Estimating psychological status from linguistic information is essential for understanding human actions across various domains. Conventional machine learning methods often face overfitting issues due to the tendency of linguistic features outnumbering data entries, which need to be collected by questionnaire survey. Our study addresses this issue by employing a multi-target feature selection (MTFS) method, which selects linguistic features relevant to multiple psychological variables to enhance the generalization ability of the prediction model. We tested MTFS against single-target feature selection (STFS) methods using several machine learning algorithms on two datasets. The results show that the MTFS method improves prediction performance, which suggests that implications from the field of social psychology can enhance the selection of relevant linguistic features.