Bayesian Method for Causal Inference in High-Dimensional Time Series with Applications to Sales Data

• Activity Number: 442
• Type: Contributed
• Date/Time: Tuesday, August 2, 2016 : 2:00 PM to 3:20 PM
• Sponsor: Section on Statistics in Marketing
• Author(s): BO NING*
• Keywords: Causal Inference ; State Space Model ; G-Wishart Prior ; Bayesian Variable Selection ; Advertisting

Abstract

We propose a novel Bayesian method for detecting causal impact in high-dimensional time-series . A Bayesian structural time-series model with spatially correlated variables are considered. A variable selection mechanism is considered within the proposed method to introduce sparsity in the model and G-Wishart prior is used on the precision matrix to give a graphical structure in the model. We adopt the stochastic search variable selection method for posterior computation and measure the causal effect by comparing the posterior distribution of the trend given the entire data and that given a part of data without observations possible affected by the causal impact. The method is shown to give useful results in simulation studies. Further the method is applied on a data set on sales to determine the effect of an advertising campaign.