Extending the application of multilevel modeling to data from the Third International Mathematics and Science Study (TIMSS

Now that TIMSS data have been gathered, researchers have begun to analyze the data and produce scholarly work on the contexts that affect student, classroom, and school outcomes. Many analysts have turned to multilevel modeling techniques to model these outcomes. TIMSS data lends itself well to the application of multi-level modeling. However, the inherent nature of the TIMSS sample design in the majority of TIMSS countries, limits the complexity of the multilevel models that can be formulated.;Only a two-level model can be formulated in the majority of TIMSS countries because the general sampling procedure involves sampling only one classroom per target grade in each school. All countries followed the general sampling procedure except the United States, Australia, Cyprus, and Sweden, which sampled two classrooms per grade in each school. While the two-level model is informative, it cannot accurately partition the total variance in the outcome into its within-classroom, between-classroom, and between-school components; the between-school and between-classroom variance will always be confounded.;This research presents an innovative procedure that uses achievement data from classrooms in adjacent grades, to allow researchers to extend the application of multilevel modeling beyond the two levels presently permitted by the TIMSS sample design, and validates its use in the countries that sampled two classrooms per grade as well as previous studies. It will allow researchers to modify the TIMSS sample in order to simulate a two-classroom per grade sample design. This modification, referred to as the pseudo-classroom procedure, allows the formulation of three-level models that can partition and model the between-classroom variance separately from the between-school variance.;Using mathematics achievement for TIMSS Population 2, this research details how the pseudo-classroom procedure can effectively model the three-level variance structure in TIMSS countries. In addition to partitioning the variance structure, this research also shows that the pseudo-classroom procedure can model the covariance structure when context variables are used to predict mathematics achievement in the United States.