Discovering model structure for partially linear models (Q2304237)

This paper proposes a model-free structure selection method to discover model structure for partial linear models. The proposed method is formulated in a framework of gradient learning, equipped with a flexible reproducing kernel Hilbert space, and an efficient proximal gradient descent algorithm is implemented to tackle the large-scale optimization problem. A combined penalty term couples functional norm and group Lasso penalties on the first and second-order gradient functions, respectively. The second-order gradient functions are examined to distinguish the linear and nonlinear variables. This examination allows to appoint the interaction terms, what helps in the post-structure selection modeling. Advantages of the method compared to a double-penalized partial linear model approach are demonstrated in two simulation examples and two applications to real data. In the first simulated example, the true regression function is additive, whereas in the second it contains a three-way interaction term. The first real dataset includes Japanese industrial chemical firms listed on the Tokyo Stock Exchange studied to check whether concentrated shareholding is associated with lower expenditure on activities with scope for managerial private benefits. The second is a dataset of owner-occupied homes in the Boston Standard Metropolitan Statistical Area studied to explain median values by census tracts. In both cases, interactions are detected by the new method.