The impact of undersampling on the predictive performance of logistic regression and machine learning algorithms: A simulation study