Estimation bayésienne du lasso adaptatif pour l'issue

Authors: Gaye, Serigne Abib
Advisor: Talbot, Denis
Abstract: In this paper, we aim to develop a new estimation method for the outcome adaptive lasso using Bayesian machinery. The research hypothesis is that our new method will significantly reduce the computational burden of the outcome adaptive lasso. Our method uses the same theoretical foundation as the outcome adaptive lasso. It therefore meets the oracle properties. For its implementation, Bayesian propensity score model is first fitted. Next, the average treatment effect is estimated using inverse probability of treatment weights. In addition, we consider a gamma distribution for the regularisation parameter λ in order to choose this parameter over a continuous set, whereas the frequentist outcome adaptive lasso uses a cross-validation procedure that selects λ among a prespecified discrete set. In fine, the method we have developed meets our expectations, and therefore makes it possible to produce inferences much faster. Indeed, it took only 41.298 seconds for this method to yield inferences, while 44.105 minutes were required for the outcome adaptive lasso. We hope that the ideas developed in this paper will significantly contribute to improve methods for selecting variables in causal inference with the support of Bayesian techniques.
Document Type: Mémoire de maîtrise
Issue Date: 2020
Open Access Date: 12 March 2020
Grantor: Université Laval
Collection:Thèses et mémoires

Files in this item:
Description SizeFormat 
36012.pdf682.58 kBAdobe PDFThumbnail
All documents in CorpusUL are protected by Copyright Act of Canada.