when we exclude the dosage probabilities that are below 0.8, the inference program performs very well and provides above ∼0.90 IQS on average. The reason is that for dosage probabilities that are lower than 0.8, there is too much uncertainty for the program to impute, which not only heavily distorts the results (poor specificity) but also makes it difficult to filter out true positives (low sensitivity). However, there is always a tradeoff when we enhance the accuracy. If we set a 0.8 threshold on dosage probabilities, the sample size dramatically reduces from 1,586 to 326 (intersection set across all markers). We further apply dTDT and MH test onto these 326 individuals with either the inferred or genotyped data but do not find significant markers at a 10− 5 level due to a small sample size (data not shown).
