Modeling physical activity outcomes: a two-part generalized-estimating-equations approach

The health benefits of physical activity are well known. Various parameters of physical activity, such as time or energy expenditure, are often assessed in observational and experimental studies. This article highlights several methodologic issues concerning the analysis of physical activity. These include non-normality, presence of many zeros, and violation of the independence assumption. Application of the standard regression model to a (log-transformed) physical activity variable may lead to spurious associations and misleading conclusions. We developed an alternative 2-part generalized-estimating-equations (GEE) approach to analyze the heterogeneous and correlated physical activity data. We first estimated a logistic GEE model for the prevalence of physical activity and factors affecting physical activity participation. We then fit a gamma GEE model to assess the effects of predictors among persons engaging in physical activity. An empirical application to an epidemiologic study of physical activity of community-dwelling older adults illustrates the proposed methodology.