Drs. Hornberger and Wrone discuss "When To Base Clinical Policies on Observational Versus Randomized Trial Data." Two problems encountered with this decision are how to balance the strengths of the research design of randomized trials with the weaknesses of costs and the time it takes to achieve results compared with database research. Database research is generally cheaper and faster but is weaker on research design-this is not a trivial problem for many areas of research. The article provides a method for estimating the costs and benefits of a randomized study. The utility of the method is well demonstrated in an example where the results convinced an HMO to sponsor a randomized trial. Although the method itself requires some time and effort, it provides direction for investigators to objectively consider a range of potential outcomes for their proposed research. The method may be useful to investigators in convincing other funding agencies to sponsor research.

Matchar, in "Linking Evidence and Clinical Decisions," describes the potential uses of another type of database called the Stroke Prevention Policy Model. Although the database is disease specific, the sources of data are wide and include epidemiologic studies, randomized clinical trials, administrative cost data, and patient interviews. These sources of data are chosen specifically because each one is the single best resource for those specific data items. Thus, the database obviates many of the limitations of administrative and clinical databases. A disadvantage is that the focus is limited to one disease area. However, as the examples show, combining this database with sophisticated analyses and simulation provides a new technology for answering clinical and policy questions surrounding stroke prevention. Such questions are extremely difficult and costly to answer by other methods.

In "Policy and Program Analysis Using Administrative Databases," Ray quickly gets our attention by listing critical problems in the area of database research. Such problems include poor data quality, lack of concurrent controls, inability to ascertain important outcomes, and incomplete data on case mix. More important, Ray provides examples of studies in which insufficient attention was given to these problems and resulted in erroneous program analysis and policy formation. Ray suggests strategies to minimize these problems in the future. However, even with these changes (which are major), he cautions that it will not be possible to evaluate the effect of many policies.

Before we give up on database research, it is important to consider that Ray reports on administrative databases. This specific type of database is primarily designed to help administrators with accounting. As such, the focus is on costs and cost categories, numbers of office visits and emergency room visits, and hospitalizations with coded diagnoses. These database were not designed to assess quality of care, case mix, or patient outcomes. Thus, although administrative databases are valuable in answering administrative questions, substantial problems arise when attempts are made to answer questions in areas for which they were not designed.

Finally, in "Linking Automated Databases for Research in Managed Care Settings," Selby describes a more optimistic role for database research. The major difference here is that Selby reports on an administrative database combined with a clinical database. That is, laboratory reports, radiographs, medications, and procedures can be accessed, as well as administrative data. With these additional data, opportunities for asking questions about quality of care, adjusting for case mix, and looking at patient outcomes now exist. This clinical database has been enriched by disease registries that have been established within it and that provide additional clinical data in a systematic manner. Further, the database is linked with state or national mortality databases, thereby providing more complete follow-up of patient outcomes. Despite the wealth of available data, Selby recognizes that some limitations remain in this type of database, including lack of data on race and ethnic group, socioeconomic status, disease-specific severity, and functional status measures. He proposes incorporating these data into the database and finding ways to link databases among a number of centers (HMOs). As if this were not enough to generate our enthusiasm, Selby provides several examples of successful studies on utilization, clinical outcomes, and cost of illness.

For persons considering database research or attempting to interpret reported results and those who want their next T-shirt to read, "Happiness is a Humongous Database," these reports are the place to start.

David M. Smith, MD

The Regenstrief Institute for Health Care; Indiana University Medical Center; Indianapolis, IN 46202