A1241
Title: Iterative distributed multinomial logistic regression
Authors: Xuetao Shi - the University of Sydney (Australia) [presenting]
Yanqin Fan - University of Washington (United States)
Yigit Okar - University of Washington (United States)
Abstract: An iterative estimator for the multinomial logistic regression model is introduced that is both asymptotically efficient and fast to compute even when the number of choices is large. In many economic applications, such as text analysis and spatial choice models, the number of discrete choices can be large. Solving for the maximum likelihood estimator via traditional optimization algorithms, such as Newton-Raphson, is infeasible because the number of arguments in the log-likelihood function is enormous. This problem is tackled by proposing an iterative estimator that optimizes the two parts of the log-likelihood function in turn. The proposed estimator allows for distributed computing, which substantially reduces the computational time. It is shown that the estimator is consistent and has the same asymptotic distribution as the maximum likelihood estimator. Via an extensive simulation study, it is shown that the iterative estimator has good finite sample performance and is extremely fast to compute.
