The proposed CLD test confirmed 280 out of 1000 causal SNP at a p-value of 0.05 (231 when the QTL effect size was reduced). The power of the CLD test is thus 23-28% and is much lower than when the SNP are used to detect QTL-SNP associations. This relatively low power reflects the fact that proving that a SNP is in complete LD is more difficult than showing that it is merely associated with the QTL. Thus, as previously reported , avoiding false discoveries results in lower power when trying to confirm causal SNP. Reducing the size of the QTL effect did not affect dramatically the power of the test, indicating that other factors, such as the LD structure in the region, are more important to the power of the test. The stringent threshold for the CLD test is the result of strong LD between the SNP in these data. Thus, the CLD test accounts for the background LD when trying to distinguish complete LD from associated SNP.
An alternative approach to find the causative SNP is the concordance test  in which the candidate SNP are genotyped in the parents of the families involved in the linkage mapping design. For this test, the QTL genotypes of the parents should be based on many offspring and be quite certain. If the SNP genotypes agree with that of the inferred QTL genotypes, it provides evidence for the SNP being causative. However, if a SNP is in strong LD with the QTL, the SNP genotypes are also expected to agree with the QTL genotypes, especially when there are only a few parents with 'almost' certain QTL genotypes. For example, in a coat colour mapping study in dogs, 37% of the candidate genes past the concordance test . Moreover, if some of the QTL genotypes are wrongly inferred, this test results in a type-I-error . The data used in this paper did not have the structure of a linkage mapping study, and thus QTL genotypes could not be inferred with high accuracy. The current data resembled that of an association study and, thus, the presented approach is suited to follow-up upon GWAS results.
The test-statistic of the CLD test is based on the assumption that, under the null-hypothesis distribution, the best SNP explains more variance of the phenotype than the second best, whereas under the alternative hypothesis the best SNP is also alike-in-state with the QTL and explains much more of the phenotypic variance. Based on the average log-likelihood values for all 1000 simulations, the difference between the best and second best SNP in these data is ~17 log-likelihood units when the causative SNP is included and only ~0.5 log-likelihood units when the causative SNP is excluded from the analysis. However, the variance between replicates is large, leading to a relatively low power when all replicates are evaluated.
In GWAS, isolated significant SNP are often distrusted, because none of the neighboring SNP confirms the presence of a QTL. In such a case of an isolated significant SNP, the CLD test would provide a positive result since its signal is so much higher than that of neighboring SNP. Here, we assumed that the previous QTL mapping study unequivocally detected a QTL in the studied region, so that regions with spurious significant SNP will not be subjected to the test.
QTL mapping cannot distinguish between a causative SNP and a SNP that is in perfect LD with the causative SNP . Thus, if two SNP are found with equally high log-likelihood values, it is not clear which of the SNP is the causative mutation, and the CLD test statistic would be zero and should not be performed. The latter effect of having a low CLD statistic if one or more SNP are in very high LD with the causative SNP appears to protect the CLD test from pointing to non-causative SNP when the causative SNP is included in the analysis. This is demonstrated by the result that none of the 22 and 44 incorrectly positioned significant SNP in Table 1 are confirmed by the CLD test.
Since higher TCLD-statistics were found for SNP with a low r2 with their nearest marker, we investigated the effect of SNP density on the power of the test. Here, we considered the highest possible density, namely sequence data, which is becoming increasingly available. We reran 1000 "ms"-simulations as described in the Methods section, but retained all the SNP that resulted from the simulated mutations. This resulted in an average of 470 SNP in the 2 cM segment, with an average r2 between adjacent markers of 0.12. The average r2 was rather low due to the often low MAF, but for 6% of the marker pairs r2 was equal to 1. The SNP closest to the middle of the 2 cM segment was designated as the QTL and an environmental effect sampled from N(0,0.5) was added to obtain phenotypes. Out of 1000 replicates, 545 had a single most significant QTL, and 402 of these had the QTL correctly identified. Out of these 402 replicates, 63 had a significant TCLD statistic (P < 0.05), resulting in a power of 16% (= 63/402). Thus, the power was substantially reduced if the marker density was increased to that of sequence data, but some level of power remained. Again none of the misplaced QTL positions passed the CLD test.
The fact that high marker densities, such as in sequence data, results in a reduction of the power of the test, may suggest that removing some SNP from the data (obviously not the putative causative SNP) will improve power. However this invalidates the CLD test, since the test assumes that some SNP from the QTL region were obtained through a SNP discovery process that is not related to the phenotypic data. Moreover, this artificial reduction of SNP density can result in false positive test results, because the TCLD statistic will be artificially increased if the second best SNP is removed and, e.g., replaced by the i-th best.
In 59 replicates, analysis I found two or more SNP with equal log-likelihood values for the most likely SNP. This was typically the causative SNP and a SNP located 1 to 3 positions away from the causative SNP. Evaluating five of these replicates showed equal haplotype combinations for every animal for the two most likely SNP, thus the two SNP were in perfect LD. Other replicates produced similar results, with the causative SNP and one close SNP returning log-likelihood values at a higher level than the rest of the SNP, although not equal. In these replicates, the analysis excluding the causative SNP returned large TCLD values and resulted in the stringent significance threshold that was used here.
As explained by Goddard and Hayes , causative SNP might be expected to show different properties to common SNP, because causative SNP may be subject to selection such that polymorphisms will typically be recent and have low minor allele frequencies. Thus they may show less LD with markers than common SNP. As a consequence, causative SNP may be expected to show less LD to common SNP in real data than in these simulations, which may improve the power of the CLD test in real data, if the causative SNP is included. However, since we tend to choose common markers for SNP genotyping experiments, the causative SNP will less likely be included in real data as long as selection is based on the minor allele frequency. Hence, all SNP in the promising regions will have to be genotyped in order to improve the probability of inclusion of the causative SNP.
When SNP are evaluated, a number of these will be coding SNP that change amino acids . The number of coding SNP is substantially smaller than the overall total number of common SNP. So far little effort has been placed on identifying coding SNP, but for the future, knowledge on which SNP are coding could be valuable when trying to identify causative mutations. Information about coding SNP will reduce the number of candidate SNP and thus improve the power of tests for causal SNP by removing the signal from non-coding SNP in LD with the causative SNP. However, non-coding SNP in regulatory regions of the genes may also be causative. If the candidate region contains several genes, information on gene function could also be used to increase the power of the test.
Including the causative SNP as a marker increased the average log-likelihood values about four times in these simulations (Figure 2). Although these simulations were quite simple, this large increase appears to be quite general, although its size may be modified by different factors, such as family structure, marker density, dataset size and QTL effect sizes. Given our general conclusion that the inclusion of the causative SNP is expected to increase the log-likelihood ratios, these factors are expected to affect mainly the power of the test.
To apply the CLD test to real data, the significance threshold must be estimated from the real data. The basic approach that is proposed is to perform a QTL analysis (i.e. analysis I), and to calculate the TCLD statistic (TCLD(real)). Then, records are simulated assuming that the SNP detected by the QTL analysis is causative, with simulated QTL variances equal to the estimates obtained from the real data analysis, where every SNP i will in turn be assigned as causative, and will be masked when analysing the data. This simulates replicated data under the null-hypothesis with an LD structure as found in the QTL region. Analysing these null-hypothesis data without including the assumed causative SNP will provide a significance threshold for the analysed data. A significance level can be obtained by counting how many of the null-hypothesis TCLD values exceed the real data TCLD(real) -value. For example, if 100 out of 1000 null-hypothesis datasets have TCLD values exceeding TCLD(real), the p-value of the real data CST is 0.1 (= 100/1000).
The relatively low power of the CLD test does not imply that it should not be used, since it is not very costly to perform and, depending on its p-value, it may provide substantial statistical evidence for a causative SNP. However, because of the low power of the test, the p-value of the real data TCLD (as described in the previous paragraph) will in most situations be quite high. Ron and Weller  suggested that the quest for the causative SNP had to be won on points rather than by knockout. Their criteria for validating causality included linkage analysis and LD mapping, positional cloning, selection of candidate genes, DNA sequencing, and statistical analysis. Their conclusion was that only an array of evidence can establish proof of causality. The critical test will be concordance and functional validation. In this setting, the CLD test may provide considerable evidence for a causative SNP, especially when a concordance test cannot be applied, but due to its high p-value, functional evidence will be needed to definitely conclude whether the SNP is causative or not.