The site variability issue for Direct Instruction Follow Through sites is examined from a variety of perspectives--meta-analysis of the Abt Associates' data, norm-referenced comparisons, and a longitudinal look at stability of effects. Much of the seemingly unexplained variability found in both the Abt report and the reports by M. Kennedy and E. House for Direct Instruction was due to: (1) varying degrees of appropriateness/inappropriateness of the local comparison groups selected, and (2) the inclusion of one site that was not implementing the model. The demographic variables for Follow Through and comparison group indicated two severe mismatches, Tupelo and West Iron County. If two sites and Grand Rapids are eliminated, much of the unexplained variability vanishes. There do appear to be some sites that are more effective than others, yet only one site (Racine) demonstrates weak, negative effects. It appeared that the use data from the pooled comparisons might present a less biased picture--especially in areas such as Reading where home and entry characteristics are fairly heavily weighted. Regression models could be used to derive a predicted or expected score for students in the experimental sample, based on demographic characteristics and pretest information. (Author/PN)