Five structural, four linguistic and twelve topic variables are used in regression analyses on results of a 50-item achievement test. The test items are related to 12 topics from the third-grade mathematics curriculum. The items reflect one of two cases of the structural variable, cognitive level; the two levels are characterized, inductive (generalization) or algorithmic thinking. Fourth- and fifth-grade students (N=120) served as subjects. Three regression analyses were carried out. The first used the 12 topic variables. Geometry was the only significant topic variable entered in the final step (p "greater than" .05, R = .5347). The second analysis added Cognitive Level to the topic variables. Cognitive Level entered first and was significant (p "greater than" .01) along with Geometry (p "greater than" .05) at the last step. At the sixth step Cognitive Level, Geometry and Fraction were significant and accounted for 39 percent of the variances in test scores. This was a 13 percent increase over step six in the first analysis. In the last analysis use those entered in the first three steps of either of the previous analyses plus the linguistic variables. Both Cognitive Level and Geometry continued to be significant in the presence of the new variables. (JP)