In a citizen opinion poll, the questionnaire list items are generally taken as binary list items. If the response is binary data, logistic regression is widely used for exploration of influential factors. Standard multiple logistic models are usually based on a linear combination of these exploratory variables. However, in a questionnaire study, the true model is rarely linear and the estimated models have poor predictive accuracy. To solve this problem, boosting methods can be used. In this paper, we propose a new boosting method for binary response cases, the logistic additive regression trees via penalized likelihood approach (EN-Bagging-MART), where an elastic net penalty is applied for the construction of regrarized logistic regression model. The performance of this method was evaluated by simulation. We have also illustrated the usefulness of the EN-Bagging-MART method in a study of satisfaction with city water.
