Outlier detection is an important part for data analysis, since outliers can infect the statistical inference evidently. The design matrix in a linear model for cross sectional data was rewritten, and an outlier detection method was proposed based on the mean shift model by using the coefficient shrink estimation. Because the selection of tuning model parameters is very important for outlier detection, a new tuning method based on a weighted tuning process was proposed. The numerical simulation results show that when the new tuning method is applied in the outlier detection procedure, two false identification probability can be decreased observably.