Binary outcomes are very common in medical and epidemiological studies. When scores are independent, logistic regression models are typically used. When outcomes are dependent, generalized estimating equations (GEE) methods are often used to analyze the correlated binary data. For any modeling procedure, an essential step is to determine goodness-of-fit (GoF) for the final model. For logistic regression, standard GoF tests have been developed and are available in most statistical software. The GoF statistics for the GEE method, however, have been developed more recently, and have not been incorporated into currently available software. Several GEE GoF statistics have been proposed (Barnhart & Williamson 1998; Horton et al. 1999; Pan 2002). The objective of this study was to compare these GEE GoF statistics using simulation data under different conditions. Sample sizes were varied, as were the possible covariates (discrete or continuous). Different models included time-dependent and time-independent covariates, quadratic components and interactions. Two or three scores per subject were generated with various correlations between scores No single GEE GoF statistic performed best under all conditions. All of them were poor at detecting the omission of the main effect for a binary time-dependent variable. Pan's statistics had the best performance in three cases: (1) detecting the omission of the interaction for binary time-dependent and time-independent covariates with two or three time-points; (2) detecting the omission of the interaction for a time-independent dichotomous variable and a time-dependent continuous variable with three time-points for exchangeable or auto-regressive correlated data; and (3) detecting the omission of a quadratic term in a time-dependent continuous variable with two time-points. Barnhart's statistics were also the most powerful statistics in the following three cases: (1) detecting the omission of the interaction for a time-independent dichotomous variable and a time-dependent continuous variable; (2) detecting the omission of the interaction for a time-independent dichotomous variable and a time-independent continuous variable; and (3) detecting the omission of the interaction for two continuous variables for two time-points. Horton's statistics had better performance in examining the omission of interaction for a time-independent dichotomous variable and a time-dependent continuous variable with three time-points independent data / acase@tulane.edu
| Identifer | oai:union.ndltd.org:TULANE/oai:http://digitallibrary.tulane.edu/:tulane_24268 |
| Date | January 2004 |
| Contributors | Lin, Hui-Yi (Author), Myers, Leann (Thesis advisor) |
| Publisher | Tulane University |
| Source Sets | Tulane University |
| Language | English |
| Detected Language | English |
| Rights | Access requires a license to the Dissertations and Theses (ProQuest) database., Copyright is in accordance with U.S. Copyright law |
Page generated in 0.0982 seconds