A two-stage procedure for the analysis of pretest-posttest data was developed and illustrated. Data were reanalyzed from a quasi-experimental evaluation that assessed the effects of an intervention program to teach elementary school children aged 10 to 12 years to watch violence on television more carefully/critically. The treatment group consisted of 221 children, and the control group included 216 children. Use of a non-linear canonical correlation analysis and the CANALS computer program transformed the pretests and posttests so that they became linearly interrelated, and possible deviations from the linear model due to non-linearity were minimized. In a second step, the resulting optimally scaled set of pretest and posttest measures was analyzed by multiple analysis of covariance (MANCOVA) to measure program effects. The variance accounted for by the MANCOVA model appeared to be increased substantially by better prediction of posttest scores, by better estimation of the effect of the treatment, or by both. The advantages of the two-stage procedure are discussed. Five tables and 16 graphs illustrate the study. An appendix provides additional data in tables. (Author/SLD)