The Impact of Ignoring Clustering Effect in Three-Level Data

• Activity Number: 650
• Type: Contributed
• Date/Time: Thursday, August 7, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Computing
• Author(s): Tina Cunningham*+ and Bob E. Johnson
• Companies: Eastern Virginia Medical School and Vanderbilt University School of Medicine
• Keywords: Three-level Design ; ICC ; Simulation ; Nested Data ; Type I Errors

Abstract

Research with three-level data is common in practice. Although the effect of ignoring the nesting factor in two-level data is well-studied, little has been discussed in the literature concerning the impact of the double nesting effects on Type I errors in three-level data. In this work, a simulation study was conducted to examine the consequence of ignoring one of the two nesting effects or both, under three-level design. We performed the simulation study using five parameters: the number of patients per each physician, the number of physicians per center, the number of centers, the first intraclass correlation measures the homogeneity between patients nested within the same physician, and the second intraclass correlation measures the agreement between patients within the same center. For each combination of parameters, 7300 sets of correlated multivariate normal data were generated and five models were fit (i) all nesting effects are ignored; (ii) all nesting effects are included; (iii) data are analyzed under independent assumption as described in (i), but the results are adjusted with the known design effect; (iv) level two is ignored; and (v) level three