A0469
Title: Fast and efficient space-time covariance estimation in large datasets by composite likelihood truncation
Authors: Zhendong Huang - RMIT University (Australia) [presenting]
Davide Ferrari - Free University of Bozen/Bolzano (Italy)
Alessandro Casa - Free University of Bozen-Bolzano (Italy)
Abstract: Pairwise composite likelihood, a linear combination of pairwise likelihood functions, is a powerful tool for estimating the variogram of space-time random fields. However, the choice of linear coefficients significantly influences the performance of the resulting estimator, both computationally and statistically. The accuracy can deteriorate dramatically when many noisy or highly correlated pairwise scores are included. A new procedure is introduced for selecting and combining pairwise likelihood components from a large set of feasible candidates, while simultaneously estimating the spatial covariance. The proposed method constructs pairwise estimating equations by minimizing an approximate distance from the full likelihood score, subject to a constraint that reflects available computational resources. This results in truncated pairwise estimating equations containing only the most informative partial likelihood score terms. Asymptotic properties of the method are studied, and numerical experiments are performed.
