Inferring Ad Influence Under Segmentation Bias

• Activity Number: 240
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Marketing
• Author(s): Jaimyoung Kwon*+ and Ram Akella and Joel Barajas and Aaron Flores and Victor Andrei and Marius Holtan
• Companies: AOL Advertising and University of California at Santa Cruz and University of California at Santa Cruz and AOL Advertising and AOL Advertising and AOL Advertising
• Keywords: internet marketing ; online advertising ; display advertising ; cost per action ; segmentation bias ; Bayesian statistics

Abstract

In evaluating the effectiveness of a cost-per-action (CPA) online advertising campaign, a key metric is the "ad influence" that measures whether an exposure to the web banner increases a web user's conversion probability. The standard method for inferring ad influence is an A/B testing, in which one divides web users into Study (shown ads) and Control groups (not shown ads) and compare the conversion probability of the two groups. The two-sample binomial model does not apply though, when (a) the ads are not served to everyone in the Study group but only to a target segment most likely to convert and (b) a black-box targeting method is used and one does not know which individuals in the Control group belong to the target. Hence one needs to estimate the "Targeting" effect as well as the "Ad Influence" effect (among targeted segment) using partial data. We formulate the problem in Imbens & Rubin (1977) framework of causal inference for randomized trial under noncompliance, treating the targeting effect as a hidden effect. The method was applied to real-world examples from 3 advertising campaigns to derive the targeting and ad influence effects reliably.