Robert Moakler, Data Science Intern, Integral Ad Science @ MLconf SEA
summary
Uses viewability (ad >50% in view for ≥1s) as a natural experiment: users with viewable impressions form treatment, non‑viewable impressions form control, outcomes measured in a 7‑day window. Fits a logistic regression on measurable confounders (ad size, page position, screen size) to predict counterfactual conversion. Tests on seven campaigns (millions of impressions, ~2k–2M conversions) show raw impression-based lift collapses dramatically after viewability and adjustment. Validates with negative controls (unrelated purchase events) to confirm bias reduction.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 0:26Is online ads effective
- 1:14Is it really effective
- 1:29How to measure effectiveness
- 4:14How to overcome selection bias
- 7:30Viewability
- 13:37WPrime
- 16:31Conclusion
related talks
30mPredictive Analytics in Social Media and Online Display Advertising
46mFrom correlation to causation: measuring ad effectiveness at scale
36mMeasuring Media Impact- Practical Geo-Lift Incrementality Testing
21mHow to Get Ahead in Advertising
16mThrough the looking glass – Following the rabbit down the data… Whole
13mClaudia Perlich, Chief Scientist, Dstillery
29mAI and ML for Computational Advertising
20mWinning the Game: Leveraging Contextual Bandits for Creative Optimization