A Monte Carlo Study of Confidence Interval Coverage for Variance Components in Multilevel Models

• In Event: AERA Poster Session 4
  In Poster Session: Division D Section 2 Poster Session 1

Abstract

Variance components are central to multilevel models but are known to have biased standard errors for small samples and non-normal distributions of cluster residuals. Statistical software such as SAS, SPSS, and R use different methods to estimate standard errors. This can lead to different confidence intervals about variance components for the same data and model, especially for small samples. In addition, the impact of symmetric and moderately heavy-tailed distributions is largely unknown. A Monte Carlo study of four methods for estimating standard errors and generating confidence intervals about variance components was performed. The findings were used to recommend the method(s) least sensitive to small samples and non-normal distributions of cluster residuals. Recommendations for data-analytic practice are given.

Authors

• Hao Jia, University of Minnesota - Twin Cities

• Yadira Peralta, Centro de Investigación y Docencia Económicas

• Michael R. Harwell, University of Minnesota