Common MMM misspecifications
The mistakes we see most often, and why a clean dashboard can still hide a broken story.
Guide stub
Open-source tooling made sophisticated MMM easier to run. It did not make good specification automatic. This stub lists themes the full guide will unpack with examples.
Outline
- Omitted confounders (promo, price, distribution, competitor)
- Wrong outcome or mismatched attribution windows
- Adstock / saturation forms that don't match the media
- Collinearity treated as “the model decided”
- Over-trusting a single optimiser seed or Pareto front point
- National models answering geo questions (and vice versa)
- Ignoring endogeneity in spend that reacts to performance
- Presenting contribution shares as causal ROI
Related: validation checklist · how we audit
Worried one of these applies?
An independent audit ranks issues by severity and gives specific fixes.