Predictive Incrementality by Experimentation (PIE) for Ad Measurement

Brett R. Gordon, Robert Moakler, Florian Zettelmeyer

arXiv 13 Apr 2023 · Econometrics

arXiv:2304.06828 · PDF · Extracted main text

Abstract

We present a novel approach to causal measurement for advertising, namely to use exogenous variation in advertising exposure (RCTs) for a subset of ad campaigns to build a model that can predict the causal effect of ad campaigns that were run without RCTs. This approach -- Predictive Incrementality by Experimentation (PIE) -- frames the task of estimating the causal effect of an ad campaign as a prediction problem, with the unit of observation being an RCT itself. In contrast, traditional causal inference approaches with observational data seek to adjust covariate imbalance at the user level. A key insight is to use post-campaign features, such as last-click conversion counts, that do not require an RCT, as features in our predictive model. We find that our PIE model recovers RCT-derived incremental conversions per dollar (ICPD) much better than the program evaluation approaches analyzed in Gordon et al. (forthcoming). The prediction errors from the best PIE model are 48%, 42%, and 62% of the RCT-based average ICPD for upper-, mid-, and lower-funnel conversion outcomes, respectively. In contrast, across the same data, the average prediction error of stratified propensity score matching exceeds 491%, and that of double/debiased machine learning exceeds 2,904%. Using a decision-making framework inspired by industry, we show that PIE leads to different decisions compared to RCTs for only 6% of upper-funnel, 7% of mid-funnel, and 13% of lower-funnel outcomes. We conclude that PIE could enable advertising platforms to scale causal ad measurement by extrapolating from a limited number of RCTs to a large set of non-experimental ad campaigns.

Citation extraction

appendix boundary found by appendix_command · 68% of the source is main text. Read the extracted text to check this.

Most heavily cited references

The works this paper leans on most, across its whole bibliography — not restricted to papers in our corpus. Ranked by composite intensity, which combines how often a work is mentioned, how many sections mention it, and how much of that falls in the main text rather than the appendix.

Showing the top 10 of 38 scored citations. 1 of these could not be matched to a bibliography entry, so only the citation key is shown.