Using the filtered dataset, the examined imputation algorithm, as previously described for allele and genotype imputation, also presents a high overall accuracy when comparing the association statistics generated using genotyped and imputed markers (Spearman's rank correlation coefficient, rS = 0.80). However, it is important to note some points. Firstly, the wide dispersion of dots around the upper left side of Figure 1 (panels A and B) indicates that, despite the overall good agreement, results that rely solely on imputed genotypes might reject more often the null hypothesis when it is in fact true. For example, at an α = 10-5 we observed that inferences based on truly genotyped markers yielded evidence for 38 markers possibly associated with type-II diabetes, whereas this number would be 73 SNPs had the same inference been based on imputed markers solely. Approximately 47% of markers that would be considered associated by imputed allelic frequencies were discordant to the evidence derived from direct genotyping. Table 1 shows that similar results are also seen for different significance thresholds, suggesting that imputed frequencies might be more prone to false-positive associations independent of a specific pre-defined significance threshold.
