EcoSta 2019: Start Registration
View Submission - EcoSta2019
Title: Estimation accuracy in analyzing multilevel data with extremely unbalanced design and highly correlated structure Authors:  Hsiu-Ting Yu - National Chengchi University (Taiwan) [presenting]
Abstract: Hierarchical Linear Models (HLM) have been widely used as the data analytical methods in psychological research to deal with the dependency naturally existed in data with a multilevel structure. However, several important methodological issues in HLMs are still lack of systematic and comprehensive investigations. Two structural properties in multilevel data possibly found in empirical research are studied: extremely unbalanced and highly correlated data structure. Systematic Monte Carlo simulation studies are conducted to examine the accuracy of parameter estimates in fixed- and random-effects. Factors examined include the number of groups, the mean group-sizes, patterns of group-sizes, and degrees of dependency in data. Results suggest that the unbalanced data structure compensate the inaccuracy in the estimates of fixed-effects parameters under the same total sample size. The unbalanced data structure also has more impact on the stability of estimates for random-effects parameters. Moreover, the number of groups affects the accuracy of parameter estimation more than the sizes of the group. Higher degrees of data correlation also lead to more underestimation of model parameters. The effects of extremely small samples are also analyzed and compared. Complete findings and possible implications will be reported and discussed.