Objective: To assess the responses to financial incentives after a change in the payment system in a capitation-style healthcare payment system over a 5-year period.
Study Design: Cross-sectional and longitudinal examination of cost, utilization, and diagnostic data.
Methods: Using Veterans Health Administration (VHA) administrative data on healthcare users between fiscal years 1998 and 2002, we calculated the proportion of new patients entering each of the payment classes, the illness burden of patients entering the payment classes, and the profitability index (a ratio of payment to costs) for each class suspected of gaming and each control class. Our main dependent variables of interest were the differences in the measures between each utilization-based class and each diagnosis-based class. We used 2 different analytic approaches to assess whether these differences increased or decreased over time.
Results: No clear evidence of gaming behavior was present in our results. A few comparisons were significant, but they did not show a consistent pattern of responses to incentives. For example, 6 of 16 comparisons of profitability index were significant, but (contrary to the hypothesis) 4 of these had a negative value for the time parameter, indicating decreasing profitability in the utilization-based classes versus the diagnosis-based classes.
Conclusions: Although the payment system could be manipulated to increase payment to VHA networks, no such consistent gaming behavior was observed. More research is needed to better understand the effects of financial incentives in other healthcare payment systems.
(Am J Manag Care. 2007;13:513-522)
Over a 5-year period, we evaluated a capitation-style healthcare systemâ€™s response to financial incentives by comparing characteristics of patients categorized to utilization-based and diagnosis-based payment classes.
We found no clear evidence that the system was being manipulated to increase healthcare payments.
Further research is needed to assess the impact of financial incentives on the delivery of healthcare.
Financial incentives within healthcare payment systems are increasingly recognized as powerful drivers of hospital service use and may dictate the healthcare services provided to patients, regardless of clinical need.1-4 Because of the direct relationship between the financing of healthcare services and the utilization of such services, changes in the payment system (eg, the institution of diagnosis-related groups for Medicare prospective payment5) can dramatically transform clinical and administrative practices.6,7 For example, Kahn et al concluded from their comparison of Medicare hospitalizations between 1981-1982 and 1985-1986 that although overall patient outcomes (eg, 30-day mortality, in-hospital mortality) were not negatively affected by the change to a prospective system, patients were more likely to be discharged prematurely.8,9 Chan et al showed that the economic incentives in Medicare's system of payment for rehabilitation hospitals encouraged a dramatic increase in length of stay during the base year.10 This is because payments were capped at the base year level, with incentives given if charges per patient were reduced in subsequent years.
Although more is known about the response of individuals11-13 or hospitals8- 10,14 to payment system incentives, less is known about responses of a healthcare delivery system to incentives in the payment system. Thus, our objective was to study the response of a large federal healthcare system to incentives to increase payments to a network (consisting of multiple integrated hospitals and outpatient clinics in a geographical area). We studied the Veterans Health Administration (VHA). The VHA is the largest integrated healthcare system in the United States, with a medical care and research budget of $32.52 billion in fiscal year (FY) 2005.15 The VHA healthcare system cares for its enrollees in Department of Veterans Affairs (VA) Medical Centers and VA Community Based Outpatient Clinics, which employ salaried physicians. The VHA budget is appropriated by Congress as a global budget and distributed to its 21 VHA networks via a form of capitation. Capitation payments are 1 of 2 flat payments per enrolled veteran meeting vesting criteria, and are paid whether the veteran uses the services or not. It is important to note that in the VHA the networks all compete for a fixed budget. Thus, if 1 network is able to increase its budget relative to other networks, those other networks will receive a smaller share of the total budget. There are only 2 ways to increase a network share of this budget. One way is to increase the number of patients cared for; a potentially more effective way to increase this share is to ensure that patients that are already cared for in the system qualify for the higher of the 2 capitation payments.
The incentives to engage in this second behavior (termed gaming) arise when the difference between the high and low payment classes is larger than the marginal cost of moving a patient over the utilization threshold (ie, when the payment/cost difference is larger for the higher payment class). Thus, there may be an incentive to move patients into those classes when this action may not be completely justified by clinical need. Because the VHA changed its method of payment to networks in 1997, the opportunity arose for a natural experiment to examine whether the presence of financial incentives affected how patients were classified. We studied payment classes for which a response to incentives was possible and might be financially advantageous, and compared them with payment classes for which there was less discretion in assigning patients to these categories (ie, gaming behavior would be less likely to occur). This work is important because in a healthcare system with a fixed global budget, efforts to increase payment for patients in this way may result in overuse of unnecessary healthcare services, effectively restricting access for others who need such services. If all units of a healthcare system are competing for a fixed budget, the incentive for networks and facilities in a globally budgeted healthcare system is to overprovide certain services without considering costs to the group. That may result in unequal distribution of services over time in the face of equivalent patient need. For example, patients needing care later in the budget cycle are competing for the residual of the fixed budget.16
Because the VHA pays healthcare providers via salary, providers do not experience incentives that are aligned with the interests of the network. In this situation, the ways that networks or hospitals can exert influence on the healthcare providers who deliver care include structural changes in service availability (eg, expanding services and hiring staff for lucrative clinical programs, downsizing less lucrative ones) and/or increasing efforts to recruit certain types of patients through outreach or screening activities (a version of cream skimming or preferred risk selection).17
We are not aware of studies evaluating responses to financial incentives in healthcare systems where payment to physicians and administrators is via salary and global budgeting. If salaried or globally budgeted healthcare delivery systems exhibit responses to incentives noted in other settings, this would suggest that even more attention should be paid to aligning incentives for improved healthcare than is currently being paid in the healthcare policy arena18,19 through pay-for-performance and other types of programs.11 Better understanding of responses to financial incentives is important, both to prevent unintended adverse effects and because using healthcare payment systems to reward quality or meet other societal goals has tremendous untapped potential.20
The measures we selected for this study follow from our hypotheses regarding the effects of gaming. We hypothesized that in profitable patient payment classes that are defined by meeting certain thresholds for health services utilization (such as bed days of care [BDOC]), the proportion and the profitability of patients in these classes would increase over time, and the illness burden of new patients would decrease over time as patients who are less ill enter the class. These changes would occur as utilization of services was increased for patients who were near meeting the utilization threshold necessary for higher payment. In concurrent control classes, which were not defined by utilization, no such behavior would occur. Because we expected that the influence the networks or hospitals can exert on the healthcare providers to effect changes in capitation classification are structural or involve recruitment of patients, we did not expect that such changes would be evident immediately after the VHA changed its payment system in 1997, but would show gradual responses over time. Therefore, we studied several years of data to assess the presence of the hypothesized response.
VHA Payment SystemThe Veterans Equitable Resource Allocation (VERA) system was implemented in April 1997 to allocate the congressionally appropriated VHA medical care budget to the VHA networks. VERA assigns patients to 1 of 45 (as of FY 2002) patient class categories. Assignment to a patient class category may be based strictly on diagnosis (eg, stroke class), a combination of utilization and diagnosis criteria (eg, BDOC for chronic posttraumatic stress disorder [PTSD]), or strictly utilization (eg, home-based primary care [HBPC]). The patient classes are designated as being either VERA basic or VERA complex. Patients in VERA basic classes (eg, ear, nose, and throat) are accessing the healthcare system for routine healthcare and do not require use of VHA special care programs, whereas patients in VERA complex classes (eg, traumatic brain injury [TBI]) utilize special care programs and may require long periods of inpatient care and rehabilitation. Each year a basic and a complex "price" is calculated.
For example, in FY 1998 the VERA basic patient allocation was $2604, and the VERA complex patient allocation was $36 960. Because of this wide disparity in payments, although only 4% of total VHA users are complex care patients, they account for more than one third of the allocated monies.
During the years of study, the majority of VHA healthcare funds were allocated by utilization criteria. For example, criteria for patients with PTSD included the requirement of at least 60 BDOC. If a PTSD patient had 59 BDOC or less, the patient was in the basic care price group ($2604). If the patient had 60 BDOC, they were in the complex care group ($36 960). This type of payment differential may provide incentives to increase certain utilization parameters to move patients from the basic care to the complex care price (regardless of need), particularly when the ratio of price paid to actual cost favors such action.
This study was approved by the Baylor College of Medicine institutional review board.
We used a retrospective matched pair design to determine the behavior over time of each of the 3 study measures (proportion of new users, network-level PI, and RRS) for each of the utilization-based classes compared with the diagnosis-based classes. For each network, each utilization-based group and diagnosis-based group combination constituted a matched pair. For each of our study measures, our primary interest was whether the difference between the classes in a pair (eg, utilization-based network-level PI minus diagnosisbased network-level PI) increased or decreased over time. For each network, the difference between the classes among patients in that network was calculated. Using the difference ensured that our results would not be affected by systemic changes over the entire VHA patient population (eg, changes in illness burden as measured by RRS). We hypothesized that the diagnostic cost group RRSs for the utilizationbased classes would decrease over time relative to those for the diagnosis-based classes. Similarly, we predicted that both network-level PI and overall proportion of patients in a given class would increase in the utilization-based classes compared with the diagnosis-based classes. An increase in network-level PI would occur if patients entering that class had increasingly lower costs over time, indicating a lower illness burden.
We selected 4 utilization-based classes (PTSD, HBPC, domiciliary, and substance abuse) and 4 diagnosis-based classes (SCI paraplegic—old injury, SCI quadriplegic—old injury, stroke, and TBI) for a total of 16 utilization/diagnosis pairs. We evaluated a total of 48 comparisons, 16 pairs and 3 outcome measures per pair. Accordingly, we used 0.05/48 = 0.00104 as the significance level in all tests. For each FY and each comparison, there were 21 data points (one for each VA network) with the following exceptions: (1) for the domiciliary analyses, we removed 2 of the 21 networks from the study years (FYs 1998-2001) because they had fewer than 5 patients in the payment class; (2) for the substance abuse analyses, we removed 1 VHA network because of low patient count across the study years examined.
The analyses were done at the network level, and the dependent variable was the difference between the utilization-based and diagnosis-based classes in the study measure. Because the repeated measures at a network over time were expected to be correlated, we used a repeated measures analysis with unstructured covariance. The only fixed effects were the intercept and time. The sign and significance of the time parameter indicated whether the difference variable was changing over time.
Because the significance of parameter estimates in the repeated measures analysis requires an assumption of normality, we used a nonparametric analysis to check the results. A derived variable analysis, also known as a summary statistic analysis, 27 was used for this purpose. The derived variable is a generated variable that characterizes the time history of the difference between the utilization-based and diagnosis-based variables at each network.27-29 The derived variable used here was the ordinary least squares (OLS) estimated slope.30 For each network, the OLS slope is the coefficient of time in an OLS model in which time is the single independent variable, and the difference (utilization-based class — diagnosis-based class) in the study measure (eg, RRS) is the dependent variable. The OLS slope provides an indication of whether the values are increasing or decreasing over time and the rate of increase or decrease.
For each outcome measure (difference between the utilization-based class and the diagnosis-based class) and each network, the OLS slope was computed. For each measure, the mean of the OLS slopes over all networks was computed. A Wilcoxon signed rank test was used to determine whether the mean was significantly different from 0.30. The derived analysis is entirely nonparametric.
Between FY 1998 and FY 2002, the total number of VHA patients funded by the VHA payment system increased from 3 195 795 to 3 780 217 patients, an increase of 18.3%. Table 2 displays the characteristics of the new patients in each of the classes we examined. The number of new patients across the classes we studied ranged from 31 897 in FY 1998 to 22 16 in FY 2002, representing a very small percentage of the overall cohort, 1.0% and 0.6%, respectively. Again, although the total proportion of patients is relatively small, the budget impact of these patients is large. Specifically, all patients in these complex payment classes accounted for on average 19% of the healthcare budget for the 5 study years we examined.
Reading horizontally across the tables reveals that the 2 tests show consistent results with a single exception (the HBPC vs SCI paraplegic—old injury RRS comparison). A few of the comparisons were significant, but they do not show a consistent pattern of responses to incentives. For example, 6 of 16 comparisons of PI were significant. However, 4 of these had negative values for the time parameter, indicating decreasing profitability in the utilization-based class relative to the diagnosis-based class. This result was contrary to our hypothesis. Several of the differences in proportion of new users were significant, but some were negative. The only comparison pair that indicated possible gaming across all 3 variables was HBPC and stroke. Because HBPC did not exhibit the same pattern compared with other control classes, this may have been a random occurrence.
We studied changes in the way patients were classified and coded over a 5-year period to assess effects of incentives at the payment system level in a healthcare system with a global budget. To control for other factors that were present during the study period, we examined how variables that might indicate gaming moved in the utilization-based classes relative to the diagnosis-based classes. We found that although it was profitable and possible to change the classification of patients by utilization criteria, we found little evidence for such gaming behavior.
It is important to note that in the VHA the networks all compete for a fixed budget. Thus, if 1 network is able to increase its budget relative to other networks, those other networks will receive a smaller share of the total budget. There are only 2 ways to increase a network share of this budget. One way is to increase the number of patients cared for; a potentially more effective way to increase this share is to ensure that patients that are already cared for in the system qualify for the higher of the 2 capitation payments. Such gaming has been shown in other studies. Chan et al showed that the economic incentives in Medicare's system of payment for rehabilitation hospitals encouraged a dramatic increase in length of stay during the base year.10 This is because payments were capped at the base-year level, with incentives given if charges per patient were reduced in subsequent years. Kahn et al concluded from their comparison of Medicare hospitalizations between 1981-1982 and 1985-1986 that patients were more likely to be discharged prematurely after the implementation of the prospective payment system.8,9
Our findings may not be surprising. First, the hospitals we studied were federal hospitals. Second, we assessed whether there was an effect at the network level, where the incentives may not be powerful enough to produce changes in categorization behavior. We also examined effects at the hospital level, but because of very small sample sizes for some of the groups, we did not report them. Physicians employed by the VHA are salaried, and so gain no additional income from participating in gaming, and they have everything to lose if they do not follow stringent compliance and auditing standards.31,32 Finally, although it was still profitable to increase the number of patients in the study groups, it may be that increases in patient numbers in these classes had already taken place prior to our study years.
Of course, we cannot make causal inferences in this observational study. It is possible that our findings could be due to other trends. However, we would expect that our diagnosis-based classes also would be subject to such effects. Thus, the comparison of utilizationbased and diagnosis-based classes should have revealed gaming if it were present.
A large body of empirical research, as well as common sense, supports the assertion that the behavior of institutions and individuals is determined in part by how they are paid. Society generally demands that professionals such as physicians place the good of their patients ahead of financial considerations. It is likely that this trust has been arranted, but it would be naive to believe that financial considerations have no impact on how health professionals care for individual patients. For policy makers, reimbursement reforms can constitute a relatively accessible lever for system change. It is sometimes easier to modify payment arrangements than to change other organizational features.33
We found that in a system with a global healthcare budget where providers are paid by salary, but where utilization and thus patient categorization could be manipulated to increase payment to networks, no such consistent gaming behavior was observed. We are not aware of other data evaluating responses to financial incentives in healthcare systems where payment to physicians and administrators is via salary and global budgeting. Better understanding of responses to financial incentives is important, both to prevent unintended adverse effects and because using healthcare payment systems to reward quality or meet other societal goals has tremendous untapped potential.20 More research is needed to better understand the effects of financial incentives in healthcare financing arrangements.
AcknowledgmentsWe thank Mark Kuebeler, MS, and Michael Thompson, MS, for expert data management and Abel Esquivel for administrative assistance in the preparation of this manuscript.
22. Rosen AK, Loveland S, Anderson JJ, et al. Evaluating diagnosisbased case-mix
23. Centers for Medicare and Medicaid Services. Medicare Advantage(Medicare+Choice, Average Adjusted per Capita Costs) rates and statistics.
24. Ash AS, Posner MA, Speckman J, Franco S,Yacht AC, Bramwell L.