This paper addresses the problem of heterogeneity of samples of learning disabled (LD) children by comparing five different systems for identifying homogeneous subgroups in terms of their ability to predict longitudinal reading and mathematics scores. One hundred and sixty LD children served as subjects. Three of the five subgrouping systems were based upon the Wechsler Intelligence Scale for Children--Revised (WISC-R), but did not predict achievement. One system based upon clinical diagnostic criteria and one based upon the magnitude of the child's achievement deficit did predict achievement. The former system predicted both reading and mathematics achievement, while the latter system predicted only reading achievement. A pairwise regression analysis demonstrated that most of the predictive variance of the system based upon the magnitude of achievement deficit was accounted for by the clinical diagnostic system. A further analysis of the diagnostic system indicated that a subgroup of LD children who showed evidence of mild socioemotional disturbance performed significantly better at each time of measurement over a 4-year period, and performed at an accelerated rate compared to a subgroup of LD children without signs of a socioemotional disturbance. These results are interpreted as support for a system of subgrouping LD children based upon effective variables. The implications of this system for research and treatment of learning disabilities are considered. A five-page reference list follows the study. (Author/JAZ)