ILLUSTRATION OF ANALYSIS TAKING INTO ACCOUNT COMPLEX SURVEY CONSIDERATIONS - THE ASSOCIATION BETWEEN WINE CONSUMPTION AND DEMENTIA IN THE PAQUID STUDY

Epidemiologists are increasingly looking to large-scale sample surveys to provide data for studies of the associations between known or susp ected risk factors and disease. More often than not, widely available statistical software packages have been used to analyze such data, par ticularly when multivariable modeling is involved. Such packages assum e that the data have resulted from simple random samples. However, whe n the survey design incorporates such features as clustering and strat ification, the results of statistical analyses based on this assumptio n can be incorrect. The authors utilized data from the PAQUID (Personn es Agees Quid) study, collected periodically from 1988 to 1996, to ill ustrate the ease of performing a ''design-based'' (vs. a ''model-based '') analysis of complex survey data, and they compared the results obt ained using both approaches. The PAQUID study is a stratified cluster sample of elderly community residents in the southwestern departments of Gironde and Dordogne, France. In the illustration presented-in whic h 3,777 community residents aged 65 years or older were selected to pe rmit identification of baseline and lifetime factors that might be rel ated to cognitive loss, dementia, and Alzheimer's disease-measures of association (such as odds ratios and their associated standard errors) were comparable for both analytical strategies. However, this may not be the case for other examples. Descriptive measures (such as estimat es of means and proportions) may be more seriously compromised by the decision to ignore the sampling design. The availability of modern sta tistical packages with survey analysis capabilities should encourage d ata analysts to perform design-based analyses whenever possible. Am J Epidemiol 1998;148:298-306.