Selection of working correlation structure in generalized estimating equations

Abstract

The well-known generalized estimating equations is a very popular approach for analyzing longitudinal data. Selecting an appropriate correlation structure in the generalized estimating equations framework is a key step for estimating parameters efficiently and deriving reliable statistical inferences. We present two new criteria for selecting the best among the candidates with any arbitrary structures, even for irregularly timed measurements. The simulation results demonstrate that the new criteria perform more similarly to EAIC and EBIC as the sample size becomes large. However, their performance is much enhanced when the sample size is small and the number of measurements is large. Finally, three real datasets are used to illustrate the proposed criteria. Copyright © 2017 John Wiley & Sons, Ltd.

Vitamin K status and physical decline in older adults—The Longitudinal Aging Study Amsterdam

“…To determine the longitudinal relationship between the dp-ucMGP tertiles and physical functioning, we used a generalized estimating equation analysis to estimate regression coefficients or in case of functional limitations odds ratios (OR) with 95% confidence intervals from the WALD-test adjusting for potential confounders. We used an exchangeable correlation structure because of the close correlation between follow-up examination within participants [22]. A minimally adjusted model included dp-ucMGP, adjusted for follow-up time (years) and baseline age and sex (model 1).…”

The longitudinal relationship between fatigue, depression, anxiety, disability, and adherence with cognitive status in patients with early multiple sclerosis treated with interferon beta-1a

“…To assess the longitudinal relationship in 5 separate association models between the physical and psychological symptoms and adherence with cognitive status, we applied a binominal generalized estimating equation (GEE) analysis expressed as relative risks and 95% confidence intervals. Due to the high correlation between follow-up measures within patients, we used an exchangeable correlation structure [ 23 ]. We present a crude model (unadjusted) and a model adjusted for education (3 categories) and time since MS diagnosis (months).…”

Using discrete choice model to elicit preference for health-care priority setting

“…This controversy centers on how well the correlation parameters can be estimated. [ 24 ] For example, if R represents the within-group correlation for square matrix of max ({ni}) × max ({ni}), the correlation specifies the structure of R. Let R ts denotes the t, s element. The independent structure is defined as…”