Examining the Effects of Class Imbalance with Missing Data on the Logistic Regression Model

The problem of class imbalance has become frequent, and missing data also are often a predicament in the social sciences. In the literature, the two problems have been debated substantially and separately. Class imbalance has been shown to adversely impact the estimation of logistic regression coefficients. Also, the missing data problem has been demonstrated to yield bias in parameter estimation in various types of statistical analyses. This simulation study focuses on the context logistic regression and investigates how class imbalance interact with missing data to impact the estimation of the coefficients and test procedures such as the Wald and likelihood ratio tests. Three factors are considered in this simulation—(1) missing data mechanisms MCAR, MAR, MNAR, (2) sample sizes—small, 100; medium, 500; and large, 1000, (3) class imbalance ratios 10%, 20%, 30%, 40%, and 50%. The results demonstrate that the combination of class imbalance and missing data problems affects the performance of the logistic regression, particularly when the imbalance ratio is as low as 10% and the sample size is also low, and potential methods for the remedies are suggested.