Threshold Regression With a Threshold Boundary
Ping Yu, Xiaodong Fan
Abstract
This article studies computation, estimation, inference, and testing for linearity in threshold regression with a threshold boundary. We first put forward a new algorithm to ease the computation of the threshold boundary, and develop the asymptotics for the least squares estimator in both the fixed-threshold-effect framework and the small-threshold-effect framework. We also show that the inverting-likelihood-ratio method is not suitable to construct confidence sets for the threshold parameters, while the nonparametric posterior interval is still applicable. We then propose a new score-type test to test for the existence of threshold effects. Comparing with the usual Wald-type test, it is computationally less intensive, and its critical values are easier to obtain by the simulation method. Simulation studies corroborate the theoretical results. We finally conduct two empirical applications in labor economics to illustrate the nonconstancy of thresholds.