Comparing Alternatives to Multilevel Analysis When the Location and Extent of Clustering Is Systematically Varied

• In Event: Methodological Issues and Applications in Multilevel Modeling

Abstract

Clustered data analysis is common and multilevel modeling is one of many possible analytic approaches. Many alternatives to multilevel modeling exist for researchers comfortable with ordinary least squares regression. However, the extent to which popular alternatives to multilevel modeling produce unbiased standard errors as the extent and location of clustering within the population model are systematically varied has not been studied. This study focuses on a model with one within-level predictor and a random effect. With few exceptions, Taylor Series linearization produced unbiased standard errors. Fixed-effects modeling is sensitive to clustering present in the relation between the outcome and predictor and produces biased standard errors under most conditions. Unsurprisingly, OLS and the corresponding design effect standard error produced substantial bias.

Authors

• Blaine Pedersen, Texas A&M University
  Presenting Author

• Oiman Kwok, Texas A&M University
  Presenting Author

• Karen E. Rambo-Hernandez, Texas A&M University
  Presenting Author