An Extensive Comparison of Quantitative Trait Loci Mapping Methods

Abstract

Background: The choices of study design and statistical approach for mapping a quantitative trait (QT) are of great importance. Larger sibships and a study design based upon phenotypically extreme siblings can be expected to have a greater statistical power. On the other hand, selected samples and/or deviation from normality can influence the robustness and power. Unfortunately, the effects of violation of multivariate normality assumptions and/or selected samples are only known for a limited number of methods. Some recommendations are available in the literature, but an extensive comparison of robustness and power under several different conditions is lacking. Methods: We compared eight freely available and commonly applied QT mapping methods in a Monte-Carlo simulation study under 36 different models and study designs (three genetic models, three selection schemes, two family structures and the possible effect of deviation from normality). Results: Empirical type I error fractions and empirical power are presented and explained as a whole and for each method separately, followed by a thorough discussion. Conclusions: The results from this extensive comparison could serve as a valuable source for the choice of the study design and the statistical approach for mapping a QT.

Parameter Estimation and Quantitative Parametric Linkage Analysis with GENEHUNTER-QMOD

“…With regard to type I error and power analysis, our results of the VCA and Haseman-Elston regression are, to a large part, consistent with the findings of Kleensang et al [28] . In that simulation study, the true type I error for the VCA also often deviates from the nominal type I error and is unacceptably liberal when the normality as-sumption does not hold.…”

“…A low power under normality violations is described as well. Haseman-Elston regression is shown to be conservative under non-normally distributed data [28] , as we can confirm with our results. The power is described as inferior to the power of the VCA when no ascertainment condition is applied, which is also consistent with our findings.…”

“…Family structure, marker and disease locus as well as number of H 0 replicates for p value calculation remained as in scenarios 1, 2 and 3. Along the lines of Kleensang et al [28] , we imposed the following ascertainment conditions. Scenario 5 contains only families with at least one sibling in the highest quartile of the original distribution (single-proband selection).…”

The Maximum-Likelihood-Binomial method revisited: a robust approach for model-free linkage analysis of quantitative traits in large sibships

“…We decided to compare this method with the original MLB-QTL method, and to use Merlin-regress as a ''gold-standard'' as it is a widely used and has been recently shown to be one of the most powerful approach under a wide range of situations [Kleensang et al, 2010]. We understand that this is a non exhaustive comparison but we adopted this strategy for two reasons.…”

Linkage and association of phospholipid transfer protein activity to LASS4

“…This analysis method requires a prespecifi ed single locus trait model and provides a lod score and p-value. Such fi xed-model analyses are very powerful if the trait model is correctly specifi ed ( 52,53 ) and are unlikely to produce false-positive evidence for linkage when the trait model is misspecifi ed (54)(55)(56)(57), especially in larger pedigrees ( 52 ). Estimation of the QTL model parameters is discussed in the supplementary material.…”