Volume 3 Supplement 7
The effect of minor allele frequency on the likelihood of obtaining false positives
© Tabangin et al; licensee BioMed Central Ltd. 2009
Published: 15 December 2009
Determining the most promising single-nucleotide polymorphisms (SNPs) presents a challenge in genome-wide association studies, when hundreds of thousands of association tests are conducted. The power to detect genetic effects is dependent on minor allele frequency (MAF), and genome-wide association studies SNP arrays include SNPs with a wide distribution of MAFs. Therefore, it is critical to understand MAF's effect on the false positive rate.
Data from the Framingham Heart Study simulated data (Problem 3, with answers) was used to examine the effects of varying MAFs on the likelihood of false positives. Replication set 1 was used to generate 1 million permutations of case/control status in unrelated individuals. Logistic regression was used to test for the association between each SNP and myocardial infarction using an additive model. We report the number of "significant" tests by MAF at α = 10-4, 10-5, and 10-6.
Common SNPs exhibited fewer false positives than expected. At α = 10-4, SNPs with MAF 25% and 50% resulted in 69.2 [95%CI: 62.8-75.6] and 70.8 [95%CI: 61.3-80.4] false positives, respectively, compared to 100 expected. Rare SNPs exhibited more variability but did not show more false-positive results than expected by chance. However, at α = 10-4, MAF = 5% exhibited significantly more false positives (105.5 [95%CI: 81-130.1]) than MAF = 25% and 50%. Similar results were seen at the other alpha values.
These results suggest that removal of low MAF SNPs from analysis due to concerns about inflated false-positive results may not be appropriate.
Correct identification of the most promising single-nucleotide polymorphisms (SNPs) for follow-up is one of the greatest challenges of conducting genome-wide association studies (GWAS). Hundreds of thousands of simultaneous association tests are typically conducted across the genome, and the resulting p-values or test statistics are compared and often ranked to determine the most promising SNPs. This procedure is conducted regardless of the specific adjustment chosen for multiple testing correction. For example, the conservative Bonferroni correction to adjust the critical significance threshold simply lowers the threshold of α to account for the number of tests conducted; any SNP with a p-value below that adjusted threshold will be considered significant. False-discovery rate (FDR) procedures explicitly rank p-values before selecting the SNPs below the FDR threshold for further consideration. Genome-wide permutation testing methods involve randomly permuting case and control status and examining the distribution of the best resulting test statistic, thus creating an empirical distribution of extreme test statistics against which to compare observed results. With each of these methods however, if multiple SNPs reach this threshold, the investigator must choose which SNPs to follow for replication.
GWAS SNP arrays necessarily include SNPs with a wide distribution of minor allele frequencies (MAFs), from nearly monomorphic (MAF <0.5%) to very common (MAF≈50%). All of the methods for determining SNP selection implicitly assume that the p-values or test statistics from individual SNPs are drawn from the same distribution, regardless of MAF. If this were true, then ranking the test statistics from SNPs of different MAF would be valid. However, the power to detect a given genetic effect with a given study size depends to a great extent on the MAF of the risk allele tested. Specifically, loci with a low MAF (<10%) have significantly lower power to detect weak genotypic risk ratios than loci with a high MAF (>40%) . Further, previous studies have demonstrated that rare genotypes are more likely to result in spurious findings . Thus, many GWAS have removed SNPs with MAF<10% [3, 4].
It is therefore critical to understand the true effect of MAF on the likelihood of obtaining a false-positive result (e.g., very low p-value) under the null hypothesis of no true association. We hypothesize that MAF affects the likelihood of a specific variant achieving significance thresholds under the null hypothesis. If MAF does affect the probability of achieving a specific significance, this would result in non-comparability of p-values (or test statistics) between SNPs with different MAFs. To address this issue we utilized the Genetic Analysis Workshop 16 simulated data set and performed extensive permutations.
One million permutations were conducted per SNP, in which case or control status was randomly assigned to an individual in the dataset, keeping the total numbers of cases and controls consistent. For each permutation, logistic regression was used to test for the association between each SNP and myocardial infarction event 1 using an additive model. To determine whether the test statistic distribution differs by allele frequency we report the number of false positive "significant" tests by MAF at α = 10-4, 10-5, or 10-6. These threshold alpha values were selected because we could expect at least one false positive by chance using one million permutations. Statistical analysis was conducted using SAS version 9.1.3 (SAS Institute, Cary, NC) on a dedicated server.
While low MAF SNPs (1% or 5%) did not have elevated false positives compared to chance, the 5% MAF SNPs had significantly more false positives than the more common (25 and 50%) MAF SNPs at the 10-4 (p < 0.03) and 10-5 (p < 0.05)thresholds. These differences were not captured by standard summary statistics, but instead, only by examining values at the tail end of the distribution.
The current study provides evidence that false-positive rate is influenced by MAF. We report that common SNPs (MAF 25-50%) result in significantly fewer false positives than expected under the null χ2 distribution, suggesting that for common SNPs the current thresholds may be too conservative. Also, we report more false positives with 5% MAF SNPs than with common SNPs at the thresholds examined. In addition, under the null hypothesis, up to 5% randomly missing data did not impact our findings. Thus, under the null hypothesis, MAF but not proportion of missing data appears to affect the likelihood of a false positive result.
These results have implications for a common analytical approach applied to large-scale SNP data. GWAS analysis often begins by discarding all genotypes for SNPs of low MAF, typically for those with MAF<10% [3, 4]. This practice results in a huge loss of data. For example, 35.5% of SNPs on the Affymetrix chips have MAFs below 10%. In addition, discarding low MAF SNPs may hamper our ability to detect rare disease-causing polymorphisms .
The reasons for discarding these SNPs may be related to genotyping issues, such as lower genotyping rates or concerns about calling accuracy. However, discarding low-MAF SNPs may also be due to perceptions about the statistical inferences that result from analyzing such SNPs. Anecdotally, investigators have expressed "distrust" of significant results arising from low-MAF SNPs. However, a recent study demonstrated that in the absence of bias in genotype error rates between cases and controls, even very small MAFs show type I error rates close to nominal levels . Our study confirms and extends this report, showing that nominally significant results occur significantly less often than expected, even for low MAF SNPs, resulting in a conservative bias. Thus, our findings do not support the removal of low MAF SNPs from analysis due to concerns about inflated false-positive results.
However, our study does raise the concern about ranking p-values generated from SNPs across vastly different MAF. Given our findings, we suggest that the χ2 threshold to establish a particular level of significance may need to be modified using empirical permutations for several MAF "bins." For example, in our analysis dataset, for an MAF of 25% and a significance threshold of 10-5, the threshold χ2 value would need to be reduced from 19.51 (from the standard χ2 distribution) to 18.79 to capture the expected 10 false positives. However, it should be noted that to determine robust estimates, χ2 adjustments, and false positive rates for a GWAS significance of 10-8, over 10 billion permutations and significant computing capability would be required. Additional studies using larger datasets will be required to fully understand this issue.
List of abbreviations used
False discovery rate
Genome-wide association studies
Minor allele frequency
We thank Dr. Michael Wagner and Mr. Prakash Velayutham for assistance with SNP selection, and Ms. Karen Noel for help in optimizing the SAS server. This work was supported by the Fifth-Third Charlotte M Schmidlapp Award (LJM) and NIH NS36695 (LJM and JGW). The Genetic Analysis Workshops are supported by NIH grant R01 GM031575 from the National Institute of General Medical Sciences.
This article has been published as part of BMC Proceedings Volume 3 Supplement 7, 2009: Genetic Analysis Workshop 16. The full contents of the supplement are available online at http://www.biomedcentral.com/1753-6561/3?issue=S7.
- Ardlie KG, Lunetta KL, Seielstad M: Testing for population subdivision and association in four case-control studies. Am J Hum Genet. 2002, 71: 304-311. 10.1086/341719.PubMed CentralView ArticlePubMedGoogle Scholar
- Lam AC, Schouten M, Aulchenko YS, Haley CS, Koning D-J: Rapid and robust association mapping of expression QTL. BMC Proc. 2007, 1 (Suppl 1): S144-10.1186/1753-6561-1-s1-s144.PubMed CentralView ArticlePubMedGoogle Scholar
- Cupples LA, Arruda HT, Benjamin EJ, D'Agostino RB, Demissie S, DeStefano AL, Dupuis J, Falls KM, Fox CS, Gottlieb DJ, Govindaraju DR, Guo CY, Heard-Costa NL, Hwang SJ, Kathiresan S, Kiel DP, Laramie JM, Larson MG, Levy D, Liu CY, Lunetta KL, Mailman MD, Manning AK, Meigs JB, Murabito JM, Newton-Cheh C, O'Connor GT, O'Donnell CJ, Pandey M, Seshadri S, Vasan RS, Wang ZY, Wilk JB, Wolf PA, Yang Q, Atwood LD: The Framingham Heart Study 100 k SNP genome-wide association study resource: overview of 17 phenotype working group reports. BMC Med Genet. 2007, 8 (Suppl 1): S1-10.1186/1471-2350-8-S1-S1.PubMed CentralView ArticlePubMedGoogle Scholar
- Florez JC, Manning AK, Dupuis J, McAteer J, Irenze K, Gianniny L, Mirel DB, Fox CS, Cupples LA, Meigs JB: A 100 k genome-wide association scan for diabetes and related traits in the Framingham Heart Study: replication and integration with other genome-wide datasets. Diabetes. 2007, 56: 3063-3074. 10.2337/db07-0451.View ArticlePubMedGoogle Scholar
- Gorlov IP, Gorlova OY, Sunyaev SR, Spitz MR, Amos CI: Shifting paradigm of association studies: value of rare single-nucleotide polymorphisms. Am J Hum Genet. 2008, 82: 100-112. 10.1016/j.ajhg.2007.09.006.PubMed CentralView ArticlePubMedGoogle Scholar
- Moskvina V, Craddock N, Holmans P, Owen MJ, O'Donovan MC: Effects of differential genotyping error rate on the type I error probability of case-control studies. Hum Hered. 2006, 61: 55-64. 10.1159/000092553.View ArticlePubMedGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.