In our study we used additive models without considering the genetic architecture of the simulated QTL; however the methods performed well in localising the true simulated QTL. Out of the four methods employed, GENMIX performed comparatively better in QTL detection. It detected all the 8 simulated QTL and 6 were mapped accurately within a 2 cM region of the QTL region. However, it also identified 4 false positives (FP). The number of tests carried out in GENMIX was approximately 7 times the number of markers and we used Bonferroni multiple testing correction for the number of total marker but not for the total number of tests (i.e. ~7 times the number of markers) which could have resulted in increased number of false positives. Besides the number of haplotypes in a lineage goes down as we moved down the tree  which can give numerical instability. Out of these four FPs in GENMIX, two (on chromosome 4 and 5) were identified by other methods at the same location (Figure 1). Divergent approaches of GWAS picking up the same FP could be due to insufficient correction for family structure. A likely explanation is that some SNPs in these two regions were positively correlated (in linkage disequilibrium) with the QTL because of linkage (within family). It is thus not straightforward to distinguish true associations from spurious, regardless of the correction for the pedigree structure. This underlines the importance of replication study before a follow-up study can be taken up for identifying causal mutation underlying a QTL.
BVS was the second best method in terms of power to identify QTL and it had less FP compared to GENMIX. It detected all the simulated QTL except the one on chromosome 5 with epistatic interactions. BVS fits all the SNPs simultaneously and given that the first epistatic QTL was fitted in the model, there was a little chance for the second one to be significant in the model. In other words, the first QTL explains most of the variation induced by both QTL because of their dependency. Especially, this can happen if the epistasis is of additive by additive nature, where most of the epistatic variance is converted to additive . In order to confirm this, we ran the MMA for all SNP on chromosome 5 where the first epistatic QTL was already in the model. As a result, the second epistatic QTL was not detected (results not shown).
The MMA identified six QTL. The two linked QTL on chromosome 2 were both identified by MMA but only the first one (the most significant) was reported in the workshop as the second QTL was not significant when fitted along with first one in the model. On the other hand RHM detected both of them but the first QTL was mapped 10 cM downstream the true QTL.
The highest significant SNP for the multi-allelic QTL on chromosome 1 (largest QTL) in MMA was 0.7 cM away from the true position, while the other methods mapped it closer to its position. No individual SNP (bi-allelic) can be in perfect LD with this QTL (multi-allelic) which might have resulted in poor precision for this QTL in MMA.
The imprinted QTL on chromosome 4 was only detected by GENMIX and BVS. The power of detection of the QTL will decrease if the model does not reflect the true genetic architecture of the QTL. However, GENMIX and BVS methods were sensitive enough to identify the imprinted QTL, though both of them model its effect as additive.
Sahana et al.  observed very high false positives when haplotypes were considered as fixed effects in the model. Because the frequency of some haplotypes can be very low, this could result in low accuracy of estimates and result in false positive when haplotypes are fitted as fixed effect. We expected this problem can be taken care by fitting haplotypes as random where the effects of the low frequent haplotypes will be regressed towards zero. However, RHM still had very high false positive rate.