Abstract

Model selection and inference are necessary steps in manypractical analyses related to economics, engineering and biomedical sciences.The growth in size and complexity of modern data, however, challenges theapplicability of traditional approaches for inference and model selection. Inthis talk, we introduce a method and related algorithms to automatically selectestimating equations from a large set of feasible candidates and simultaneouslycarry out parameter estimation. The proposed approach minimizes the distancebetween the linear combination of candidate estimating equations and the fulllikelihood score equation subject to a weighted L1-norm penalty. The selectedestimating equations are sparse in the sense that they contain only the mostinformative model parameters, while overly noisy or redundant ones are dropped.We show that our method satis¿es the so-called oracle properties byguaranteeing consistency for model selection while retaining optimalstatistical accuracy for the selected parameters in large samples. Theproperties of our methodology are illustrated through numerical examples onsimulated and real data.