Is GEE always valid? – Exploring the discrepancies between the score test results and least square resultsView Presentation Stewart J Anderson , University of Pittsburgh, Graduate School of Public Health, Department of Biostatistics
Michael Dunbar, University of Pittsburgh, Department of Psychology
*Xiaoxue Li, University of Pittsburgh, Graduate School of Public Health, Department of Biostatistics
Sarah Scholl, University of Pittsburgh, Department of Psychology
Saul Shiffman, University of Pittsburgh, Department of Psychology
Keywords: GEE, score test, least square means
Generalized estimating equations (GEE) provide a powerful statistical method for analyzing longitudinal binary outcomes. In GEE, a score test is often the preferred tool for making inference. However, we found that score test results are sometimes inconsistent with least square means. We demonstrate this phenomenon in an analysis comparing the difference between intermittent smokers (ITS) and daily smokers (DS) in their smoking behavior by the day of week (DOW). We modeled the probability of smoking as a function of DOW, smoker type, and the interaction between the two. We hypothesized that smoker type is a modifier of DOW in explaining the probability of smoking. The score test gave non-significant interaction results. However, by examining the model based probabilities of DOW, ITS tended to smoke much more on Fridays and Saturdays than on other days of the week whereas DS had a consistent pattern of smoking on all days. The inconsistency may be due to the magnitude of the group x DOW interaction effect being much smaller than the mean probability differences between ITS and DS, thus diminishing the sensitivity of the score test to detect differences in smoking patterns.