The regression-based discrepancy definition of learning disabilities has been suggested by Rutter and Yule as an improvement of the well-known and much criticized achievement-intelligence discrepancy definition, whereby the examinee's predicted reading attainment is substituted for the intelligence score in the discrepancy expression. Even though the regression-based discrepancy definition has been with us for more than 30 years, critical examination of this approach is scarce. This article fills this lacuna by examining the implications of two variables in the model on the diagnosis of learning disabilities: (a) the effect of predictive validity on the proportion of examinees identified as learning disabled, and (b) the effect of the predictor's identity on the identity of the examinees diagnosed with learning disabilities. Implications of these effects concerning the validity of the regression-based discrepancy model and of the results of its implementation are discussed. (Contains 5 figures and 2 notes.)