Link Functions in Multilevel Modeling for Binary Data

• In Event: 29.036 - Marvelous Multilevel Modeling

Abstract

This study examined the performance of three link functions (logit, log-log, and complementary log-log) in generalized multilevel modeling for binary data. A skewed binary outcome distribution was created by manipulating five different proportions of ones along with four combinations of level-1 and level-2 sample sizes. The performance of each link function was evaluated with relative bias, model fit, 95% coverage rates, and power. The findings indicated that the logit link showed the least amount of relative bias in most conditions; however, the log-log link showed better model fit and performed as well as the logit link in terms of coverage rates and power, while maintaining the acceptable level of relative bias, under the proportion conditions of 50% to 80%.

Authors

• HyeSun Lee, University of Nebraska - Lincoln

• Lesa Hoffman, The University of Kansas