Unintended Consequences: Medicare’s Impact on the Diagnosis of Non-Enrollees

It has been well-established that prospective payment increases upcoding among patients covered by Medicare. This paper develops a model of hospital diagnosing behavior when the hospital is faced with two groups of patients (those covered by traditional insurance and those in a prospective payment system) and two types of disease (higher reimbursing and lower reimbursing). The results suggests that if physicians are unable to discriminate between patients on the basis of insurance type, hospitals with higher proportions of Medicare PPS patients are more likely to diagnose all patients into higher reimbursing diagnostic categories. I test the implications of the model with data from California hospital discharges between 1999 and 2005 and find significant evidence that increases in the proportion of Medicare patients at a hospital lead to more severe diagnoses for all patients regardless of insurance type. Medicare, Fraud, Upcoding, Insurance, DRGs, PPS JEL: I11, I18, L51, H51 Unintended Consequences (Working Paper, March 10) 3 Introduction In the United States, government involvement in health care is substantial and growing. Medicare is the largest health insurer in the United States with over 45.2 million beneficiaries and total expenditures of $468 billion in 2006 (Centers for Medicare & Medicaid Services 2009). These figures are expected to grow as baby boomers reach qualifying age. As the role of government in health care continues to be debated, and is likely to expand, the influence government exerts on medical practice is becoming increasingly important. Most research in this area has focused on the beneficiaries of publically financed care, to the neglect of privately insured individuals. It is natural that health economists have focused on patients enrolled in publically funded plans when evaluating innovations in such plans. This paper takes the analysis a step further, however, examining the unintended secondary effects on those thought to be outside of the public plan's purview. Specifically, this paper examines the effects of Medicare’s prospective payment system (PPS) on the diagnosis of non-Medicare participants. The Medicare prospective payment system (PPS) has been in place since 1983 and relies on the assignment of patients to diagnosis related groups (DRGs) based on diagnostic codes. Special software has been developed to assign the 10,000+ possible diagnostic codes into about 500 DRGs. Each diagnostic group corresponds to a fixed amount that Medicare agrees to pay the hospital for treatment. The payment amount is roughly the expected average cost of treatment for the particular DRG, controlling for hospital characteristics. If a hospital can deliver the treatment at a lower cost, it keeps the difference. Similarly, if the actual cost of treatment is higher than the expected cost, the hospital faces a loss. The idea of this payment system is to 1 There are some important exceptions. See, for example Mitchell et al. (2000), Rice et al. (1999), Finkelstein (2007), Yip (1998). Unintended Consequences (Working Paper, March 10) 4 provide hospitals with an incentive structure that controls costs. It has been studied by health economists, legal experts, and health industry professionals. The existing literature focuses on the effects of service competition among hospitals to attract PPS patients with the most profitable conditions, and the incentives for fraud (Lorence & Spink 2002; Dafny 2005; Silverman & Skinner 2004; Becker et al. 2005; Steinbusch et al. 2007). This paper extends the scope of analysis of prospective payment systems to include patients who may use the same health resources as those enrolled in the system, but whose care is paid for through some other method. Specifically, this paper examines how the incentives to commit a particular type of fraud, called upcoding, may affect patients that are not enrolled in Medicare. Upcoding Background Upcoding (also known as DRG creep) occurs when healthcare providers assign diagnoses which are not supported by medical facts, in order to increase reimbursement. This type of behavior is well documented in the literature. The first discussion of upcoding in the literature was by Simborg (1981) who noted that in the presence of the prospective payment system similar sets of patients and treatments were being coded into higher weighted categories. Since that time, several papers have focused on whether the change in hospitals case-mix were caused by actual changes in patient health, changes in coding protocol or were the result of upcoding (Steinwald & Dummit 1989; Carter et al. 1990; Hsia et al. 1992). The results from these early studies suggested that DRG creep was likely a product of increased illness and coding complexity, rather than motivated by profit Unintended Consequences (Working Paper, March 10) 5 seeking behavior. More recent studies, however, suggest that the profit motive may be more responsible for upcoding than originally thought. Interest in upcoding was reignited in the late 1990s when a series of costly lawsuits, an investigation by the New York Times (Gottlieb & Barbanel 1997), and a targeted investigation by the Department of Justice revealed substantial evidence of upcoding among large providers. These lawsuits greatly increased the expected financial and reputational cost of Medicare upcoding, and may have encouraged researchers to reexamine the evidence. In a survey of health care information managers conducted in the wake of the Department of Justice lawsuits, Lorence and Spink (2002) found that many mangers were routinely pressured to upcode reimbursement data used in claims filing. This type of upcoding is most easily accomplished when there is little documentary evidence of the specific illness diagnosed; a situation that occurs with surprising frequency. Psaty et al. (1999) reviewed the medical records of 485 subjects with a primary or secondary Medicare discharge diagnosis of heart failure and found that 37.5 percent reflected false positives. They concluded that if the results apply to the entire Medicare population, upcoding of this one diagnosis may have cost Medicare up to $933 million in 1993. Similarly, McCarthy et al. (2000) reviewed the medical records of 485 elderly Medicare recipients and found that 30 percent of medical and 19 percent of surgical patients lacked any documented evidence, such as lab results, supporting their diagnosis. Silverman and Skinner (2004) find that for-profit hospitals are significantly more likely to upcode than not-for-profit hospitals. They also find that hospitals under financial Unintended Consequences (Working Paper, March 10) 6 distress upcode less than others. In analyzing an exogenous change to the Medicare PPS, Dafny (2005) found little evidence that hospitals change the intensity of care in response to changes in the PPS system, but they do employ sophisticated coding strategies to maximize reimbursement. Norton et al. (2002) found that if the prospective payment system increases the average price of an inpatient stay, the length of stay is likely to increase while changing the marginal price of inpatient stay has little effect. Finally, Becker, Kessler and McClellan (2005) use state by state data on fraud detection expenditures, social security records, hospital characteristics and Medicare claims data to analyze how Medicare providers respond to law enforcement. They find that increased enforcement leads certain types of patients and hospitals to have lower billings, without adverse consequences for patients’ health outcomes. While previous studies have documented the effects of the prospective payment system on Medicare participants, they have ignored how upcoding incentives may affect patients which use the same resources as Medicare PPS patients, but are not enrolled in PPS insurance plans. This paper seeks to fill this gap in the literature by building a theoretical model of upcoding from the perspective of the provider, and then testing the model using data from California from 1999-2005. Medicare’s PPS and non-Medicare Patients The incentive to upcode Medicare patients is relatively straightforward: a higher reimbursing diagnosis results in a higher profit, provided that the patient is treated the same as if they were diagnosed into the lower reimbursing diagnosis. Upcoding of non-prospective 2 Silverman and Skinner conjecture that this may be because financially stable hospitals may be able to afford better coding software or training of coding consultants. Unintended Consequences (Working Paper, March 10) 7 payment system patients is more difficult to explain because reimbursement is not directly tied to diagnosis; hospitals are instead reimbursed based on the actual cost of treatment. There are several mechanisms, however, that may cause the incentive to misdiagnose Medicare patients to spill over to non-Medicare patients. Four such mechanisms are described below. First, the sheer size of Medicare in the insurance market allows it to shape clinical medicine. As the largest single insurance organization in the nation, it would be virtually impossible for Medicare to have no effect on clinical medicine. This market power is a double edged sword: well designed policy will contain costs not only for Medicare patients, but non-Medicare patients as well; poorly designed policy may raise medical costs for Medicare and non-Medicare patients alike. Second, because medical treatment is based on the patient’s symptoms and not her insurance type, the rise of medical algorithms for use in the diagnosis and treatment of disease limits the ability of physicians to discriminate based on the patients' insurance type. This means that algorithms for diagnosing and treating Medicare patients will be applied to non-Medicare patients as well. If these procedures are designed at the hospital level to maximize profits by placing Medicare recipients into higher DRGs, it may affect non-Medicare patients as well. Third, coding (the translation of patient charts into standardized codes for reimbursement) is usually performed by someone other than a physician. If aggressive coding is occurring at the coding level, with coders that are trained to code the most remunerative diagnoses, these effects are not likely to be confined to Medicare recipients. 3 One way this could happen is if a diagnostic test which produces a high number of false positives for higher reimbursing diagnoses is used instead of a more sensitive test. Unintended Consequences (Working Paper, March 10) 8 Fourth, the risk of regulatory scrutiny is likely to be raised if treatment patterns vary according to insurance type. In order to avoid detection, fraudulent providers may choose to treat Medicare and non-Medicare patients the same. For example, because different DRGs exist for diagnoses with and without complications (e.g. respiratory infection with complications/ respiratory infection without complications), hospitals may have an incentive to perform additional diagnostic tests in order to find complications. If regulators note that these tests are only being performed for Medicare patients, it will likely raise suspicion of fraudulent activity. If the expected marginal benefit of these tests for the patient is less than the expected marginal cost, it may result in inefficiencies in the treatment of non-Medicare patients as well as Medicare patients. Parallels with Tax Evasion As noted by Silverman and Skinner (2004), upcoding is similar in many respects to tax evasion. In both types of fraud, there are positive returns if the fraud goes undetected, but negative returns if the fraud is detected. Also, in both types of fraud, the probability of detection may be related to the magnitude of the fraud (e.g. an individual in a high paying profession that claims little taxable income and a provider that claims all patients have the higher reimbursing DRG are both likely to be investigated). Finally, in both tax evasion and Medicare upcoding, there is a cost to detecting fraud which precludes the regulator from auditing every tax payer or provider. I therefore draw on the tax evasion literature in the development of my model of Medicare fraud. Unintended Consequences (Working Paper, March 10) 9 A complete model of upcoding should include a model of policy optimization from the point of view of the regulator; however, we learn from the tax evasion literature that there are several problems with this approach. As noted by Andreoni et. al. (1998), models of tax evasion that consider the regulator’s optimal audit strategy can essentially be divided in to two groups: those that assume the regulator can announce a rule before the game begins and those that suggest the regulator cannot commit to an audit rule. Models of the first type generally imply that all audited taxpayers will report their earning honestly, a fact which is clearly contradicted by the empirical evidence. Models of the second type lead to multiple equilibria and it is difficult to arrive at simple and testable conclusions. A more general problem with game theoretic models in this area is that most models assume that taxpayers are informed about how the regulator arrives at the decision to audit-a tenuous assumption at best given that the auditing strategies of most governments are closely guarded secrets. This criticism can be applied to Medicare fraud as well, as there is no definitive criteria regarding what will trigger investigation. Because of these issues, my model of upcoding follows the more general trend in the tax evasion literature of modeling the provider’s behavior under what the provider may reasonably assume about the regulator’s behavior. Model Each provider maximizes their expected profit function which is dependent on the proportion of patients that are covered by Medicare PPS, reimbursement under PPS and feefor-service, the gains to fraud, the probability of fraud detection and the costs of detected fraud. The expected profit function is Unintended Consequences (Working Paper, March 10) 10 * [ ( *) (1 ( *)) ( ) ( ) ) (1 ( ) ( )) ( ) {( ) ( ) ( ( ) } )] . c A i B i i i