Objective: To estimate the frequency with which patients are incorrectly used as the unit of analysis among statistical calculations in published studies of physicians' patient care behavior.
Design: Retrospective review of studies published during 1980-1990. ARTICLES: 54 articles retrieved by a computerized search using medical subject headings for physicians and study characteristics. Article selection criteria included the requirement that the physician should have been the correct unit of analysis.
Intervention: Presence of the error was determined by consensus using published criteria.
Main results: The error was present in 38 articles (70%). The number of study physicians was reported in 35 articles (65%). The error was found in 57% of articles that reported the number of study physicians and in 95% of those that did not. The error rate was not lower among articles published more recently nor among those published in journals with higher rates of article citations in the medical literature.
Conclusion: The unit of analysis error occurs frequently and can generate artificially low p values. Failure to report the number of study physicians can be a clue that this type of error has been made.