I'm not sure normally distributed liability is so important for GWAS (though it may be for computational modeling of disease). Imagine a disease caused by a single recessive allele at 50% frequency. Disease risk is not normally distributed at all, but the GWAS approach would be fantastically successful. Ditto for perfectly defined phenotypes: imagine a "disease" that is actually two diseases, and that there are common SNPs that influence either one (but not both). Now you just need twice the sample size (more or less) to identify the relevant genes. But it's true knowing the right disease subtypes a priori would be helpful. If diseases are really caused by different rare, partially penetrant mutations in different genes in every individual, you're right that GWAS isn't the best way to look for those mutations. That's a very special model though!

Well, it's not exactly the same model as mental retardation, deafness, etc, right? In that the latter involve mutations of very large effect (such that they're approximately Mendelian), while common diseases don't show that inheritance pattern. And one of the most common causes of deafness is an allele at 1-2% frequency in a gene that accounts (with other mutations) for about 50% of the cases. That's definitely a "rare" allele, but not so rare that it's individual or family specific. (In fact, if you were to do a population-based GWAS of deafness, you might well get that gene). I think people probably think individual/family-specific alleles contribute to disease risk. But to argue that they're *the* source of disease risk seems like quite the leap. Not common variation at all (what is this schizophrenia GWAS picking up, if not common variants)? Not rare variants at 1% frequency?