HOW GOOGLE AND UBER MEASURE WHAT CANNOT BE RANDOMIZED

Summary from the TLDR Marketing newsletter

Major marketing investments that can't be randomized still need causal measurement, rather than post-hoc narratives. Google's CausalImpact and Uber's synthetic control methods show how you can estimate the effect of national launches, sponsorships, and other whole market interventions by building counterfactuals from historical data or comparable markets. These methods are weaker than randomized tests, so you should preregister the analysis, establish a strong pre period, run placebo tests, and report uncertainty intervals. Synthetic control should be a last resort for spending that cannot support randomization.