I am working on a GWAS dataset containing 920 individuals with genotype information on ~1.5M SNPs (genotyped on Illumina 2.5omni chip; no imputed SNPs). I am testing several different phenotypes in this dataset, but everytime I run an association analysis (logistic regression), my QQ plots look deflated, i.e. the P values are systematically less significant than the expected distribution. I have tested several different phenotypes, and get very similar results every time.
I tried running an association analysis using randomly generated phenotype labels (equal number of cases and controls), and the QQ plot is still deflated:
I have tried the following solutions:
- adjustment for PCA coordinates
- increasing minor allele frequency threshold to only include SNPs with maf>0.05
- LD-based pruning down to 240.000 SNPs
None of which remove the deflation.
I searched for literature on this, and while some studies report this type of QQ plot, none seem to actually address the reason for the deflation. From what I can tell, deflation is known to occur in studies using imputed SNPs or data on copy number variation, but should not be seen in a dataset such as mine.
Does anyone know why my plots look like this, and whether this reflects a problem in data?
Best Answer
This is an answer to the old question, but many people still run into this issue.
Ok, so I guess the answer is that if your QQ plot looks like that, and the distribution of -log(P) values is linearly related to the expected values to some point, than the effects of the loci are truthfully lower than expected from the normal distribution. This is expected in highly quantitative traits, and directly related to the trait heritability. However, you can apply some less stringent provisional threshold (usually 3 or 4). According to the results of Bian and Holland (2017), the predictive abilities of the loci cut-off with threshold of 4 and the Bonferoni threshold were not notably different in poligenic traits.