Observational vs Experimental Methods in Advertising

Summary

The three experiments in Activity Bias in Advertising provide a powerful case study of observational methods failing spectacularly. Even sophisticated regression and matching techniques cannot recover the true causal effect when the fundamental selection mechanism is violated.

The Failure of Regression Controls

From Experiment 1 (Table 2 in the paper):

ModelControlsEstimated Effect
(0)None1198%
(1)Day dummies894%
(2)+ Session dummies871%
(3)+ Page views, minutes spent872%
TruthRCT5.4%

Adding more controls barely reduces the bias. The fundamental problem: ad exposure is a proxy for being an active internet user on that day, and no set of behavioral controls fully captures this.

Why Matching Fails

  • Propensity score matching tries to find “similar” users who were/weren’t exposed
  • But in display advertising, exposure is determined by visiting specific pages
  • The “matched” control group is fundamentally different from the treatment group in unobserved ways
  • This is precisely the selection problem — exposed users are selected on activity

Implications for Practice

  1. RCTs are essential for measuring advertising effectiveness — observational estimates are not just noisy but systematically biased upward
  2. Activity bias is not unique to advertising — any setting where exposure correlates with baseline activity will have similar issues
  3. The magnitude of bias can be enormous (100x+), not just a modest overestimate
  4. More data doesn’t help if the identification strategy is wrong — this is a bias problem, not a variance problem

Connection to Econometrics

This paper provides a vivid illustration of concepts from Regression and the CEF and Instrumental Variables. The failure here is that there is no valid instrument and the Conditional Independence Assumption is violated. The experimental design (random assignment of ad exposure) is the only reliable solution.

See Also