Adjusting for clustering in survey research

The current widespread application of cluster sampling in survey design presents concerns to the researcher regarding the chosen unit of study. The statistical power of a study could be spuriously increased or decreased, sample size could be altered dramatically, and the uniqueness of the individual unit under study (e.g., a patient) could be lost and substituted by average values of characteristics depending on which unit of study is selected. The 1985 National Ambulatory Medical Care Survey of the U.S. National Center for Health Statistics was used to illustrate research methodologies that would deflate spuriously high results due to clustered sampling. Using the individual patient as the study unit, correction factors ranging from 1.99 to 35.40 were calculated in order to deflate exaggerated t-tests, chi-square, and F values of predictor variables. The effect of clustering was more marked on (1) continuous rather than binary outcome variables, as the former provide a richer environment to form clusters; and (2) on outcome variables relating to the practice-related or personal style of physicians. Using the physician (who represented a cluster of patients) as the unit of study, it was realized that correction factors were relevant to physician-specific predictors but not patient-specific predictors. Using design effect correction factors developed from a simple univariate analysis of variance, pharmacoepidemiologists can analyze accurately the currently available large survey databases of clustered samples.