P-value multiple-testing adjustment in the context of correlated tests and association mapping

PEÑA MALAVERA, A.; BALZARINI, M.; GUTIERREZ, L.

Florencia

Conferencia; XXVIIth International Biometric Conference; 2014

International Biometric Society

Linear Mixed Models (LMM) have successfully been used to evaluate marker-trait associations in Genome-Wide Association Studies (GWAS). They take into account population structure originated by the correlation among genotypes. In GWAS there are as many hypothesis tests as markers examined; therefore, corrected p-values are commonly used in these studies. Some of the multiple-testing corrections used in QTL studies are Bonferroni (B), Benjamini and Hochberg (BH), Benjamini and Yekutielli (BY) for independent tests, and Li and Ji (LJ) for correlated tests. However, these adjusting procedures work reasonably well in classical QTL studies where individuals are equally correlated. In GWAS with population structure and unequal relationships among individuals, these methods might not perform appropriately. We propose to correct for multiple testing based on the number of effective tests (similar to LJ) but accounting for population structure (called Modified Li and Ji: MLJ). The LJ method proposes to correct for multiple testing by estimating the number of effective tests using the number of significant axis from a singular value decomposition of the correlation matrix across markers. We propose to use a modification of this approach by using the marker regression coefficients corrected by population structure to account for linkage disequilibrium instead of the un-corrected correlation matrix. In this work different multiple-correction methods are compared in a simulated data-set. Two models were previously adjusted: 1) a regression model without correction for population structure (i.e. naive model), and 2) a regression model including a set of principal components, measuring structure as covariables with random coefficients (LMM). The methods were compared based on the false positive (FP) rate, False Discovery Rate (FDR) and the power (φ) to detect QTL. Two population structure scenarios were used (high population structure, FST=0.11; and low population structure FST=0.04). All methods produced lower FP and kept φ when population structure was previously fitted. As expected, B showed the best FP rate, but the lowest power. Meanwhile, MLJ performed better than BH and BY in both scenarios yielding smaller FP (0.10) compared to 0.34 without correcting for multiple testing or 0.24 and 0.12 using a BH or BY, respectively. Additionally, MLJ produced smaller FDR than LJ. With high genetic divergence the reduction in the number of hypothesis tests (96%) was stronger than the decrease (93%) in the number of test for the scenario with low genetic divergence.