Graph estimation for Gaussian data zero-inflated by double truncation

Abstract: We consider the problem of graph estimation in a zero-inflated Gaussian model. In this model, zero-inflation is obtained by double truncation (right and left) of a Gaussian vector. The goal is to recover the latent graph structure of the Gaussian vector with observations of the zero-inflated truncated vector. We propose a two step estimation procedure. The first step consists in estimating each term of the covariance matrix by maximising the corresponding bivariate marginal log-likelihood of the truncated vector. The second one uses the graphical lasso procedure to estimate the precision matrix sparsity, which encodes the graph structure. We then state some theoretical convergence results about the convergence rate of the covariance matrix and precision matrix estimators. These results allow us to establish consistency of our procedure with respect to graph structure recovery. We also present some simulation studies to corroborate the efficiency of our procedure.