MMMs are super helpful in providing an objective and quantitative view of marketing. Instead of guesswork, executives use MMM to holistically measure and optimize the impact of their marketing.
Therefore, it is natural that many companies attempt to develop their in-house MMM. In this article, I will discuss some of the technical challenges of developing an MMM system.
This post uses Challenges And Opportunities In Media Mix Modeling, by David Chan and Michael Perry as a reference.
At the very high level, here are the crucial challenges when developing an MMM system:
1 - Data limitation
- Small data
- Limited data range
- Information content
- Data granularity
- Data quality
- Correlated input variables
2 - Selection Bias
- Ad targeting
- Seasonality
- Funnel effect
3 - Model uncertainty
- Structural uncertainty
- Parameter uncertainty
Data Limitation
The first challenge that we face when building an MMM is the size of the data. The "Small Data" problem is prevalent in MMM, which is surprising given that everyone talks about Big Data. However, when modeling an MMM, we likely need more data quickly.
Three years of weekly data at the national level, equivalent to 156 data points. With only 156 data points, modelers must measure a dozen ad channels and consider seasonality, price, promotion, lagged effect, and more. This will make fitting a model challenging, if possible. Modelers need to make assumptions and use heuristics to go about model fitting.
Another critical data limitation is about the range of the data. The amount of spending YoY is typically confined within a range, but the MMM model is expected to provide insight outside that historical spend range. In other words, they must extrapolate, which comes with uncertainty.
Companies never change their marketing mix rapidly. The elements of a marketing mix remain constant or very similar over a range of five to ten years. This is good because companies can stay consistent and focus on execution. But it's making modeling this data difficult, given that the data for every year is almost identical to other years.
"Another major data limitation is "related input variables."Advertisers often distribute spend across ad channels in a correlated manner, complicating regression model fitting and creating uncertainty about each channel's true impact.
Selection Bias
Let's explain "Selection Bias" with an example. Paid Search targets a segment of the population that is already interested in similar queries. In this situation, demand is correlated with another confounding yet unobserved variable.
Another example of the Selection Bias arises in Seasonality and Funnel Effect.
The funnel Effect occurs when an ad channel impacts another channel. For example, a TV campaign drives up search volume, indirectly impacting Paid Search.
Model Uncertainty
Last but not least, "Model uncertainty" is another critical challenge MMM modelers face.
Given the size of the data and the complexity of MMM models, it's likely to have multiple model fits that explain the data equally well. The issue is when these models have differing views on the optimum mix.
The variance of sales is typically larger than the variance of media spending, which makes measuring advertising effectiveness difficult.
Conclusion
Building an MMM system is a complex and challenging task. The technical challenges of MMM are not limited to the data limitation, selection bias, and model uncertainty.
If you're interested in learning more about Marketing Mix Models and how they can help you business, ELIYA can help. Contact us today to learn more about our MMM solutions and how they can benefit your brand.
