A0814
Title: Sure screening for interaction effect in generalized linear models
Authors: Yuta Umezu - Nagasaki University (Japan) [presenting]
Abstract: In regression models, detecting interactions between several covariates is an important task in many areas. However, when interaction effects are modelled, the model's size becomes vastly large even if the number of covariates is relatively small. Consequently, handling high dimensionality is needed, which means the number of parameters may be much larger than the sample size. Detecting such interactions in generalized linear models based on marginal screening are investigated. To this end, a marginal maximum likelihood estimator is considered in which the Lasso penalty is adopted only for the coefficient of the interaction term. As a result, a simple but effective screening rule can be obtained. Moreover, the screening rule enjoys good theoretical property, the so-called sure screening property; truly active interactions with probability converging to one under appropriate conditions can be detected. Several simulation studies and a real data example for checking the method's performance will also be presented.
