Bayesian vs. frequentist MMM: why modern media mix models went Bayesian

Bayesian vs. frequentist MMM: why modern media mix models went Bayesian

Media mix modeling has quietly shifted from classical regression to Bayesian frameworks (Google's Meridian, Meta's Robyn). The choice isn't fashion — it changes what the model can survive.

The core problem MMM faces
MMM fits conversions against spend with few data points (weekly history) and many correlated channels. Classical (frequentist) regression in this regime is fragile: multicollinearity makes coefficients swing wildly, and the model happily returns a negative coefficient for a channel you know works, because the data alone can't pin it down.

What Bayesian adds
Bayesian MMM lets you encode priors — informed beliefs before seeing data. You can constrain spend coefficients to be non-negative (advertising shouldn't reduce sales), set sensible priors on adstock (how ad effect carries over weeks) and saturation (diminishing returns) curves, and the model blends prior with data. With sparse data, priors stabilize estimates; with abundant data, they get overruled.

The honest caveat
Priors are also where bias enters. A team that priors search to be highly effective will get a model that finds search effective. The discipline is making priors explicit and defensible, ideally calibrated from past experiments.

Bottom line for practitioners: Prefer Bayesian MMM when data is thin and channels are collinear — which is nearly always. It prevents the absurd negative-coefficient failures and produces credible intervals, not false-precision point estimates. But audit the priors as carefully as the results: a Bayesian MMM is only as honest as the beliefs you fed it, and calibrating those priors with real lift tests is what separates measurement from confirmation bias.