B1304
Title: A back-fitting based MCEM algorithm for scalable estimation in multinomial probit model with multilayer network linkages
Authors: Gourab Mukherjee - University of Southern California (United States) [presenting]
Abstract: A new multinomial probit model is proposed for analyzing nominal responses in cross-sectional data. The proposed model uses covariate information and a generalized additive modelling framework to integrate multilayered network linkages between consumers to predict a focal consumer's choice. The key feature of the proposed weighted regression-based additive model is the ability to use {\it multiple} idiosyncratic localized characteristics of the network layers to shrink the model coefficients "locally''. Incorporating this feature can help improve the model's prediction accuracy, especially when the data is cross-sectional and information about an individual consumer is scarce. However, parameter estimation using extant approaches scales poorly. Therefore, a novel Monte-Carlo Expectation-Maximization (MCEM) - based approach is developed that substitutes the computationally expensive E-step in the classical EM algorithm with an efficient Gibbs sampling-based evaluation and implements the M-step using a fast back-fitting method. The proposed MCEM algorithm's convergence properties are established, providing evidence supporting its scalability and providing a distributed computing-based implementation that yields parameter estimates and their standard errors. The proposed method is applied to predict demand for compact cars in the Sacramento market, focusing on the probability of buying a hybrid car.
