Inference by Stochastic Optimization: A Free-Lunch Bootstrap

Speaker Name: 
Jean-Jacques Forneron
Speaker Title: 
Assistant Professor
Speaker Organization: 
Boston University
Start Time: 
Thursday, December 5, 2019 - 1:40pm
End Time: 
Thursday, December 5, 2019 - 3:00pm
Location: 
E2-499
Organizer: 
Assistant Professor, Jessie Li, Econ

Abstract:

This paper proposes a Stochastic Newton-Raphson (SNR) algorithm which delivers asymptotically valid Bootstrap draws and point estimates in a single run. This algorithm generates draws that take the form of a Markov-Chain generated by the gradient and hessian computed on batches of data that are re sampled at each iteration. We show that these draws yield both accurate estimates and asymptotically valid frequentist inferences. This is particularly attractive in settings where the model needs to be re-estimated many times to compute standard errors. SNR performs well in simulations. Furthermore, a simple modification of the baseline algorithm produces graphically appealing synopses of data irregularities. Sensitivity of the estimates to outliers are illustrated in several applications.

Event Type: 
Event