The Initial Effects of Physician Compensation Programs in Taiwan Hospitals: Implications for Staff Model HMOs

 2003 Kluwer Academic Publishers. Manufactured in The Netherlands.

The Initial Effects of Physician Compensation Programs

in Taiwan Hospitals:

Implications for Staff Model HMOs

Department of Accountancy, National Taipei University, Taiwan, ROC SHUEN-ZEN LIU

Department of Accounting, College of Management, National Taiwan University, Taiwan, ROC JAMES C. ROMEIS∗

Health Services Research, School of Public Health, Saint Louis University, St. Louis, MO 63104-1314, USA

E-mail: romeisjc@slu.edu CHIH-LIANG YAUNG

Institution of Health Care Organization Administration, College of Public Health, National Taiwan University, Taiwan, ROC

Abstract. This paper examines whether a Physician Compensation Program (PCP), which was based on the responsibility centers system,

improved departmental efficiency in a large Taiwan teaching hospital. PCPs in Taiwan may have implications for staff-model HMOs. Monthly financial data and related information for 58 departments in the 5 months following the introduction of the program (the PCP period) and the corresponding 5 months before the introduction of the program (the pre-PCP period) were provided by the case hospital. The Data Envelopment Analysis (DEA) model is used to measure the operational efficiency of each department in the case hospital over the two periods. We first use asymptotic DEA-based tests to examine whether differences in efficiency scores between the two periods are significant. Then, a multi-factor tobit model is used to examine factors that might explain the observed differences in efficiency. The data of each month in the PCP period (November 1996–March 1997) and the pre-PCP period (November 1995–March 1996) are used to calculate efficiency scores and control for monthly effects. We find that average efficiency improves after the implementation of the PCP, with or without controlling for other related factors. Physicians’ seniority and percentage of physicians’ service time in the department are associated with improved efficiency. Finally, departments with higher profits and fewer numbers of employees are associated with higher efficiency. The findings suggest that to achieve an increase in hospital efficiency in Taiwan, responsibility centers should be integrated with formal physician compensation programs. Such results have implications for staff model HMOs in the US and their variants in countries with national health insurance.

Keywords: Physician Compensation, staff-model HMO, Data Envelopment Analysis (DEA)

1. Introduction

If they are lucky, the health services research community gets to observe macro level changes in a country’s delivery sys-tem once in their professional careers. This paper is the first of a series of reports that cover Taiwan’s implementation of National Health Insurance and has implications for countries that have policy interest in staff model HMO organizational structures [22].

Taiwan’s National Health Insurance (NHI) program was established on 1 March, 1995. It provides health care cover-age to almost all residents (roughly 97% in 1999) in Taiwan. This is a remarkable change from implementation where ap-proximately 40% of the population (mostly children, women and older adults) had no insured access to care. Within the first year of implementation approximately 95% had cover-age. The reimbursement scheme under the NHI is mainly ∗_{Corresponding author.}

fee-for-service with some diagnoses reimbursed using DRGs methods, e.g., referred to as case payment in Taiwan. The long-term goal of the NHI is moving towards a fully capitated system. One important feature of Taiwan’s delivery system is its commitment to large hospital-based organizations. Unlike the US with its commitment to ambulatory care and primary care gatekeepers, Taiwan is committed to large tertiary care institutions and sees no need to develop primary care at this time.

The implementation of NHI did not cause Taiwanese hos-pitals to change organizational structures. By statute, su-perintendents or the CEOs must be physicians and therefore loosely resemble US hospitals when they were administra-tively dominated by physicians. From a more contemporary perspective, another way of conceptualizing the administra-tive structures is a staff-model HMO, e.g., in addition to the central administrators, staff physicians are salaried employ-ees of the hospital who organize and provide specialty and sub-specialty care.

The issue of hospital–physician integration on efficiency and financial performance has generated considerable inter-ests in recent literature because based on US experiences physicians incur most of operating costs in hospitals through their clinical judgments. Wilensky and Rossiter [23,24] re-viewed the finding from a series of studies on demand induce-ment using data from the National Medical Care Expenditure Survey (NMCES) of the National Center for Health Services Research. They stated that 90% of health care expenditures were physician initiated or at least physician controlled. Goes and Zhan [16] found that higher degree of financial integra-tion between hospitals and physicians (e.g., closer hospital supervision over physician decisions) was related to lower operating costs. The purpose of this paper is to enhance our understanding regarding the impact of hospital–physician integration strategy using a field study of a Taiwan hospi-tal. Specifically, this paper evaluates the initial effects on de-partmental efficiency after the implementation of a physician compensation program (PCP). In addition to the effects of the PCP, departmental efficiency may be influenced by factors re-lated to departmental characteristics and physicians’ profiles. Hence, those factors are also investigated and controlled. The results appear to have implications for staff-model HMO organizational structures because of the domination by physi-cian executives. Based on Goes and Zhan’s [16] measure-ment, the staff-model represents the closest type of hospital– physician integration and should be highly efficient. However, we believe that agency problems still exist when physicians only receive flat salaries. Thus, this study pro-vides a good opportunity to examine the effects of more sub-tle arrangements between physicians and hospitals within a staff-model context.

1.1. The case hospital

The case hospital is affiliated with a national university and is recognized as one of the most prestigious public teaching hospitals in Taiwan. The case hospital currently has 4,693 employees (including 873 full time physicians, 450 part time physicians, 1,308 nurses, 96 pharmacists and 1,966 other staff positions). It treats approximately 6,000 outpatients per day and has roughly 2,000 beds; the occupancy rate is about 84% with an average LOS 11.5 days.

The case hospital had experienced consecutive operational losses for decades and viewed the losses as natural conse-quences of high quality teaching and research. It received governmental subsidy to cover the losses each year. But as the government tightened the subsidy, competition in the health care market increased and National Health Insurance (NHI) was on the horizon, the case hospital decided to overhaul its hospital–physician integration strategy with a goal of enhanc-ing efficiency.

In December 1993, the case hospital formed a task team to design a responsibility centers system in which depart-ments were dividend into cost centers and profit centers to collect financial data. The task team was led by the vice-superintendent (a physician) and included directors from

re-lated administrative functions (e.g., accounting, information system, etc.). Physician representatives were consulted in the process when controversial issues regarding medical costs or revenues allocation occurred. In August 1995, the superinten-dent held four hospital-wide seminars to formally introduce the mechanism of responsibility centers and gather feedback, with over 400 participants. At the same time, financial reports of responsibility centers were distributed to each responsibil-ity center. Many physicians greeted initial financial reports in doubt and frustration because they felt some accounting data were inaccurate or misleading.

For example, in several incidents the amount of medical services provided by each physician in the same responsibility center was seriously distorted. Further investigation revealed that the mistake was because physicians shared the same com-puter account in clinical practices. With the expectation of upcoming performance evaluations linking to the accounting data, many physicians showed high interest in correcting er-rors in the reports or confronting the task team regarding how accounting data were generated. From August to October 1995, intensive revisions concerning financial reporting of re-sponsibility centers were conducted to ensure integrity of the system. As the accuracy of financial data became satisfac-tory and most physicians were familiar with the operation of responsibility centers, the case hospital started building con-sensus on how to link pay with clinical performance. It for-mally implemented a physician compensation program (PCP) in November 1996.

Prior to the PCP, physicians in the case hospital were paid based on seniority and rank (e.g., superintendent and departmental directors received additional pay); physicians’ compensation was not related to their performance. Un-der the PCP, physicians’ compensation consists of base salary plus incentives. The maximum amount of incen-tives available to physicians equals 10% of total hospi-tal medical revenues, under the condition of no opera-tion loss in that year. For each physician, the incentive is based on his/her contribution score. The score is deter-mined by 70% weight of individual contribution (as meas-ured by the ranking of total medical revenues generated by the physician minus his base salary in the hospital) and 30% weight of departmental contribution (as measured by the ranking of departmental profit per physician in the hos-pital). The program was designed to encourage physicians to increase clinical revenues while controlling departmental costs through team work, important steps to enhance prof-itability under a reimbursement scheme mainly based on fee-for-service. Although some profit centers were deleted in the sample selection process, most of the major ones are included in the sample (e.g., cardiology, gastroenterology, hepatology, nephrology, ontology, thoracic surgery, cardiac surgery, or-thopaedic surgery, pediatrics, gynecology, obstetrics, etc.). Deleted profit centers tended to be smaller and less signifi-cant to the overall mission and revenue of the case hospital (e.g., cosmetic surgery). Thus, we believe that the sampled departments reasonably represent the operation of the PCP.

2. Research design

2.1. The Data Envelopment Analysis (DEA) model

Most literature within DEA deals with the issues of oper-ational efficiency, pure technical efficiency, allocative effi-ciency, and scale efficiency among decision-making units in the same period [12,14,15]. The deterministic DEA model is constructed to measure the efficiency of each department relative to the performance of other departments in the case hospital.

We model departments in the case hospital as multi-input, single output Decision-Making Units (DMUs), which attempt to maximize the achievable output given inputs and technol-ogy. The output measure of efficiency (φ) can be evaluated for any observation j . Here, “o” denotes a focal department. Each department, in turn, becomes a focal department when its efficiency score is computed as the solution to the envel-opment problem, which is the dual to the linear programming problem as follows [14]: max φ,λ φ subject to φyom
J
j=1 λjyjm, J
j=1 λjxjn
xon, λj  0. (1)

Here, λj is an intensity or activity variable and yjm, xjnare the outputs and inputs of the j th department. In equation (1), the performance of a department is evaluated in terms of the ability to expand its output subject to the constraints imposed by the best-observed practice. If expansion is possible for a department, its optimal φ∗ is greater than 1; otherwise, its optimal φ∗ equals 1. The DEA model above imposes an as-sumption of constant returns to scale. Banker et al. [6] sug-gests an extension of the model to account for the case of vari-able returns to scale by adding to equation (1) the constraint ofJ_j=1λj = 1. As the implication of efficiency will be different under different returns to scale assumptions [2], the paper tests what kind of returns to scale the sample depart-ments are first. In this study, we use DEA-based statistical tests to examine the returns to scale assumption. The tests have been used to evaluate returns to scale for software devel-opment projects [21] and software maintenance projects [8].

2.2. Hypothesis testing in the DEA analysis

Nonparametric tests are used to test for efficiency differ-ences between two groups of DMUs in the DEA literature. Banker [10] suggested statistical tests for inefficiency differ-ences between two groups of DMUs. We follow Banker et al. [17] to test whether departmental efficiency of the case hospital improved in the PCP period. The null hypothesis is that there is no difference in average departmental efficiency

between the pre-PCP and PCP periods. The alternative hy-pothesis is that departments are on average less efficient in the pre-PCP period when compared with the PCP period. To test the hypothesis, let i represents a department in the over-all data set. Set I of departments consists of two subsets, T1 and T2. Department i is an element of T1 if the department is in the pre-PCP period; otherwise, it is an element of T2. We denote the efficiency of department i in group Tj by λj_i and allow for the possibility that the probability distribution of λ1_i differs from that of λ2_i. Specifically, the null hypothesis is that there is no difference in average departmental efficiency be-tween T1 and T2. The alternative hypothesis is that group T1 is on average less efficient than group T2.

2.3. The Tobit regression

We also follow prior literature [3,11,20] and use the tobit model to control for factors that may explain the observed dif-ferences in inefficiency across departments in addition to the PCP. Since efficiency scores computed from the DEA model are censored at one, an OLS regression would produce bi-ased and inconsistent parameter estimates [9]. Tobit analysis assumes that the dependent variable has a number of its val-ues clustered at a limiting value. A convenient normalization often used in the literature is to assume a censoring point at zero. The general formulation is defined as follows [9]:

y_i∗= Bxi+ εi,

yi= y_i∗, if y_i∗>0,

yi= 0, otherwise.

In our study, the dependent variable yi is (φ− 1). Here,

φ denotes the efficiency score. The independent variables

xi represent factors that may be related to departmental ef-ficiency. We also examine the possible heteroscedasticity in our tobit model. Specifically, if heteroscedasticity is ignored in the model, the resulting estimates are not even consis-tent. In addition, it is more practicable to make reasonable assumptions on the nature of heteroscedasticity and estimate the model accordingly rather than to claim that estimates are inconsistent if heteroscedasticity is ignored [18].

3. Sample selection and variable measurement

3.1. Sample and data sources

The case hospital provided us with monthly financial data and related information for all profit centers during the two peri-ods: November 1995 to March 1996 and November 1996 to March 1997. The PCP was implemented in October 1996, and hence we designated the 5 months from November 1995 to March 1996 as the pre-PCP period and the 5 months from November 1996 to March 1997 as the PCP period. Here, we will follow prior literature [9,11,18] and compute an inter-temporal frontier. Our main objective is to evaluate the im-pact of the PCP on the case hospital by measuring the change

Table 1

Descriptive statistics of the inputs and output.a

Month Mean Standard deviation

Inputs Output Inputs Output

Personnel Medicine Depreciation Total Personnel Medicine Depreciation Total

costs costs costs revenue costs costs costs revenue

95/11 68,750 32,188 20,000 262,500 65,938 55,625 25,625 231,563 95/12 66,250 29,688 21,250 251,563 63,438 54,063 26,563 217,813 96/01 65,938 35,313 21,250 254,375 64,375 58,125 26,563 213,438 96/02 68,125 27,188 21,563 202,188 64,375 48,438 26,563 168,750 96/03 68,875 20,000 21,875 254,375 62,188 30,000 26,875 220,313 96/11 71,250 11,250 21,563 274,063 67,188 17,500 27,188 229,375 96/12 70,625 22,188 20,938 280,000 66,563 34,688 26,563 227,500 97/01 74,688 20,000 20,938 275,625 71,250 28,125 26,250 220,000 97/02 74,063 16,875 20,938 208,438 68,438 26,250 26,250 156,875 97/03 75,625 19,375 19,688 275,938 70,313 25,625 25,938 220,625

a_{The analysis has been converted to US dollars. One US dollar is equivalent to roughly 32 New Taiwanese Dollars (NTDs).}

in departmental efficiency between the pre-PCP and PCP pe-riods. We merge the data of each month between different periods into one data set and calculate efficiency scores for the entire data set to control for monthly effect. For exam-ple, all observations are combined into one data set and rela-tive efficiency of each department in November 1995 and No-vember 1996 is calculated. We replicate the computation five times by varying different months between the two periods. The procedure is appropriate as revenue of the case hospital has monthly effects. For example, revenue was much smaller in February because of the Chinese New Year holidays. Be-sides, case mix in the hospital may be quite different in dif-ferent months due to seasonal factors. Although aggregating the five months data together could be an alternative and may help to average out the monthly fluctuation, we believe that performing the analysis for each month in the pre-PCP pe-riod and PCP pepe-riod separately enables us to better remove the monthly effects.

The initial data set included 102 profit centers in the pre-PCP period and 103 profit centers in the pre-PCP period. We delete any center that only serves administrative functions or its inputs or outputs have a value of zero. Because the DEA model requires a complete data set, departments not in the database for ten months are also eliminated. The procedure results in a sample of 58 profit centers (departments).

3.2. Selecting input and output variables

Based on prior literature and consulting executives in the case hospital, we identify the output and inputs that are relevant to the evaluation of efficiency in our study. In this paper, we focus on revenue efficiency, the ratio of observed revenue to maximum possible revenue, among profit centers in the case hospital. We do not use profit ratio to measure operational performance because profits are subject to distortion if indi-rect costs are allocated inappropriately. In addition, we do not use physical units (e.g., number of beds and inpatient days) of output and inputs as performance measures because some departments in the case hospital lack such measures. Total revenue is used as a proxy for output, measured by the

sum of inpatient revenue, outpatient revenue, and emergency care revenue. Considerations for quality of care and teaching are absent here because no well-accepted measures are avail-able. Thus, we implicitly assume the services provided by each department have the same quality outcomes. The out-put is assumed to be produced by three inout-puts: labor, meas-ured by personnel costs; material, measmeas-ured by medical sup-plies and medicine costs; capital, measured by buildings and equipment depreciation expenses. We do not include costs allocated from service departments as part of inputs because the cost accounting system used in the case hospital can be distorted. Specifically, many cost allocation bases employed were based on departmental revenue, not real cost drivers. That is, departments with higher revenues bear higher indi-rect costs, regardless of actual resource utilization. As the cost allocation rules work against departments with improved revenue, we believe that including service department costs in the evaluation of efficiency (i.e., construct the frontier) could be seriously biased.

The descriptive statistics of inputs and output in the DEA model are given in table 1. To adjust for the potential effect of reimbursement rates and price level changes, total revenue and medicine costs are deflated by consumer price index and salary costs are deflated by annual salary adjustment rate of the case hospital.

The results in table 1 indicate that average medicine costs of responsibility centers in the PCP period are much lower than those in the pre-PCP period. For example, medicine costs decreased from $32,1881in November 1995 to $11,250 in November 1996 (a 65% decrease). However, average per-sonnel and depreciation expenses of the responsibility centers are in the range of $65,938–$75,625 and $19,688–$21,875, respectively, during the ten months, showing the case hospi-tal may have little control on labor and capihospi-tal costs in the short run. An increase in average revenue of the responsibil-ity centers also indicates that physicians increase outputs in the PCP period (e.g., $251,563 in December 1995 increased 1_{The analysis has been converted to US dollars. One US dollar is equivalent}

to $280,000 in December 1996, an 11% increase). Also, stan-dard deviations of the inputs and output are quite large. Sev-eral factors may contribute to the result: (a) difference in size across departments, (b) difference in the nature of care provided (e.g., dental department uses very small amount of medicine while internal medicine department requires large amount of medicine), (c) the effect of the differences in NHI reimbursement rates.

3.3. Variables to be used in the Tobit model

In the Tobit model, the dependent variable is (φ− 1). Here,

φ denotes the efficiency score computed from the DEA model. The independent variables, which may explain dif-ferences in departmental efficiency, are discussed as follows.

3.3.1. Departmental profits (PRO)

PRO plays a vital role in influencing physicians’ resource al-location behavior because departmental profits directly link to physician compensation in the PCP. Thus, physicians should have incentives to select inputs that minimize costs and max-imize outputs. Such efforts should improve departmental ef-ficiency. However, an improvement in efficiency may not be identical with profits in our study. Note that departmen-tal profits (PRO) are defined as todepartmen-tal revenues minus direct costs (inputs) and indirect costs (costs allocated from ser-vice departments). In contrast, revenue efficiency measures how to use minimum mix of inputs for maximum revenue. Although the allocation of indirect costs is distorted, a part of physicians’ compensation is nevertheless based on depart-mental profits. The relationship between departdepart-mental profits and efficiency provides an opportunity for the case hospital to evaluate the appropriateness of the compensation scheme. Specifically, physicians may be better evaluated based on ef-ficiency not the distorted profits when the DEA-based effi-ciency measures and accounting profits show conflicting re-sults. However, for practical purposes, it is unlikely that the efficiency scores are used in the PCP in the case hospi-tal or elsewhere for performance evaluation purposes because DEA is too complex to be easily understood and implemented for personal monitoring by physicians, staff and supervisors. Thus, if efficiency scores and department profits (accounting numbers) move in the same direction in the Tobit analysis (i.e., profitability, although quite noisy, on average is consis-tent with efficiency), the poconsis-tential dysfunctional effects re-sulted from distorted indirect costs would be less severe.

3.3.2. Departmental size (PEO)

Studies have found that hospital size influences clinical effi-ciency, but the direction of the effect is mixed. For example, Zuckerman et al. [25] found that extending the size of hos-pital is useful to enhance efficiency; whereas Lo et al. [19] indicated that increasing the number of beds decreased effi-ciency of general hospitals in Taiwan. Prior literature has used the number of beds, revenue, and the number of employee as proxy for hospital size. Since the study is conducted at mental level, some of the profit centers (e.g., dental

depart-ment and some departdepart-ments which conduct medical tests) do not have beds. To avoid deleting too many departments with-out beds, we decide not to use the number of beds as proxy of size. In addition, departmental revenue has already been used as the output in our DEA model. As a result, in this paper we have no choice but to use the number of employees (PEO) as proxy for size of the department. The number of employees and labor costs (one of inputs) are likely positively correlated. As there is no alternative proxy for department size, the size proxy used in our study could have measurement errors.

3.3.3. Percentage of teaching physicians (TEA)

Because the case hospital is affiliated with a major national university, some physicians also serve as faculty in the Col-lege of Medicine. Ament et al. [1] suggested that physicians with teaching positions may focus more on teaching and re-search instead of cost control. They may spend less cost-containment efforts for the hardest-to-treat patients because they value research. Thus, departments with higher percent-age of non-teaching physicians could be more efficient. Here the percentage of teaching physicians (TEA) is measured by the proportion of physicians who also teach at the College of Medicine out of the total number of physicians in the depart-ment.

3.3.4. Physicians’ seniority (SEN)

Coughlin and Patel [13] found that over time physicians re-call more cues that are critical and can utilize more informa-tion that is relevant. Chilingerian [10] also found that there is a positive association between physician’s seniority and effi-cient utilization of clinical resources. Hence, we hypothesize that departments with physicians of higher average seniority become more efficient. As the case hospital has maintained a policy to recruit the best graduates from the medical school that it affiliates, almost all physicians in our sample have this single employer in their professional career. Thus, seniority of the physicians (SEN) can be measured by the their employ-ment time in the case hospital. Specifically, SEN is defined as years from the first date the physician served in the case hospital up to the first date of the month under study.

3.3.5. Percentage of physicians’ service time (TIME)

Some physicians in the case hospital serve in more than one department because of their specialty. Hence, we hypothesize that departments with higher average percentage of physician service time are more efficient than those departments with lower average percentage. Percentage of physicians’ service time (TIME) is computed by proportion of time the physician served in the department in the month under study. For ex-ample, if the physician spent the same amount of time in two departments, the percentage of the physician’s service time is 50% in each of the two departments.

3.3.6. Number of medical orders (ORD)

In this paper, we use the number of medical orders as a proxy of complexity of medical services in the department. Prior studies have found a negative association between the

num-Table 2

Descriptive statistics of explanatory variables in the Tobit model.a

Month Mean Standard deviation

PRO PEO TEA SEN TIME ORD PRO PEO TEA SEN TIME ORD

95/11 53,349 30 0.64 15.18 0.62 315 135,703 13 0.25 5.44 0.31 245 95/12 35,766 28 0.64 15.58 0.62 319 123,689 11 0.25 7.33 0.31 236 96/01 33,843 28 0.64 14.82 0.62 321 115,516 11 0.25 4.64 0.31 251 96/02 −7,934 29 0.64 14.54 0.65 328 98,474 23 0.25 4.89 0.27 245 96/03 39,609 31 0.64 15.82 0.61 332 122,313 13 0.25 7.36 0.31 246 96/11 36,003 38 0.64 15.50 0.64 337 88,086 28 0.24 7.33 0.28 257 96/12 40,324 37 0.63 15.07 0.67 343 102,007 27 0.23 5.05 0.28 264 97/01 20,864 39 0.62 15.67 0.64 361 93,874 28 0.25 7.34 0.28 243 97/02 −10,965 38 0.62 15.74 0.62 367 85,611 28 0.25 7.35 0.34 246 97/03 67,528 39 0.60 14.53 0.66 382 114,688 28 0.23 4.80 0.27 259

a_{PRO: departmental profits. PRO has been converted to US dollars. One US dollar is equivalent to roughly 32 New Taiwanese Dollars (NTDs); PEO:}

number of employees; TEA: the proportion of teaching physicians out of total physicians of the department; SEN: the average of physicians’ seniority in the department; TIME: the average percentage of physicians’ service time in the department; ORD: the number of medical orders provided by the department. ber of DRGs treated and the efficient utilization of clinical

re-sources [10]. A large number of medical orders provided by a department may result in service complexity and influence physicians’ ability to process medical information efficiently.

3.3.7. Period variable (D)

As we expect differences in revenue efficiency exist between the pre-PCP (November 1995 to March 1996) and the PCP periods (November 1996 to March 1997), we will use a dummy variable D to differentiate the two periods. Here, D = 1 if observation is in the PCP period, and = 0 other-wise.

The above discussion can be summarized in the following Tobit model:

(φ− 1) = α0+ α1PRO+ α2PEO+ α3TEA+ α4SEN + α5TIME+ α6ORD+ α7D+ ε, (2) where

φ the efficiency score is computed from equation (1) un-der appropriate returns to scale assumptions.

PRO departmental profits.

PEO the number of employees (including physicians, nurses and ancillary labors).

TEA the percentage of teaching physicians.

SEN the average of physicians’ seniority in the department. TIME the average percentage of physicians’ service time in

the department.

ORD the numbers of medical orders provided by the depart-ment.

D the PCP period vs. the pre-PCP period.

ε error term.

Descriptive statistics of the explanatory variables are pre-sented in table 2. We find a very clear monthly effect that profits in February were negative in the pre-PCP and PCP periods (the effect of Chinese New Year holidays). In ad-dition, many physicians taught at the College of Medicine (about 60%) and on average served in the case hospital for over 15 years. Thus, physicians in the case hospital were heavily involved in teaching and had a very stable turnover rate.

Table 3

The mean efficiency scores for the CRS and VRS models.a

Comparison setsb CRS model VRS model

Pre-PCP PCP Pre-PCP PCP C1 7.47 5.88 5.38 4.18 C2 5.41 3.40 3.98 2.50 C3 6.75 2.87 5.39 2.08 C4 5.60 3.23 3.97 2.31 C5 3.91 2.94 2.92 2.13

a_{CRS= constant returns to scale, VRS = variable returns to scale.} b_{C1 denotes period November 1995 and November 1996; C2 – period}

De-cember 1995 and DeDe-cember 1996; C3 – period January 1996 and January 1997; C4 – period February 1996 and February 1997; C5 – period March 1996 and March 1997.

4. Results

In the DEA analysis, the implication of efficiency will be dif-ferent under difdif-ferent returns to scale assumptions. Thus, we will first test which returns to scale assumptions best match our data. Then, under the specific returns to scale assumption, we test for differences in efficiency between the two periods. Finally, we use the tobit model to better control for factors that might explain the observed differences in efficiency in addition to the PCP.

The mean efficiency scores for the constant returns to scale (CRS) and variable returns to scale (VRS) models are pre-sented in table 3. As can be seen clearly in table 3, in the PCP period the mean efficiency scores are lower (i.e., more efficient) than those in the pre-PCP period regardless of the scale of return assumptions. Although the result provides an intuitive check for the efficiency effect of the PCP, we con-duct the following formal tests to obtain a statistically sound conclusion.

4.1. Tests of returns to scale

Because different DEA models provide estimators for tech-nical efficiency under different assumptions about returns to scale, in this study we use a DEA-based statistical test to

Table 4

Tests of returns to scale assumption.a Efficiency deviation Null Alternative Fstatisticsb distribution hypothesis hypothesis C1 C2 C3 C4 C5 Half-normal CRS VRS 1.37c1.71d1.31c1.62d1.82e Exponential CRS VRS 1.5e 1.52e1.39d1.6e 1.7e

a_{Banker (1993) suggested statistical tests for inefficiency differences}

be-tween two groups of decision-making units (DMUs). The test statistics assume an exponential or a half-normal distribution of inefficiency scores. CRS= constant returns to scale; VRS = variable returns to scale.

b_{C1 denotes period November 1995 and November 1996; C2 – period}

De-cember 1995 and DeDe-cember 1996; C3 – period January 1996 and January 1997; C4 – period February 1996 and February 1997; C5 – period March 1996 and March 1997.

c_{F-statistics indicates significant at 10% level.} d_{F-statistics indicates significant at 5% level.} e_{F-statistics indicates significant at 1% level.}

examine the returns to scale assumption. The test has been used to evaluate returns to scale for software development projects [5] and software maintenance projects [4]. In ta-ble 4, the null hypothesis of constant returns to scale is re-jected under exponential or half-normal distribution in both the pre-PCP and PCP periods. Thus efficiency scores in this paper are computed under the assumption of variable returns to scale.

4.2. Results of the efficiency effects of implementing the PCP

We follow Banker et al. [7] to test whether departmental effi-ciency of the case hospital improved in the PCP period. The null hypothesis is that there is no difference in average depart-mental efficiency between the pre-PCP and PCP periods. The alternative hypothesis is that departments are on average less efficient in the pre-PCP period when compared with the PCP period. The results in table 5 show that the null hypothesis under the assumptions of half-normal and exponential distrib-utions is rejected in all comparison sets. The finding is consis-tent with our expectation that the PCP enhanced departmental efficiency in the case hospital.

Note that efficiency scores are a performance index. They measure the performance of each department relative to the performance of other departments in the case hospital. Most of prior studies employed data on inputs and outputs to gen-erate a vector of efficiency scores. They used the results to examine whether the decision-making units under investiga-tion are efficient or not. However, in this study we investigate the efficiency effects of the PCP across two periods. As a re-sult, we will focus on test scores (t-statistics) instead of the size of scores to examine if the PCP resulted in an efficiency improvement.

4.3. The Tobit regression

Because departmental efficiency may be influenced by factors related to departmental characteristics and physicians’ pro-files in addition to the effect of PCP, the tobit model is used to better control for the effects of those factors. In our

to-Table 5

Results of Banker’s DEA-based tests for efficiency differences between the two periods.a

Comparison sets F-test statisticsb

Exponential Half-normal C1 1.37c 2.33d C2 1.98d 12.55d C3 4.05d 72.40d C4 2.28d 14.67d C5 1.52c 2.70d

a_{Banker (1993) suggested statistical tests for inefficiency differences}

be-tween two groups of decision-making units (DMUs). The test statistics assume an exponential or a half-normal distribution of inefficiency scores.

b_{C1 denotes period November 1995 and November 1996; C2 – period}

De-cember 1995 and DeDe-cember 1996; C3 – period January 1996 and January 1997; C4 – period February 1996 and February 1997; C5 – period March 1996 and March 1997.

c_{F-statistics indicates significant at 10% level.} d_{F-statistics indicates significant at 1% level.}

bit model, the hypothesis of homoscedasticity is rejected (re-sults of statistical tests are available upon request from the au-thors). To avoid inconsistent estimates, in this subsection we only consider results in the heteroscedastic tobit model. The estimates are presented in table 6. Note that because higher efficiency scores (i.e., the dependent variable) mean a lack of efficiency, we should interpret a positive sign in coefficients as inefficiency, a negative sign as higher levels of efficiency.

Several factors are associated with efficiency. First, the av-erage of physicians’ seniority (SEN) has a significant positive effect on efficiency in all comparison months. The result indi-cates that there is a positive association between physician’s seniority and a more efficient utilization of clinical resources. Although seniority increases salary and may have an adverse effect on profit, we find a negative correlation (r = −0.38) between the seniority of physicians and personnel costs. The result indicates a possible substitution between departmental physician labor costs and other personnel costs (nurses and ancillary labors costs). In addition, active senior physicians may attract more patients and thus generate higher revenue for their department, as evident by a positive relationship (r = 0.24) between the seniority of physicians and depart-mental revenue.

Second, the average percentage of physicians’ service time (TIME) shows a significant positive effect in all comparison months. This means that departments with more physicians who concentrate their efforts in that department are more efficient. An in-depth interview with several senior physi-cian administrators in the case hospital indicated that direc-tors of many responsibility centers attempted to attract more full-time physicians with expertise matched with their depart-ments in order to build a better sense of team and better mix of skills. TIME could be a measure concerning the success of such efforts.

Third, the result that the effect of PCP period (D) vari-able on efficiency is significant and positive for all com-parison sets confirms our expectation that average efficiency in the PCP period is significantly higher even after

con-Table 6

Results of heteroscedastic Tobit model for factors that might explain the inefficiency.

Explanatory Coefficient and t-statisticsb

variablesa C1 C2 C3 C4 C5 Constant 1.73 8.43 2.22 2.52 7.30 (0.54) (4.21) (1.23) (1.26) (3.63) PRO −1.50 −4.48 −0.44 −0.97 −3.52 (−2.58)c (−1.39) (−0.65) (−1.43) (−3.11)c PEO 1.14 1.98 2.14 3.06 1.64 (0.86) (2.47)d (2.19)d (3.26)c (0.80) TEA −3.00 −5.43 −0.98 −1.95 1.60 (−0.98) (−0.67) (−0.73) (−1.22) (0.51) SEN −0.93 −0.40 −0.91 −0.49 −0.19 (−2.56)d (−2.25)d (−2.72)c (−2.63)c (−2.73)c TIME −16.18 −10.30 −18.77 −11.03 −4.32 (−2.39)d (−3.23)c (−3.06)c (−3.29)c (−3.49)c ORD 0.27 0.87 0.75 −0.86 −0.36 (1.28) (1.46) (0.09) (−0.69) (−1.69)e D −1.51 −1.65 −1.92 −1.32 −2.79 (−2.47)d (−2.52)d (−3.50)c (−2.42)d (−3.37)c Chi-squaref 27.72c 31.49c 30.50c 31.42c 30.45c

a_{PRO: departmental profits; PEO: number of employees, TEA: the}

pro-portion of teaching physicians out of total physicians of the department; SEN: the average of physicians’ seniority in the department; TIME: the average percentage of physicians’ service time in the department; ORD: the number of medical orders provided by the department; D: the dummy variable for time period. Here, D= 0 indicates the pre-PCP period; D = 1 indicates the PCP period.

b_{C1 denotes the period November 1995 and November 1996; C2 – the}

period December 1995 and December 1996; C3 – the period January 1996 and January 1997; C4 – the period February 1996 and February 1997; C5 – period March 1996 and March 1997.

c_{Significance at 1% level.} d_{Significance at 5% level.} e_{Significance at 10% level.}

f_{Chi-square is based on a likelihood ratio test which tests the joint}

signif-icance of the independent variables [10]. This statistic is calculated by −2 log LR, where log LR is the difference between the maximized value of the likelihood function for the full model and the maximized value if all coefficients except the intercept are zero. The statistic tests the signifi-cance of the Tobit model and is similar to an F -test in standard regression.

trolling for other factors that could affect departmental ef-ficiency. Finally, we find relatively weak evidence con-cerning the efficiency implications of department profit and size. For example, profits (PRO) shows a significant pos-itive effect on efficiency only in C1 and C5. That is, the PCP that provides departmental members with monetary in-centives seems to also enhance efficiency. In addition, the number of employees (PEO) has a significant negative ef-fect on efficiency in three comparison sets (e.g., C2, C3, and C4), i.e., the smaller the department, the better its ef-ficiency. Although the signs of PRO and PEO are consis-tent across C1 and C5, a lack of uniform statistical signif-icance among different months indicate that the variables were less robust. Several reasons may be related to the re-sult. For example, there is considerable noise concerning the measure of departmental profits and size, as discussed earlier. In addition, case mix could be different for

differ-ent months between the Pre-PCP and PCP periods and our month-by-month analysis may not be able to remove such an effect.

4.4. Sensitivity tests

To ensure the robustness of our results, the following sensi-tivity tests are conducted.

4.4.1. Differences in production technology among departments

In prior analysis, we did not control for differences in produc-tion technology among different departments. To remedy the deficiency, we assume differences exist in production tech-nology between surgery departments and other departments. We use an additional dummy variable D1 (D1 equals 1, if the observation is from a surgery department, and 0 otherwise) to differentiate surgery departments and other departments in the tobit analysis. The results (not reported here) indicate that (1) surgery departments are less efficient than other depart-ments in C2, C3 and C4; they may be more complex in clin-ical procedures or subject to more severe distortion in reim-bursement. (2) The number of employees becomes insignifi-cant in all comparison sets. (3) Average efficiency in the PCP period is still significantly higher even after controlling for the difference in production technology among departments.

In addition to the distinction of surgical vs. non-surgical departments, we conducted in-depth interviews with five se-nior administrators (four of them were physicians) in the case hospital to clarify how the diversity of technology and rev-enue structure across departments may affect revrev-enue/profit generation capability. We gathered their expert opinions by asking them to rank the 58 sample departments based on rel-ative ease of generating revenues or profits using a five-point Likert scale. We conducted further tests in the tobit model based on the ranking. For example, if a department was viewed as relatively more difficult to generate revenue/profit resulting from differences in NHI reimbursement rates or physicians’ flexibility in controlling costs, a dummy variable equals 1; otherwise, the dummy variable equals 0. We in-cluded the additional dummy variables in the tobit model and found them not statistically significant in all periods (i.e., C1– C5), nor did they affected the sign or significance level of other variables.

4.4.2. Aggregating the observations over the five months

We performed the analyses on each month separately to con-trol for monthly effects. Aggregating the observations over our sample period can average out the noise in the monthly values. Thus, we also analyze the aggregating data. The re-sults (available from the authors upon request) indicate that Banker’s DEA-based tests [3] still strongly show an efficiency gain in the PCP period. However, the Tobit model yields a weaker result. For example, department size (PEO) is no longer associated with efficiency; the seniority of physicians (SEN) and the effect of PCP (D) are statistically significant at 10% level.

4.4.3. The role of indirect costs

We did not include indirect costs allocated from service de-partments as part of inputs because such costs were subject to distortion. As mentioned earlier, many service costs were al-located based on departmental revenue, not cost drivers. The practice works against departments with improved revenues. Because the percentage of indirect costs out of the total costs was large for the case hospital (about 49% in the pre-PCP pe-riod and 47% in the PCP pepe-riod), in the sensitivity analysis we also include costs allocated from service departments as another input in the DEA model. The findings (not reported here) are consistent with our earlier conclusion that the PCP improved departmental efficiency in the case hospital; how-ever, as we expected, the result are generally weaker in terms of statistical significance.

5. Discussion

Staff-model HMOs rely heavily on a salary-based method of payment. Thus, experiences from the case hospital, where physicians are traditionally salaried employees, are useful to better design the hospital–physician integration strategy. For example, the case hospital found that a flat salary or a salary based on seniority and rank does not provide the right incen-tives for physicians to expand services and control costs un-der a reimbursement scheme which is mainly based on fee-for-service but with small portion of cases paid under DRGs. More subtle incentive schemes for physicians are necessary to promote efficiency. In this paper, we provide evidence that the PCP, which is based on individual as well as team (i.e., re-sponsibility centers system) incentives, did induce physicians to enhance efficiency in the case hospital. In addition, het-eroscedastic tobit regression is used to control for the effects of departmental characteristics and physicians’ profiles. The results indicate that average efficiency in the PCP period is significantly higher than that in the pre-PCP period after con-trolling for other factors that may also influence efficiency. We also find that the average of physicians’ seniority and per-centage of physicians’ service time in the department are pos-itively associated with efficiency. Besides, departments with higher profits and fewer numbers of employees are associated with higher efficiency.

We find that PCP is not only associated with an increase in overall hospital revenue but also a shift in revenue struc-ture. Specifically, outpatient revenue of the case hospital in-creased from 34.7 to 41.1% of total revenue during the sam-ple period. Although the ratio was well below the national average of hospitals (about 65.2%), during the study period its increase rate (6.4%) was much higher than the national average rate (about 1.2%). Two possible reasons may be re-lated to the phenomenon. First, the reimbursement rates for outpatient services had been more favorable when compared with inpatient services in Taiwan. Second, inpatient services were more likely constrained by the number of beds which were inflexible to expand; by contrast, it was easier to expand outpatient services simply by hiring more doctors or inducing

doctors to see more patients (e.g., extending hours of outpa-tient services or cutting time with each paoutpa-tient). We believe that the PCP, with a strong focus of increasing revenue, may be related to such behavior.

Relative to the implementation of NHI, the results indicate that efficiency increased in all months of the PCP period when compared with that of the pre-PCP period. Efforts to improve efficiency in Taiwanese hospitals had been underway well be-fore the change in policy and thus this study was unable to detect any NHI effect. What is instructive is that such a mas-sive change in policy may not have immediate and widespread effects in particular hospitals.

Several limitations exist in the paper. First, the sample de-partments differ in technology and complexity. As a result, their revenue or profit generation capability may be quite dif-ferent. Our sensitivity tests based on surgical/non-surgical distinction or a survey of hospital administrators may not ap-propriately control for all potential differences. Second, be-cause of data availability we only use the 5 months before the implementation of the PCP and the 5 months after the im-plementation of the PCP to test its initial effects in the study. A longer period of comparison is certainly desirable besides the initial effects that we examined. Third, the study is un-able to distinguish whether the efficiency gain came from the effects of PCP or an advance in medical technology dur-ing the sample period, as both would cause a shift in the revenue frontier of the observed departments. Fourth, the study fails to measure the quality of output. Future research should consider the effects of disease severity as well as qual-ity of teaching and research on the productivqual-ity of hospital resource. Fifth, we only consider financial data, future re-search may consider additional variables (e.g., physicians’ concern for quality, departmental management style, etc.) to enrich our understanding of the hospital–physician integra-tion strategies. Finally, the PCP is not the only strategy that the case hospital adopted to influence physicians’ behavior (e.g., TQM can be effective as well). Future research may want to account for the simultaneous existence of multiple management control systems in hospitals. We believe that cross-hospitals research design will be a more appropriate ap-proach to address the issue.

Acknowledgements

The authors acknowledge helpful comments and suggestions provided by two anonymous reviewers. The second author appreciates the research grants from the National Science Council (NSC87-2416-H-002-036) and the National Taiwan University Hospital (NTU82-02) in Taiwan.

References

[1] R.P. Ament, E.J. Kobrinski and W.R. Wood, Case mix complexity dif-ferences between teaching and nonteaching hospitals, Journal of Med-ical Education 56 (1981) 894–903.

[2] R.D. Banker, Hypothesis tests using data envelopment analysis, The Journal of Productivity Analysis 7 (1996) 139–159.

[3] R.D. Banker, Maximum likelihood, consistency and data envelopment analysis: A statistical foundation, Management Science 39 (1993) 1265–1273.

[4] R.D. Banker and S.A. Slaughter, A field study of scale economies in software maintenance, Management Science 43 (1997) 1709–1725. [5] R.D. Banker, H.H. Chang and C.F. Kemerer, Evidence on economies of

scale in software development, Information and Software Technology 36 (1994) 275–282.

[6] R.D. Banker, A. Charnes and W.W. Cooper, Some models for estimat-ing technical and scale inefficiency in data envelopment analysis, Man-agement Science 30(9) (1984) 1078–1092.

[7] R.D. Banker, R.J. Kauffman and R.C. Morey, Measuring gains in op-erational efficiency from information technology: A study of the Posi-tran development at Hardee’s Inc., Journal of Management Information Systems 7 (1990) 29–54.

[8] K. Chakraborty, B. Biswas and W.C. Lewis, Measurement of technical efficiency in public education: A stochastic and nonstochastic produc-tion funcproduc-tion approach, Southern Economic Journal 67 (2001) 889– 905.

[9] H.H. Chang, Determinants of hospital efficiency: The case of central government-owned hospitals in Taiwan, Omega-International Journal of Management Science 26(2) (1998) 307–317.

[10] J.A. Chilingerian, Evaluating physician efficiency in hospitals: A mul-tivariate analysis of best practices, European Journal of Operational Re-search 80 (1995) 548–574.

[11] T.N. Chirikos and A.M. Sear, Measuring hospital efficiency: A compar-ison of two approaches, Health Services Research 34(6) (2000) 1389– 1408.

[12] T.D. Coelli, S.P. Rao and G.E. Battese, An Introduction to Efficiency

and Productivity Analysis (Kluwer Academic Publishers, 1998).

[13] D. Coughlin and V. Patel, Processing of critical information by physi-cians and medical students, Journal of Medical Education 62 (1987) 818–829.

[14] R.S. Fare, S. Grosskopf and C.A.K. Lovell, Production Frontier (Cam-bridge University Press, New York, 1994).

[15] H.O. Fried, C.A.K. Lovell and S.S. Schmidt, eds., The Measurement of

Productive Efficiency: Techniques and Applications (Oxford University

Press, New York, 1993)

[16] J.B. Goes and C.L. Zhan, The effects of hospital–physician integration strategies on hospital financial performance, Health Services Research 30 (1995) 507–530.

[17] W.H. Greene, Econometric Analysis (Macmillan Publishing Company, 1997).

[18] L. Hjalmarsson, S.C. Kumbhakar and A. Heshmati, DEA, DFA and SFA: A comparison, The Journal of Productivity Analysis 7 (1996) 303–327.

[19] J.C. Lo, K.S. Shih and K.L. Chen, Technical efficiency of the general hospitals in Taiwan – an application of DEA, Economic (Institute of Economics, Academia Sinica, Taiwan, ROC) 24(3) (1996) 375–396. [20] G.S. Maddla, Limited-Dependent and Qualitative Variables in

Econo-metrics (Cambridge University Press, 1983).

[21] T.A. McCarty and S. Yaisawarng, Technical efficiency in New Jer-sey School Districts, in: The Measurement of Productive Efficiency:

Techniques and Applications, eds. H.O. Fried, C.A.K. Lovell and S.S.

Schmidt (Oxford University Press, New York, 1993).

[22] R.E. Santerre and S.P. Neun, Health Economics: Theory, Insights, and

Industry Studies (Irwin, Chicago, 1996).

[23] G.R. Wilensky and L.F. Rossiter, A reexamination of the use of physi-cian services: The role of physiphysi-cian-initiated demand, Inquiry 20 (1983) 162–172.

[24] G.R. Wilensky and L.F. Rossiter, The relative importance of physician-induced demand in the demand for medical care, Milbank Memorial Fund Quarterly 61(2) (1983) 252–271.

[25] S. Zuckerman, J. Hadley and L. Iezzoni, Measuring hospital efficiency with frontier cost functions, Journal of Health Economics 13 (1994) 255–280.