Prediction of attrition in large longitudinal studies: tree-based methods versus multinomial logistic models