As we march through the information age, we increasingly see research that draws information from administrative data sets. We are also witnessing an explosion of clinical guidelines supported by administrative data. With growing frequency, those to whom we must be accountable gauge our professional performance with measures culled from administrative data. In addition, we will probably add to and interact with such databases every day during the course of routine patient care. After all, whenever we check off codes on encounter forms or enter data into electronic medical records, we are helping to create a large database. Thus, the proceedings of the sixth Regenstrief Conference are definitely relevant to Annals clinician-readers.

Work drawn from large databases now often appears in clinical journals, such as The Lancet, JAMA, and The New England Journal of Medicine. A recent study used computer records of patients from 143 general practices in the United Kingdom to explore the risk for venous thromboembolism and various oral contraceptives [1]. An electronic database of discharges from 78 hospitals in 23 states permitted the development of a prediction rule to identify patients with community-acquired pneumonia who were at low risk for adverse outcomes [2]. Electronic pharmacy files, telephone billing records, and registries of motor vehicle accidents permitted creative investigations of the associations of motor vehicle crashes and benzodiazepines [3] and cellular telephones [4].

However, one need not venture off the pages of Annals for examples of the pertinence of large databases to clinicians. An increasing number of clinical research studies originate in administrative files. Investigators often begin by turning to one administrative data set or another to identify cases with codes that indicate patients who have the clinical condition of interest. For example, Ashton and colleagues [5] used the computerized hospital discharge database of the Veterans Affairs medical system to evaluate eligible participants and their outcomes for a study of factors related to hospital readmission. To identify participants eligible for aspirin after acute myocardial infarction, Krumholz and colleagues [6] used Medicare claims. Often, as these authors did, researchers must augment administrative data with the collection of primary clinical data. At other times, when the electronic files are sufficiently rich, the investigators need not leave them to answer the research question.

Researchers can explore certain clinical issues much more thoroughly and efficiently by using large, electronic databases than by using smaller sets of primary data. For example, Annals recently published a study on the increasing incidence of osteoporotic ankle fractures [7]. The study data came from Finland's national hospital discharge register, a single database that tracks all of the hospitalizations of that country's 5 million inhabitants. Closer to home, a population-based database of linked medical records for Rochester and Olmstead County, Minnesota, permitted an understanding of the changing epidemiology of primary hyperparathyroidism [8]. Careful epidemiologic study of these clinical issues would have been impossible without the fertile data available in these population-based electronic data sets. Many of the selection biases that plague researchers dissipate when the researchers can study entire populations by carefully using administrative data. If we are unaware of the potential strengths and limitations of data in large databases, we will be ill-prepared to evaluate research that uses these data.

Perhaps the most prevalent role of large databases in the 1990s, and the one that causes clinicians the most worry, is the measurement of quality of care. Electronic records clearly can offer a tool for improving and monitoring the processes and outcomes of care. These files can demonstrate whether clinicians' practice parallels or deviates from recommended guidelines. For example, data from cardiac surgery registries in New York State and Ontario allowed a comparison of rates of coronary artery bypass grafting in these two areas [9]. This research offered insight into how to increase the appropriate use of this procedure. Pine and colleagues [10] showed how the augmentation of administrative files with clinical laboratory data can improve the ability to predict adverse outcomes from administrative files.

However, investigators and administrators may misuse large databases. In 1995, Localio and colleagues [11] showed how a managed care program's analysis of hospital discharge data had a 60% chance that at least 1 of 22 hospitals studied would be falsely labeled as a high-mortality hospital. The authors warned that "the risk posed by improper analysis to the reputations of skilled providers and to the health of the patients they serve leads us ... to advocate stringent criteria for judging the appropriateness of analytic methods for comparing hospitals and physicians based on patient outcomes." The supplement accompanying this issue of Annals contains further warnings about potential errors in the use of large databases. These papers contain valuable lessons for clinicians who wish to understand how insurers and others may have derived measures of clinicians' own performance. Such an understanding is necessary equipment for contesting faulty measures.

Annals authors have previously discussed the power of electronic medical records [12]. The supplement advances this discussion by describing the clinically relevant information that is already captured in computers. The supplement papers offer guidance to clinicians, researchers, administrators, and other interested parties on improving the usefulness of these data.

Why are we seeing more medical research that uses large health care databases? Cynics might say we are using these databases simply because they are there, sitting on computer drives waiting for someone to use them. Others may believe that managed care organizations and other payers have concocted these files just so they can hang them over clinicians' heads. Advocates of health services research would probably argue that these databases are powerful tools with which we can study health care systems in an effort to improve them. Each of these factors likely explains, in part, the increasing presence of the large database in medicine.

The papers in the supplement do not unequivocally extol the merits of large electronic databases. The authors of the papers know these databases intimately and offer as much caution as promise. Readers can learn about these huge data files so that they can both respect their limitations and draw on their strengths. Whether large databases will ultimately prove to be as useful or as dangerous as various parties believe remains to be seen. What is certain is that we will be seeing more of them in our roles as consumers of research, subjects of performance measures, and data collectors helping to complete our patients' increasingly electronic medical records.