Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
11 Claims Data and Effectiveness: Acute Myocardial Infarction and Other Examples Barbara ]. McNeil The question of effectiveness of medical treatment is an extremely im- portant one and one that will benefit from close collaboration between physicians and social scientists. In this chapter, however, I confine my discussion to a limited aspect of that collaboration- that is, to the analysis of claims data, particularly analysis of Medicare claims data. My discussion is based on the claims data as they exist today. It is important to note, however, that since these data have begun to be used for prospective payment, their accuracy has improved considerably. I think we can expect improvements of similar magnitude once these data are used to a greater extent for research on effectiveness and outcomes, particularly as they relate to medical technology. The original definition of medical technology from the Office of Tech- nology Assessment (OTA) considers two types of technologies. The first is any medical device, drug, or surgical procedure used in the care of patients. The second is any organizational or support system within which medical care is delivered. It is unlikely that claims data in their current form will be usable in the latter, so I will restrict my comments to the first type of technology. STRENGTHS OF CLAIMS DATA Large claims data bases have a number of strengths. To illustrate these I draw upon the experience of many other researchers and on my own experience as a researcher and as a commissioner with the Prospective Payment Assessment Commission (ProPAC). The following list illustrates the most notable strengths. It applies principally to Medicare Part A data (primarily hospitalization data) because that is where most of our experience has been thus far. Part B (ambulatory) data, particularly when linked to Part A data, expand these 65
66 EFFECTIVENESS AND OUTCOMES IN HEALTH CARE strengths still further. Such linkage is costly, however, and initial efforts are just being completed. Strengths of claims data include the following: 1. They can be used to provide usage rates. 2. They can be used to indicate variations in use of technology by geog- raphy, hospital type (e.g., teaching, nonteaching; urban, rural), age, sex, and so on. This is the area in which John Wennberg has worked so successfully over the years. 3. They can be linked to mortality data in order to define mortality rates as a function of the above items and as a function of key diagnostic and procedure codes. This is the basis of the Health Care Financing Administration's (HCFA) initiative in providing mortality rate data to hospitals. 4. When linked with the Medicare Cost Report, claims data can be used to estimate the costs of hospitalization. Comparative data can also be obtained across types of institutions. If patients' records were linked over time and Part B data were linked with Part A data, they could be used to provide information on the costs of an episode of care. (This assumes that it is possible to define an episode accurately.) 5. They can provide information on home health services. Although this list of strengths is long, for our initial activities in the Effectiveness Initiative, we will largely be talking about items 1, 2, and 3. GENERAL LIMITATIONS There are four serious limitations to these Medicare claims data. First, there is very limited information on comorbidity and disease severity. Thus, it is difficult, if not impossible, to define an "inception cohort" that is, a homogeneous group of patients whose identity is clearly and reproducibly defined at a particular time and who are then followed into the future. Second, there is limited information on socioeconomic status, and much recent literature has shown that socioeconomic status correlates well with usage of certain health services and medical technologies. Third, data on outcome are sparse. Currently, they allow us to measure mortality rates and readmission rates; however, it is not always possible to determine whether a readmission is related to the prior admission, is a consequence of suboptimal care, or is an unrelated event. Because much of medical care is designed to reduce morbidity rather than mortality, omission of data on postdischarge functioning of the patient and on alleviation of the symptoms that generated the hospitalization limits the usefulness of current outcomes data to research. Moreover, as we think about incorporating outcomes data, we should think about obtaining data at times after discharge that reflect the expected results of the hospitalization. For example, outcomes
IOM CLINICAL CONDITION WORKSHOPS 67 after a cholecystectomy should probably be obtained at 3 months, but out- comes after hip replacement surgery should probably wait for 6 to 12 months. Fourth, many codes used to describe diagnoses and procedures are nonspe- cific, as discussed below. Recent work by Lisa Iezzoni and her colleagues on coding of acute myocardial infarction illustrate some of these limitations (1~. This study reports that more than one-quarter of the patients assigned an acute myocardial infarction code from the International Classification of Diseases (ICD-9- CM) at the time of discharge did not have the condition or receive active treatment for the condition during hospitalization. Miscoding resulted most often when patients were admitted with a "rule-out infarction" diagnosis. Misspecification (that is, the physician failed to note explicitly the absence of acute myocardial infarction) or failure of the medical abstracter to note subsequent explicitly documented exclusion of the infarction resulted in the largest number of coding errors. Admission of patients for cardiac catheterization with coronary angiography within 8 weeks of acute myocardial infarction (thus technically permitting the acute myocardial infarction code) was cited as another major reason for misclassification. The difficulties raised by the coding guidelines for the ICD-9-CM and the diagnosis-related group (DRG) codes are further compounded when a secondary diagnosis of acute myocardial infarction is used to assign the infarction DRG to cases where another cardiac condition is the principal diagnoses. The study supports the conclusion that previous hospital discharge data on acute myocardial infarction lack sufficient validity in themselves to define an inception cohort for effectiveness and outcomes research. As coding rules change over the next year, however, to minimize some of the above-mentioned problems, identification of an inception cohort from the discharge codes will become more accurate. Inaccuracy of diagnostic codes is not unique to acute myocardial infarction. In the next section, I amplify on the four general limitations of claims data in the context of assessment of three technologies: diagnostic devices, drugs, and clinical trials. LIMITATIONS FOR DIAGNOSTIC DEVICES This is probably the area in which claims data are likely to be least useful, in the absence of significant changes. The first limitation derives from the fact that, for inpatients, Medicare claims files code for only three procedures. Ill patients usually have significantly more than three diagnostic procedures, and hence the list of coded diagnostic procedures is frequently incomplete and biased. It is biased because sicker patients will not have room on the claim for the diagnostic code, whereas healthier patients will.
68 EFFECTIVENESS AND OUTCOMES IN HEALTH CARE An example of this phenomenon occurred when ProPAC tried to track the use of magnetic resonance imaging (MRI) among Medicare beneficiaries. There were far fewer MRIs reported than we estimated had been done. In addition, there were far fewer done on sicker patients in the DRGs most likely to make use of MRI. This is analogous to the phenomenon described by Stephen Jencks regarding concurrent diagnoses among ill patients (21. The second problem in the evaluation of the effectiveness of diagnostic devices relates to time-lags between the use of new technologies and the development of codes for them. Development of codes can take years, thus preventing us from identifying the use of new devices. Although this is a limitation primarily of inpatient records, it can occur in outpatient records as well. Examples of coding omissions that hinder evaluation include MRI, electrophysiological studies, and positron-emission tomography (PET). Third, claims data do not provide any information on the type of equipment used. For imaging technologies this is critical: major differences in effectiveness can result from use of older generations of equipment. Fourth, it is seldom possible to differentiate between tests done for diagnosis and those done for screening. This is obviously important in the case of mammography. Finally, there is no correlation of diagnostic test results with information from an independent source (for example, pathology). It is important to emphasize that, to the extent that we have information from inpatient care (ICD-9-CM codes) and ambulatory sources (ICD-9-CM or Current Procedural Terminology [CPI] codes) some of the above problems can be alleviated. In any case, the limitations described above regarding inpatient data have prompted the National Cancer Institute to conduct a major prospective study of the effectiveness of diagnostic imaging procedures in patients with one of five types of cancer. Nine institutions are currently collaborating in this study, and six more are expected to be added next year. LIMITATIONS FOR DRUGS The problems of claims data for drugs are similar to those for diagnostic devices. Codes for new drugs may lag their availability by many years. The classic example of this relates to thrombolytic therapy. Most physicians, policymakers, and researchers identified this as an extremely important area for study two years ago; however, there were no codes for thrombolytic therapy. There are still no codes for the therapy per se it can be identified (and then not always) only when done in connection with an angioplasty. Drugs are very complicated to evaluate because of multiple doses and multiple forms, and it is going to be tricky to get information on outpatient drug use. The repeal of the Medicare Catastrophic Coverage Act, with its drug coverage, will make information on Medicare beneficiaries more difficult to obtain. However, a number of researchers have been extraordinarily
IOM CLINICAL CONDITION WORKSHOPS 69 successful in using claims data from selected states (for example, New Jersey) (3~. LIMITATIONS FOR THERAPIES There has been a tremendous amount of discussion about the use of claims data for therapies, and much of it has been very negative. I think we should recognize, however, that a number of useful things can be accomplished with claims data for therapy. For one, we may not need to resort to randomized trials for all interventions. Some limitations remain, however. The first one is that the coding for therapy is not always current. For example, two years ago ProPAC was interested in studying cochlear implants as a new therapy for patients with deafness. At the time there was no way of identifying these patients from hospitalization claims data alone. The second problem with the coding for therapies is that the code may not be specific enough. This is particularly troublesome for ICD-9-CM codes used on inpatient records. CPT codes are considerably more specific in reporting procedures although they have little or no diagnostic information. Thus, if bills for physician services or outpatient services are linked with hospitalization records, specificity is improved. Failing that linkage, there are problems in four areas: 1. The ICD-9-CM codes do not reflect refinements in a procedure (for example, a cementless instead of a cement hip prosthesis). 2. The codes frequently do not indicate whether a procedure was a repeat one (for example, a first or a second coronary artery bypass graft). Linking patient records over many years (for example, 10 years) would solve this problem if the payer were the same during the entire period. 3. The codes are sometimes incomplete. A one-year study of total parenteral nutrition (TPN) conducted by ProPAC illustrates this. At that time it was believed that DRGs 296 and 182 (nutritional disorders and miscellaneous digestive disorders) would contain many patients having TPN. A review indicated that approximately 1,200 patients that year (less that 1 percent of all patients in those DRGs) were identified from the claims records. Independent estimates suggested a number more like 100,000 to 200,000 patients. In this case, as with MRI, sicker patients had enough other procedures done to them that TEN never reached the claims records. 4. Claims data seldom allow identification of an inception cohort. This was mentioned under general limitations, but I repeat it here because of its particular importance for evaluation of therapies. Elliot Fisher and John Wennberg emphasize this in their discussion of the claims analyses of transurethral prostatectomies (41. In general, it will be easier to define an inception cohort for an acute event, such as acute myocardial infarction, than for a chronic one.
70 EFFECTIVENESS AND OUTCOMES IN HEALTH CARE CONCLUSION Finally, where are we? I think we are at a point in the careers of a number of health services researchers that is really quite rosy. We have a data base that is constantly being improved and will continue to be improved as a result of our research interests. Over the short term, I believe that these data will be used primarily for generating hypothesis. Our resultant analy- ses and studies will have an obvious impact on our ability to measure the effectiveness of medical practice. Over a longer term, it is likely that some of our results will be used to identify access problems. Who is not getting what? For what reason? To accomplish both short- and long-term objectives, we must work closely with policymakers on activities related to improving the data base and the training of individuals capable of using it. REFERENCES 1. Iezzoni, L.I., Burnside' S., Sickles, L., et al. Coding of Acute Myocardial Infarction: Clinical and Policy Implications. Annals of Internal Medicine 109:745-751, 1988. 2. Jencks, S.J. Issues in the Use of Large Data Bases for Effectiveness Re- search. Pp. 94-104 in Effectiveness and Outcomes in Health Care. Heithoff, K.A. and Lohr, K.N., eds. Washington, DC: National Academy Press, 1990. 3. Avorn, J., Dreyer, P., Connelly, K., et al. Use of Psychoactive Drugs and Quality of Care in Rest Homes. New England Journal of Medicine 320:227-232, 1989. 4. Fisher, E.S. and Wennberg, J.E. Administrative Data in Effectiveness Studies: The Prostatectomy Assessment. Pp. 80-93 in Electiveness and Outcomes in Health Care. Heithoff, K.A. and Lohr, K.N., eds. Washington, D.C.: National Academy Press, 1990.
