A1484
Title: Approximate unbiased sparse estimating functions in high-dimensions
Authors: Giulia Bertagnolli - Free University of Bozen-Bolzano (Italy) [presenting]
Alessandro Casa - Free University of Bozen-Bolzano (Italy)
Davide Ferrari - University of Bolzano (Italy)
Abstract: In many statistical settings, specifying the full likelihood is challenging if not impossible, which has motivated the development of composite likelihood methods. In high-dimensional contexts, there is often an additional need for sparse estimation to reduce model complexity and identify a limited set of significant parameters. Both challenges are addressed within a unified and general framework. Starting from an unbiased estimating function, we solve a penalized minimization problem: The resulting optimal estimating function is sparse, approximately unbiased, and achieves minimal dispersion distance from the classical score function. By imposing sparsity directly on the set of estimating functions, we significantly reduce the number of estimating equations to be solved. We also propose an efficient algorithm for solving the minimization problem, with two main computational advantages: it operates in a block-wise fashion and avoids inverting large covariance matrices.
