Unpopular Opinion on Data Imputation in MMM

Domain Expert Imputation > Statistical Imputation

No dataset is perfect, including the ones we use in Marketing Mix Modeling (MMM).
Sometimes, despite our best efforts of accurate data collection, we have to reluctantly resort to imputation techniques for missing data.

Now, if you know me, you know I am a strong advocate for statistical methods. In most cases, they outperform gut feel or guesswork by a wide margin.

But when it comes to data imputation in MMM, I have often found domain-informed imputations to be more accurate than purely statistical ones !!

They say that MMM is art + science. I guess data imputation is one area where this becomes clearly visible.

Why Statistical Methods underperform relative to Domain Experts ?

Statistical methods often falls short in data imputation because statistical models lack context.
They don't know the business cycles, the market shifts, or the nuances of promotional calendars.

But a domain expert does.

Let’s say you are missing December sales data. The rest of the year ranges between $2M–$3M. Any statistical method will likely impute something within that band.

But a domain expert might know that December historically sees a huge sales spike often exceeding $4M. His/her imputation, grounded in context, hence ends up being far more accurate.

So, does that make statistical imputation methods useless?

Not at all.

In fact, techniques like MICE (Multiple Imputation by Chained Equations) benefit significantly when seeded with good, domain-informed initial guesses.
You get faster, more accurate convergence compared to starting with blind imputations.

I will dive deeper into MICE in an upcoming post.

But the takeaway is this:
▪️In MMM, don't disregard domain expertise.
▪️Domain Expertise + Statistics is a powerful combination.

---

Thanks for reading.

Need help imputing your MMM data and building a good MMM model ? Click here