Educational research on the factors of student achievement has been limited by its failure to consider the multilevel or hierarchical nature of most data. This study used a nonexperimental regression-based procedure, hierarchical linear modeling (HLM), to empirically develop a predictive model of fifth-grade achievement in reading and mathematics for a statewide data set at both the individual student and school district levels. The database was comprised of reading and mathematics achievement test scores of 86,227 elementary students in Pennsylvania who were enrolled in third grade in 1986 and in fifth grade in 1988. Findings indicate that only a small portion of the variability in individual achievement is potentially explainable by district-level factors. HLM was also used to identify district-level factors that explain the variation in district mean achievement and within-district relationships. For example, a small effect of class size was revealed in increasing the within-district relationship between prior ability and student achievement. These results permit the formulation of a wider range of policy inferences than is possible with conventional regression analyses. One figure and six tables are included. (27 references) (LMI)