Assessing the use of GEE methods for analyzing family data: the Strong Heart Family Study

• Activity Number: 442 - Disease Prediction, Statistical Methods for Genetic Epidemiology and Mis
• Type: Contributed
• Date/Time: Thursday, August 12, 2021 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Xi Chen* and Ying Zhang and Amanda M. Fretts and Tauqeer Ali and Jason Umans and Richard Devereux and Elisa Lee and Shelley Cole and Daniel E Zhao
• Companies: University of Oklahoma Health Sciences Center and University of Oklahoma Health Sciences Center and University of Washington and University of Oklahoma Health Sciences Center and Georgetown University and Weil Cornell Medicine and The University of Oklahoma Health Sciences Center and Texas Biomedical Research Institute and OUHSC Hudson College of Public Health
• Keywords: family study; generalized estimating equations; cluster data; kinship matrix; simulation

Abstract

Abstract: In family studies, unbalanced family size and complicated pedigree structures pose challenges in the data analysis. Although the generalized estimating equations (GEE) have been commonly used, there is no systematic study to evaluate the GEE performance on family data. To close the knowledge gap, we performed a simulation study for family data with continuous, binary and frequency data type. Kinship matrix, derived by the SHFS Genetic Center, was used to estimate the correlation within each family. The SHFS data included 2,964 individuals from 91 families. The size of family ranges from 1 to 113. Coverage probability and relative bias were used to evaluate the GEE performance. Our simulation results showed that GEE with independent correlation structure worked well for continuous data. However, if the assumption of a generalized linear mixed effect model (GLMM) holds true, GEE did not work as well for the frequency and binary data as it produced biased parameter estimates and coverage below the nominal level. We recommend a model selection procedure such as cross-validation should be carried out to determine whether a GEE or a GLMM is a better fit for the data.