LD score regression distinguishes confounding from polygenicity in genome-wide association studies