Sparse estimation in high-dimensional zero-inflated Poisson regression model

Abstract

In the high-dimensional zero-inflated Poisson (ZIP) regression model, the traditional Gauss-Newton iteration method and Expectation Maximization (EM) algorithm cannot do the variable selection. We propose an Expectation-Maximization-Regularization (EMR) algorithm. In the E step, we calculate the expected value of the implicit variable under the existing estimated value. In the MR step, the “regularization + likelihood function” method is used. We introduce the elastic net penalty and use the coordinate descent method to obtain the sparse estimate. The simulation results show that the EMR method makes the parameter estimation and variable selection simultaneously, and the estimation accuracy is high. Finally, applying the EMR method to forecast the number of municipal engineering Public-Private-Partnership (PPP) projects in Hebei Province, this model has a strong explanation.