The purpose of this study was not only to investigate critical factors for taking English mediated instruction class (hereinafter referred to as EMI class) but also to analyze the effects of EMI on university students' academic self-efficacy and self-concept for English communication by using KELS data of 2011-2012. First of all, selection mechanism for EMI class was found using logistic regression. Following this process, propensity score matching strategy, combined with two types of regression, i.e., OLS and logistic regression, was used for analyzing the treatment effects of EMI Class. In addition, two types of sensitivity analysis, i.e., Rosenbaum Bounds and Mantel-Haenszel Bounds, were used to estimate the robustness of the treatment effects against the hidden bias. The results of this study were as follows: First, students' English standardized test score of the university entrance exam, the total credits of taking EMI class in the previous year, and academic self-concept of last year were significant factors for choosing EMI class. Second, the effects of EMI class on students' academic selfefficacy and self-concept for English communication were not statistically significant after controlling for "selection bias". Third, sensitivity analysis showed that the average treatment effect of EMI class was robust against the effect of unobserved variables. Based on these statistical results, some policy implications were suggested in this paper.