The logic of using a gain score approach versus longitudinal causal models is studied in this secondary analysis of a complex data base. The gain score model used by the Federal Reserve Bank and the School District of Philadelphia in their "What Works in Reading?" study is successively refined using the LISREL structural equation program. First the Philadelphia data base is described and then difficulties of using gain score models are discussed. Regression estimates of the different models are described. Procedures dealing with identification, specification, and collinearity are examined. A sensitivity analysis of measurement and specification error shows the degree to which estimated parameters are affected by researchers' assumptions. The reanalysis shows improvements in the understanding of achievement test data and the logic of how to analyze data bases with longitudinal dependent variables. (Author)