Illustration of analysis taking into account complex survey considerations: the association between wine consumption and dementia in the PAQUID study. Personnes Ages Quid

Abstract

Epidemiologists are increasingly looking to large-scale sample surveys to provide data for studies of the associations between known or suspected risk factors and disease. More often than not, widely available statistical software packages have been used to analyze such data, particularly when multivariable modeling is involved. Such packages assume that the data have resulted from simple random samples. However, when the survey design incorporates such features as clustering and stratification, the results of statistical analyses based on this assumption can be incorrect. The authors utilized data from the PAQUID (Personnes Agees Quid) study, collected periodically from 1988 to 1996, to illustrate the ease of performing a "design-based" (vs. a "model-based") analysis of complex survey data, and they compared the results obtained 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 which 3,777 community residents aged 65 years or older were selected to permit identification of baseline and lifetime factors that might be related 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 estimates of means and proportions) may be more seriously compromised by the decision to ignore the sampling design. The availability of modern statistical packages with survey analysis capabilities should encourage data analysts to perform design-based analyses whenever possible.