The Flemish part of Belgium was one of the participants in TIMSS 2003 (Trends in International Mathematics and Science Study). Besides the international data collection (which included a mathematics and science test for the students and questionnaires for the students, teachers and principals), in Flanders there were additional questions for all the questionnaires and a numerical and spatial intelligence test for the students in secondary schools and a supplementary questionnaire for their parents. In Flanders, the TIMSS sample design differed from the general design: instead of taking one class in each of the selected schools, two classes per school were selected. Because of this design, a multilevel model with an intermediate level could be implemented. Whereas in other countries only a division between one higher level (the school level, which is at the same time the teacher and the class level) and a lower level (student level) can be made, in Flanders the higher level can be subdivided in a school and a class level. The Flemish dataset contains valid data of more than 5000 students in almost 300 classes and about 140 secondary schools. In an 'empty'-model, this is a model without the explanatory variables, the variance of the scores on the mathematics test can be divided. In other words, it can be determined how much of the total variance is due to the student level, how much is due to the class level and how much to the school level. To explain the variance on the distinguished levels, different variables derived from the (extended) questionnaires can be entered in the model. In this paper, we examine particularly the effect of the extent to which the teachers create a constructive learning environment. Research indicated that group composition, learning environment (i.c. constructivism) and class climate characteristics have significant effects on the achievement. An important question concerns the relationship between group composition and learning environment: 'Do more able class groups get better education?' The results of the multilevel analyses with the explanatory variables at the different levels will be presented.