OBJECTIVE: Evaluate the importance of accounting for correlations between observations when analyzing repeated measures of health care utilization. METHODS: Poisson regression was applied with and without generalized estimating equations to analyze ?-agonist use, health care contacts, and nocturnal awakenings in chronic asthmatics in two large similarly designed, 12 week, randomized trials of Montelukast, a once daily leukotriene receptor antagonist, compared to placebo. Data on ?-agonist use and nocturnal awakenings were recorded on daily diary cards. Health care contacts included visits to physicians, the emergency room, and hospitalizations documented as not mandated by the protocols. RESULTS: Montelukast compared to placebo decreased the rate of days of ?-agonist use by 9.6% (p<.0001), nights with awakenings by 22.5% (p<.0001) and health care contacts by 23.0% (p=.09) when analyzed without generalized estimating equations. Use of generalized estimating equations did not change the parameter estimates for any of the health outcomes measured, but did increase the standard errors and confidence intervals in all cases. The increase in the standard errors ranged from 19.1%-355.1%. ?-agonist use and nights with awakenings remained statistically significant. However, health care contacts became less significant (p=.16). CONCLUSION: Failure to account for correlations between observations may lead to erroneous conclusions based on significance levels that are too small. In large samples, exceeding 100-200 patients, generalized estimating equations provide one method of obtaining correct significance levels and confidence intervals when data are correlated. The example above included daily values for 12 weeks in 1300 patients. Other methods of adjusting for correlated data are needed in smaller samples.