Starting off the day on complex traits. @asiepel assures us he's not @BeEngelhardt but she's late because she's voting! First talk is David Balding #probgen18
DB: begins by explaining LD score regression. Intuition: the more LD a SNP has, the more reach it has, and so test statistics will be bigger. Slope is measure of heritability. Points out that there's binning that can be misleading #probgen18
DB: LD score regression is easy and uses summary statistics so it's very popular. But it's "sort of wrong", slope is not good measure of heritability and intercept is not good measure of confounding #probgen18
DB: issue is that each SNP is weighted only by LD, meaning implicitly that all SNPs are the same but we don't think that's true. Note that SNPs with high LD are likely to take multiple causal SNPs #probgen18
DB: we have prior information about heritability from properties of the SNP (e.g. MAF). Shows different types of LD structure that "seem" like they should have different heritability but LD type methods treat them the same. #probgen18
DB: used to do GWAS one SNP at a time, but now do whole genome. But we have too many predictors (SNPs) relative to our sample size. Usually do some kind of shrinkage to control sparsity. Claim is that gaussian shrinkage is a bad prior and we can construct better priors #probgen18
DB: new model keeps computational efficiency of Gaussian prior, but uses weights to down weight SNPs that are highly redundant. This results in throwing out ~90% of the data! Recently introduced a new parameter to account for MAF #probgen18
DB: aside that standardizing genotypes is a bad idea because it implicitly assumes a relationship between MAF and effect size #probgen18
DB: use simulated data to check model. Turns out that the method that works best is the one that the data was simulated under! So need to check against data. Do a mixture model to see which is preferred. Sims show this works for model choice #probgen18
DB: application to real data results in strong preference for the new model. Some data suggests the hybrid model is better, but this may be a result of QC issues. New method estimates higher heritability, hopefully because picking up more signal in the data #probgen18
DB: LD score originally introduced to deal with confounding. But we think GWAS over corrects for confounding, and you can "reflate" test statistics and get more hits using this method #probgen18
DB: two models give very different results about where heritability is. "Our results are boring, kind of a heat death". Cf LD score results which show more interesting genomic features #probgen18
