Does Bayesian MMM 'fancy up' Marketing Attribution?
Bayesian MMM makes MMM unnecessarily complex
I recently stumbled upon a tweet response to "What controversial statistics take will have you like this (image of a man surrounded by men with spears ready to attack him).
Ed Kroc in his responses said the following about Bayesian methods.
📌 "Bayesians should be forced to scientifically (not mathematically) justify their priors."
📌 "A prior should be scientifically/substantively defensible, otherwise why bother with Bayes?
If you're truly ambivalent about the estimand, then use a flat prior, so let the likelihood drive everything, so again: why bother with Bayes?"
📌 "Bayes is usually used as a way to "fancy up" a problem without sincerely engaging with its complexities.
I prefer the Bayesian approach theoretically, but I don't trust the average non-stat person to do it well."
Bayesians should be forced to scientifically (not mathematically) justify their priors.
I totally agree with all of the above.
Bayesian methods are indeed used to 'fancy up' a problem. It is used to give an impression that something deeply scientific is happening.
Bayesian will showcases chart after chart of sampling from a distribution as if something ground breaking is happening 😅 .
I will give credit to Bayesians for succeeding in the marketing efforts though and for creating an impression that - Bayesian methods are intuitive and easy.
Coming to the field of Marketing Mix Modeling (MMM), Bayesian MMM is just used to "fancy up" the marketing measurement and attribution problem.
You will hear high sounding words like 'Parameter recovery', 'time varying parameters', 'Uncertainty Measurement' etc.
At the end of the day, MMM much like a multi linear regression problem is about estimation of parameters. One need not make estimation of parameters any more difficult, complex and compute intensive than it could already get.
But Bayesian MMM do exactly that and it provides tremendous room for manipulation. MMMs are typically small data problem (relatively speaking). And the priors almost always overwhelm the evidence in the data. Not to mention the other big Achilles heel of Bayesian MMM - Multicollinearity.
If you want a detailed explanation of the shortcomings of Bayesian MMM - the links are in resources.
Lastly I find the meme highly relevant. Frequentist MMM is like Yusuf Dikec (Right). Gets the job done without any fuzz or "fancy-ing up" 😎.
With Bayesian MMM the question begs - why take the complex, convoluted, compute intensive and error prone route.
Link to the tweet response in resources.
P.S: Just to clarify, I am using the meme in jest. No disrespect intended to the South Korean shooter.
Congratulations to both Turkish and South Korean athletes on the Silver Olympic medals.
Resources:
Tweet link: https://x.com/ed_kroc/status/1817018046038278556
Which technique provides for greater manipulation in MMM - Bayesian or Frequentist?
https://open.substack.com/pub/arymalabs/p/which-technique-provides-for-greater?r=2p7455&utm_campaign=post&utm_medium=webTwo key problems that ails Bayesian MMM:
https://open.substack.com/pub/arymalabs/p/two-key-problems-that-ails-bayesian?r=2p7455&utm_campaign=post&utm_medium=webReasons to go Bayesian debunked:
https://open.substack.com/pub/arymalabs/p/reasons-to-go-bayesian-debunked?r=2p7455&utm_campaign=post&utm_medium=webBayesian Marketing Mix Modeling's Stating the obvious problem
https://open.substack.com/pub/arymalabs/p/bayesian-marketing-mix-modelings?r=2p7455&utm_campaign=post&utm_medium=webAdopting MMM for the first time? Use Frequentist MMM.
https://open.substack.com/pub/arymalabs/p/adopting-mmm-for-the-first-time-use?r=2p7455&utm_campaign=post&utm_medium=web
Thanks for reading.
For consulting and help with MMM implementation, Click here
Stay tuned for more articles on MMM.
 | A guest post by
|