Counterfactual Inference Using Behavioral Models: An Application to Search Advertising
When designing online advertising markets, it is challenging to use A/B testing to learn about the effects of policy changes. An alternative is to build a statistical model and infer advertiser preferences, and use this model to estimate the effect of counterfactual policies. An important characteristic of a successful model is that it accounts for heterogeneity in advertiser objectives as well as in the magnitudes of the parameters of their preferences. We apply this method to a data set of search advertisers using a large-scale Bayesian model. Results include the finding that advertisers who use "exact match" bidding tend to value impressions rather than clicks. We also evaluate revenue/efficiency tradeoffs involved in improved targeting of advertisements to users.