rate increases, the probability of not accepting a new Medicaid patient decreases. This is consistent with the findings in the linear models reported in panel A of table 3. The results using imputed values in the second column show very similar results, with a coefficient of -0.015 and a standard error of 0.005.

Taken together with the evidence from the previous sections on prices, these results provide strong evidence that physicians act as profit-maximizers and apparently do not practice cost shifting. In some respects, these results are very troubling. According to the popular "cost shifting" story, a lowering in government prices on Medicaid or Medicare reimbursements leads to a higher price of privately insured patients, but implicitly the assumption remains that the poor being served by the government program will be served regardless. These results suggest otherwise. The evidence reported here suggests that physicians act as profit-maximizers, implying that if the government lowers the reimbursement for the poor, physicians will serve fewer of them.

V. ANALYSIS AND CONCLUSION

Higher Medicaid reimbursement rates are positively correlated with higher fees to privately insured patients for a physician's services, contradicting the popular notion that physicians engage in cost shifting. Using instrumental variables, the analysis checks for exogeneity of the Medicaid rates and finds little evidence that the rates are correlated with the error term. This strengthens the argument that causality runs from the Medicaid rates to the setting of fees rather than vice versa.

While useful, the data do not allow for a full specification of the production or cost function. Therefore, one must view these results with some caution since a more complete specification might reveal interesting interdependencies among procedures in the setting of fees and their relationship with costs. However, the theoretical arguments against cost shifting will be similar with multi-product firms. A more complete specification might lead to different magnitudes for the effect of Medicaid on fees than found in this study, but the signs probably would not change.

The direct implication is that physicians' fees move in the direction of government reimbursements: a cut in Medicaid reimbursements tends to lower fees for non-government patients. Although this study uses only Medicaid information due to the cross-sectional nature of the data, the implications for the other large federal health program, Medicare, are similar. Profit-maximization implies that fees ought to move in the direction of Medicare reimbursements rather than in the opposite direction as suggested by proponents of cost shifting.

Additionally, lower Medicaid reimbursements tend to lead physicians to treat fewer Medicaid patients. Given the results of this paper, the pool of Medicaid patients does not appear "fixed" but rather varies with the reimbursement rate being offered by the government. Obviously, these results one must view with some caution due to limitations inherent in the data, but this general topic of the relationship between reimbursement levels and the utilization of Medicaid services by the eligible population (the "take-up" rate) suggests a fruitful avenue for future research.

A somewhat more indirect implication is that the market for physician services might suffer inefficiency if in fact physicians are acting as local monopolies. Monopoly behavior has well-known inefficiency properties. Thus, analysts might focus on how to make physicians services a more competitive industry - perhaps by eliminating barriers to entry.

ABBREVIATION

PPCIS: Physicians' Practice Costs and Income Survey, expanded version

The author gratefully acknowledges comments of Michael Grossman, Val Lambson, Tom McGuire, and an anonymous referee and the financial support of a Brigham Young University research grant.

REFERENCES

Coulam, Robert F., and Gary L. Gaumer, "Medicare's Prospective Payment System: A Critical Appraisal," Health Care Financing Review: 1991 Annual Supplement, 45-77.

Cromwell, Jerry, Killard W. Adamache, Carol Ammering, William Bartosch, and Ann Boulis, "Equity of the Medicaid Program to the Poor Versus Taxpayers" Health Care Financing Review, Spring 1995, 75-104.

Dranove, David, "Pricing by Non-Profit Institutions: The Case of Hospital Cost-Shifting," Journal of Health Economics, 7:1, 1988, 47-57.

Feldstein, Paul J., Health Care Economics, Delmar Publishers Inc. 4th edition, 1995.

Greene, William H., "Sample Selection Bias as a Specification Error: Comment," Econometrica, 49:3, 1981, 795-798.

-----, Econometric Analysis, Macmillan Publishing Company, 1995.

Hadley, Jack, and Stephen Zuckerman, "Changes in Hospitals' Charges to Privately Insured Patients during PPS," Project Report 3429-07, The Urban Institute, Washington, D.C., 1990.

Hausman, Jerry, "Specification Tests in Econometrics," Econometrica, 46:6, 1978, 1251-1271.

Health Care Financing Administration, Office of the Actuary, Health Care Financing Program Statistics: Analysis of State Medicaid Program Characteristics, 1986, Washington, D.C., 1987.

-----, Health Care Financing Program Statistics: The Medicare and Medicaid Data Book, 1983.

Heckman, James J., "Sample Selection Bias as a Specification Error," Econometrica, 47:1, 1979, 153-161.

McGuire, Thomas G., and Mark V. Pauly, "Physicians Response to Fee Changes With Multiple Payers," Journal of Health Economics, 10:4, 1991, 385-410.

Morrisey, Michael A., Cost-Shifting in Health Care, AEI Press, Washington D.C., 1994.

Morrisey, Michael A., and Frank A. Sloan, "Hospital Cost Shifting and the Medicare Prospective Payment System," Working paper, University of Alabama at Birmingham, 1989.

Morrisey, Michael A., Frank A. Sloan, and J. Valvona, "Shifting Medicare Patients Out of the Hospital," Health Affairs, 7:4, 1988, 52-64.

Sloan, Frank A., Janet Mitchell, and Jerry Cromwell, "Physician Participation in State Medicaid Program," Journal of Human Resources, 12(Supp.), 1978, 211-245.

Zuckerman, Stephen, and John Holahan, "PPS Waivers: Implications for Medicare, Medicaid, and Commercial Insurers," Journal of Health Politics, Policy, and Law, 13:4, 1988, 663-681.

Printer friendly Cite/link Email Feedback | |

Author: | Showalter, Mark H. |
---|---|

Publication: | Contemporary Economic Policy |

Date: | Apr 1, 1997 |

Words: | 5467 |

Previous Article: | Federal water policy reform: implications for irrigated farms in California. |

Next Article: | Social security reform in China: the case of old-age insurance. |

Topics: |