The feasibility of constructing composite scores which will yield pretest measures having all the properties required by the special regression model is explored as an alternative to the single pretest score usually used in student selection for Elementary Secondary Education Act Title I compensatory education programs. Reading data, including Stanford Achievement Test scores, obtained from students in grades 2, 3, 4, and 6 in four school districts are analyzed from a technical and a practical perspective. Mathematics data obtained from students in grades 2 and 3 in one school district are also analyzed. Although composites do not seem to increase the accuracy of estimating the no-treatment expectation in regression designs variables such as teacher ratings and placement in a reading series can be quite reliable. A procedure requiring the simple sums of two or three carefully selected variables has obvious practical advantages over elaborate score transformations. It can yield pretest measures having all the properties necessary for the special regression model. If, however, the scaling of a variable or score inflation on a rating causes floor or ceiling effects in the composite scores, problems are created when project gains are estimated with the special regression model. (RL)