The concept under investigation was the reliability of estimates of mean scores of groups under various assumptions of multiple-matrix sampling when reliabilities are computed according to procedures based on generalizability theory. Four different cases were compared with respect to the generalizability coefficients depending upon whether pupils and/or items were assumed to have been sampled from finite or infinite populations. In the experimental design, schools were nested within districts, students were nested within schools, items were nested within forms, and forms were crossed with schools within districts. Tests of reading, written expression, and mathematics were used in the study for students in grade 6 and in grade 12. Each test consisted of from 128 to 198 items, allocated without duplication into 16 or 18 approximately parallel forms with 8 to 11 items per form. Thirty-one school districts were selected for the sixty grade sample and 33 for the twelfth grade. Tables present the random effects analysis of variance design, the estimated variance components for each grade, and the resulting generalizability coefficients. (Author/CTM)