Impact of genotyping errors on the type I error rate and the power of haplotype-based association methods

Abstract

Background: We investigated the influence of genotyping errors on the type I error rate and empirical power of two haplotype based association methods applied to candidate regions. We compared the performance of the Mantel Statistic Using Haplotype Sharing and the haplotype frequency based score test with that of the Armitage trend test.Our study is based on 1000 replication of simulated case-control data settings with 500 cases and 500 controls, respectively. One of the examined markers was set to be the disease locus with a simulated odds ratio of 3. Differential and non-differential genotyping errors were introduced following a misclassification model with varying mean error rates per locus in the range of 0.2% to 15.6%.

Results: We found that the type I error rate of all three test statistics hold the nominal significance level in the presence of non-differential genotyping errors and low error rates. For high and differential error rates, the type I error rate of all three test statistics was inflated, even when genetic markers not in Hardy-Weinberg Equilibrium were removed. The empirical power of all three association test statistics remained high at around 89% to 94% when genotyping error rates were low, but decreased to 48% to 80% for high and non-differential genotyping error rates.

Conclusion: Currently realistic genotyping error rates for candidate gene analysis (mean error rate per locus of 0.2%) pose no significant problem for the type I error rate as well as the power of all three investigated test statistics.

Figure 1

Estimated haplotype distributions. Number of estimated haplotypes and their distribution according to haplotype frequencies for different error rates. The number of rare haplotypes is increased with increasing mean error rate per locus.

Figure 2

Results on Type I error rate. Type I error rate of the Armitage trend test and the Mantel Statistic Using Haplotype Sharing with no (dots), differential (triangulars) and nondifferential (squares) genotype errors incorporated following the unrestricted misclassification model for different error rates per locus. With a small error rate of a) 0.2% both pointwise tests hold the nominal significance level of 0.05 (horizontal line), but with a higher error rate of b) 8% or c) 15.6% the type I error rate is dramatically increased for differential genotyping errors.

Figure 3

Results on Type I error rate in dependence on sample size. Type I error rate of the Armitage trend test (dotted lines) and the Mantel Statistic Using Haplotype Sharing (dashed lines) in the presence of differential genotyping errors (5% in cases, no errors in controls) for different sample sizes (100, 200, 300, 400, 500 case-control pairs).

Figure 4

Results on empirical power. Empirical power of the Armitage trend test and the Mantel statistic Using Haplotype Sharing for no (dots), differential (triangulars) and nondifferential (squares) genotype errors incorporated following the unrestricted misclassification model for different error rates per locus. With a small error rate of a) 0.2% both pointwise tests achieve a power of around 90–100%. With higher error rates of b) 8% or c) 15.6% the power is reduced for nondifferential genotyping errors.

Figure 5

Results on Type I error rate regarding deviations from HWE. Type I error rate of the Armitage trend test and the Mantel Statistic Using Haplotype Sharing with no (dots) and differential (triangulars) genotyping errors (mean error rate per locus 15.6%). The type I error rate is only slightly decreased, when markers not in HWE are excluded from the analysis.