Bayesian vs frequentist MMM

The practical difference, minus philosophy

Skip the seminar-room debate; for a buyer it reduces to three operational differences. First, uncertainty: Bayesian outputs carry credible intervals natively, so "is search really better than social" has an answerable form (the intervals overlap or they do not). Second, prior knowledge: priors let experiment results and physical plausibility enter the model formally, which is how calibration works in Meridian and PyMC-Marketing.¹² Third, cost: Bayesian sampling is computationally heavier and demands more statistical skill from whoever runs and explains it.

Why Robyn stayed ridge

Robyn's bet is different: ridge regression, fast to fit, wrapped in an evolutionary search over transform hyperparameters, producing many candidate models for a human to choose among.³ It trades posterior distributions for speed and accessibility; uncertainty is explored through the spread of candidate models rather than expressed within one. Neither approach is free of judgment; they relocate it. Bayesian judgment lives in the priors, Robyn's in the model-selection step. The question for any team is which kind of judgment it can exercise and defend.

What it changes when evaluating vendors

If a vendor is Bayesian, the diligence questions are prior questions: which priors are informative, what informed them, show prior-versus-posterior for the big channels. If a vendor is frequentist or ridge-based, the questions move to stability and selection: how do estimates move between refreshes, and who picks the final model by what criterion. Both sets appear in the 12 vendor questions. A vendor of either school that cannot answer its own school's questions has chosen its method for marketing reasons, which is at least on-brand.