A graphic and tabular variable deduction method in logistic regression. Issue 16 (18th August 2022)

• Record Type: Journal Article
• Authors: Zeng, Guoping
• Is Part Of: Communications in statistics. Volume 51:Issue 16(2022)
• Page Start: 5412
• Page End: 5427
• Publication Date: 2022-08-18
• Subjects: Logistic regression -- logit plot -- missing values -- outliers -- summary table -- variable deduction; Primary 62-07 -- Secondary 62G08; Mathematical statistics -- Periodicals; Mathematics; Statistics; 519.2
• DOI: 10.1080/03610926.2020.1839499
• Languages: English
• ISSNs: 0361-0926

Abstract

Abstract: Logistic regression assumes linearity between independent variables and log odds. Currently, there are two methods to check the linearity as part of variable deduction. The first method is to use the well-known Box–Tidwell test, which runs logistic regression with interaction terms for all independent variables. The second method estimates log odds by running logistic regression for all independent variables. However, these two methods have some common issues. Firstly, they don't treat missing values, leading loss of information. Secondly, they don't treat outliers. Thirdly, both run logistic regression for all independent variables, which is infeasible for big data. In this article, we propose a graphic and tabular variable deduction method by simultaneously checking the linearity assumption of each independent variable using a logit plot and treating missing values and outliers using a summary table. Our novel method is easy to implement for any real-life predictive modeling using logistic regression.