In microarray experiments, numerous genes are tested at the same time, and some of them are with low variability. When detecting differentially expressed genes under two biological states, traditional t-test statistics would be very large because of the small denominator. Thus, many of these genes will be mistakenly declared significance via traditional t-test. The adjusted t statistic is to deal with this problem by adding a reasonable constant to the denominator of traditional t-test statistic. This paper suggests a method to select the added constant with data under a mixture model. Expectation-Maximization algorithm is used to estimate MLE of parameters of incomplete data and cope with the mixture components. The criterion of selecting appropriate adjustment is based on the goodness-of -fit test statistic for middle portion data and the postulated model. The simulation results confirm that adjusted t statistics perform better for significance analysis of microarrays.