This paper proposes an improvement to Taguchi's T-method. In product development, it is difficult to gather sufficient samples at the beginning of the development process. Taguchi's T-method was proposed as a method for constructing an estimation model in such a situation. However, if an outlier exists in the training data, Taguchi's T-method cannot obtain a correct estimation model. This is because Taguchi's T-method uses the least squares. Therefore, in this paper, we propose an improved version of Taguchi's T-method that does not degrade the accuracy of estimation by using a median-median line even for small training samples with outliers. The Effectiveness of the proposed method is confirmed through experiments. We confirmed that the accuracy of the proposed method is superior to that of similar methods.