General Departmental Seminar Series
Omitted Variables in Multilevel Models
Jee-Seon Kim, PhD,
Department of Educational Psychology,
University of Wisconsin
November 19, 2004, 12 - 1 pm in room G5/113 Clinical Sciences Center (600 Highland Ave.)
There is a strong and growing interest in multilevel modeling within the social and behavioral sciences. The hierarchical structure of multilevel models allows researchers to hypothesize relationships at various levels of the hierarchy, and thus both study and control for contextual influences on the outcome (e.g., student within class, class within school, and school within district, etc.). This talk provides an introduction to multilevel models as studied in the social sciences, and discusses issues related to omitted variables in a multilevel modeling framework. The omission of variables is a critical model misspecification problem in multilevel analysis. In educational research, essential characteristics of teachers and schools are often unmeasured or inaccessible, and their omission may yield biased and inconsistent regression coefficient estimators at any level in the model. This talk presents a number of robust and efficient estimators and a battery of statistical procedures for testing the severity of omitted variable bias in multilevel analysis. The proposed methods extend the well-known (in econometrics) Hausman test for panel data models and provide several options for studying various forms of misspecification in multilevel models. Methods for omitted variable testing and robust estimation are illustrated through a longitudinal data set involving state test scores collected for students in Dallas elementary schools annually between 1994 and 2000. Simulation studies summarize finite sample properties of these tests and estimators.
Return to seminar list