OBJECTIVES: The purpose of this study was to develop and test an algorithm for estimating disease severity for economic studies involving administrative cost database analyses. The intent was to go beyond the traditional three-group classification (mild, moderate, severe) and develop a more refined four-group system consistent with NHLBI Guidelines for estimating asthma severity: mild intermittent (MI), mild persistent (MP), moderate persistent (MoP), severe persistent (SP). METHODS: A retrospective cohort design was used, employing administrative claims from a fee-for-service provider of health care benefits to approximately 3.5 million federal employees in the U.S. Patients with asthma were defined by: ? 1 medical encounters for asthma (exclusive of COPD and allergic rhinitis) during 1994-5 and continuous enrollment in the health plan for 24 months; 22,833 patients with asthma were identified. Four algorithms were developed using NHLBI guidelines for pharmacologic intervention with adjustments for practice patterns during 1994-5. Evaluation was based on the observed severity distribution in the population and results from bivariate and multivariate analyses in random samples, examining the relationship between severity level and the following clinically-grounded variables: deaths, hospitalizations, ER visits, and use of an allergist/pulmonologist. RESULTS: The final algorithm was a 2-step procedure based on ?-2 agonist and oral steroid use. Severity distribution in the population was as follows: 69.5% MI, 16.9% MP, 11.1% MoP, and 2.5% SP. The final, population-based bivariate analyses showed a logical stair-step pattern with significant relationships between severity and death rates, number of hospitalizations, and specialist use (p < .001). The ordinal logistic regression model was also significant (p < .0001; chi-square=597.45, df=4), with a Goodman-Kruskal Gamma statistic of 0.25. CONCLUSION: Results suggest the algorithm is useful for classifying patients into four levels of severity using an administrative cost database. Clinical validation of the algorithm is warranted.