Aggregating Level 1 Covariates in Two-Level Contextual Analysis: Problem With the Group-Mean Approach and Exploration of Alternative Methods

• In Event: 61.087 - Poster Session 12
  In Poster Session: 61.087-4 - Multilevel Modeling: Complex Survey Data and Missing Data

Abstract

A popular practice in HLM is to average the scores of a level-1 covariate by cluster and create a level-2 covariate. This practice, however, may lead to substantial biased coefficients estimates at level-2, especially when group size is small.

This study aims to: 1. examine how aggregating level-1 covariates influences the bias of contextual effects; and 2. compare the effectiveness of four alternative methods (two-stage estimation, latent covariance model, post-hoc adjustment, and Bayesian approach) in correcting these biases.

Simulation found that, under the current conditions studied, latent covariance model and Bayesian approach generally reduced bias, while post-hoc adjustment worked well when ICC(X) can be reliably estimated. The complete results will provide practical guidelines to handle bias arising from aggregated covariates.

Authors

• Hui Jiang, The Ohio State University

• Robert Nichols, The Ohio State University

• Susan Anderson Mauck, The Ohio State University - Columbus

• Ann A. O'Connell, The Ohio State University