Dr. Risch [ 19901 is to be commended for his authoritative investigation and review of the many potential problems associated with linkage analysis of complex diseases. Such a comprehensive report has long been due. Here, I would like to expand on a technique briefly referred to towards the end of the Etiologic Heterogeneity section in that paper, that is, on the practice of maximizing the lod score over different disease definitions.
In his section on phenotypic variation, Dr. Risch discusses the three diagnostic schemes, S, SS, and SSF (from narrow to broad), used by Sherrington et al. [1988], the maximum lod score being 6.5 for SSF vs. 3.2 for S . He notes that an additional lod score of 3.3 was obtained with the inclusion of the non-schizophrenia phenotypes in the SSF scheme. One has to be aware that ''inclusion" here does not refer to adding additional individuals but to a change of affectation status. The non-schizophrenia individuals (considered affected under scheme SSF) were not disregarded in the narrow (S) diagnostic scheme; rather, they were called unaffected. There must be considerable misclassification, under either scheme S or SSF, and perhaps also under SS.
A solution often taken to resolve such misclassification is selecting that diagnostic scheme as the most plausible that leads to the highest lod score. This technique does not seem to be based on a sound scientific principle as it, in effect, maximizes the investigator's evidence for linkage, which is the very thing he hopes to find. It appears unavoidable that this method must lead to an inflated lod score; that is, the maximum lod score observed is associated with a higher P value than in ordinary linkage analyses.
In discussions of these problems, I have been given the response: "But it seems to work." However, what "works" is essentially that the lod score of 3 was exceeded. Unless we know the empirical significance level ( P value) associated with such a find-