The Implications of Misspecifying Cluster Membership When Robustly Estimating Standard Errors to Account for Correlated Data

• Activity Number: 659
• Type: Contributed
• Date/Time: Thursday, August 2, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Manisha Desai*+ and Susan Bryson and Thomas Robinson
• Companies: Stanford University and Stanford University and Stanford University
• Keywords: robust estimators ; sandwich estimators ; clustered observations ; intracluster correlation ; correlated data

Abstract

The use of robust estimators (REs) of standard errors for handling clustered data has increased steadily over the years. For well-defined clusters, REs are robust to misspecification of the correlation structure. We conducted a simulation study to investigate the sensitivity of results to assumptions about clustering membership. REs were biased when clustering membership was misspecified. REs that assumed clustering of independent data led to type I error rates of up to 40%. REs assuming partial and complete misspecification of membership (where some and no knowledge of true membership were incorporated into assumptions) for data generated from a large number of clusters (50) with a high ICC (0.50) yielded type I error rates that ranged from 8%-12% and 18%-52%, respectively; assuming independence gave a type I error rate of 17%. When the ICC was weak (=0.01), nominal levels of type I error were achieved when clustering was ignored, whereas REs that even partially misspecified membership led to inflated type I error rates. Based on our findings, we provide practical guidelines for the application of REs in the presence of clustering.