Seminar: Tim Au, Bayesian Inference for the Causal Impact of Online Advertising Using Linear Gaussian State Space Models
Speaker: Tim Au, Data Science Manager at Google
Title: Bayesian Inference for the Causal Impact of Online Advertising Using Linear Gaussian State Space Models
Inferring the causal impact of online advertising has become progressively more important, with advertisers increasingly relying on comparative case studies to better understand the implications of their marketing strategies. Methods proposed for the analysis of these studies have generally assumed that the intervention of interest only affects a single treated unit (e.g., an entire country). However, marketing interventions frequently affect multiple treated units (e.g., several countries), and analyzing situations such as these has often entailed either applying a method separately to each treated unit or applying a method once to the aggregate of all the treated units. In this talk, we instead consider jointly modeling all of the units together in a multivariate linear Gaussian state space model to more efficiently estimate the counterfactuals, where Bayesian inference facilitates the evaluation of heterogeneous unit-level treatment effects, aggregate-level treatment effects, and other potential quantities of interest. The advantages of our proposed model-based approach are further illustrated using numerical simulations and a real case study from the online advertising context, although we note that our framework is also equally applicable to other areas as well.