Many had written off MMM (Marketing Mix Models) not so long ago. But like a phoenix, MMM keeps rising from the ashes.🔥
Thanks to:
GDPR, Apple 14 privacy, Google’s cookie deprecation and ever growing pressure on marketing departments to justify their budget allocation, MMM is back to life again. 🚀
But how can we prevent MMM’s death (again)?
In my opinion, the following caveats are important to keep in mind.
⚙ Don’t treat MMM like AutoML
I have heard prospective clients wanting to fully automate the MMM pipeline. Automation is possible in parts but not fully. The variables that are selected automatically based on superficial correlation (or similar metric) are rarely relevant. Models specified automatically are always misspecified. One may have a high accuracy model but it would not make domain sense.
🔃 Not respecting the causal structure
MMM is all about attribution. To know what moved the needle of your KPI, one must causally prove the relationship. Only then will your attribution be accurate. Mere correlation will not suffice. It is imperative that we employ all the econometric tools possible to get right the causal structure.
📈 Pay attention to endogeneity
Following from the last point, endogeneity is another problem which one should entangle if they want a good attribution. You don’t want endogeneity to muddy the water and thereby spoil the reflections (insights).
🛠 DIY MMMs
There are many good open source libraries out there. However, unless you are a skilled MMM modeler with good grasp of statistics/econometrics, building MMM models on your own using these tools is a recipe for disaster. Many organizations approach us after they have tried their hand at these libraries but did not get satisfactory results.
When we take over the project, we find that it is not the fault of the tool but the person !! Not to mention the time wasted and the opportunity cost of getting delayed insights.
