ISM Research Memorandum
No.
1202
Title:
Re-assessment of Multiple Testing Strategies for More Efficient Genome-wide Association Studies
Author(s):
Otani, Takahiro (The Institute of Statistical Mathematics);
Noma, Hisashi (The Institute of Statistical Mathematics);
Nishino, Jo (Nagoya University Graduate School of Medicine);
Matsui, Shigeyuki (Nagoya University Graduate School of Medicine)
Key words:
multiple testing; genome-wide association study; statistical power; familywise error rate; false discovery rate
Abstract:
Although enormous costs have been dedicated to discovering relevant disease-related genetic variants, especially in genome-wide association studies (GWASs), only a small fraction of estimated heritability can be explained by these results. This is the so-called missing heritability problem. The conventional use of overly conservative multiple testing strategies based on controlling the familywise error rate (FWER), in particular with a genome-wide significance threshold of P<5x10^-8, is one of the most important issues from a statistical perspective. To help resolve this problem, we performed comprehensive re-assessments of current available strategies using recently published, extremely large-scale GWAS datasets of rheumatoid arthritis and schizophrenia (>50000 subjects). The estimates of statistical powers averaged for all disease-related genetic variants of the standard FWER-based strategy were only 0.09% for the rheumatoid arthritis data and 0.04% for the schizophrenia data. To design more efficient strategies, we also conducted an extensive comparison of multiple testing strategies by applying false discovery rate (FDR)-controlling procedures to these datasets and simulations, and found that the FDR-based procedures achieved higher powers than the FWER-based strategy, even at a strict FDR level (e.g., FDR=1%). We also discuss an alternative useful measure, namely gpartial power,h which is an averaged power for detecting the clinically and biologically more meaningful genetic factors with the largest effects. Simulation results suggest that the FDR-based procedures can achieve sufficient partial power (>80%) for detecting these factors (odds ratios of >1.05) with 80000 subjects, and thus this may be a useful measure for defining realistic objectives of future GWASs.