Statistical inference for generalized additive models: simultaneous confidence corridors and variable selection

Abstract

In spite of widespread use of generalized additive models (GAMs) to remedy the “curse of dimensionality”, there is no well-grounded methodology developed for simultaneous inference and variable selection for GAM in existing literature. However, both are essential in enhancing the capability of statistical models. To this end, we establish simultaneous confidence corridors (SCCs) and a type of Bayesian information criterion (BIC) through the spline-backfitted kernel smoothing techniques proposed in recent articles. To characterize the global features of each non-parametric components, SCCs are constructed for testing their overall trends and entire shapes. By extending the BIC in additive models with identity/trivial link, an asymptotically consistent BIC approach for variable selection is built up in GAM to improve the parsimony of model without loss of prediction accuracy. Simulations and a real example corroborate the above findings.