For decades, the Frequentist approach to Marketing Mix Modeling (MMM) was the industry standard. But as marketing journey became more complex and fragmented, it started to show its cracks. Here’s why a Bayesian approach to Marketing Mix Modeling has emerged as the stronger approach.
At a high level, Frequentist statistics treats probability as purely data-driven. No assumptions, no prior knowledge, just the data. It outputs point estimates, quantified using confidence intervals. Bayesian statistics, on the other hand, treats probability as a belief that gets updated as new data arrives.
At Impression, we consider many factors before choosing the best approach for all client needs, but Bayesian is almost always our approach of choice. Let me explain why.
Frequentist methodology: Rigid, straightforward, confined
The core limitation of the traditional Frequentist approach to MMM is its rigidity. Coefficients are treated as static, unknown values. The model’s understanding is strictly confined to the data it is specifically provided with.
For a marketer, this creates some familiar headaches. If you scale media spend across multiple channels at the same time, the model gets confused. We call this multicollinearity. The model ends up assigning contribution to the wrong channel because it can’t determine which channel drove the increase in revenue.
And if you’ve just run an incrementality test proving your channel delivers a ROAS of 3? The model doesn’t care. It starts fresh every time, with no memory of what you already know.
Despite its issues, the Frequentist approach still has its uses. If you are looking for a straightforward analysis and have a large, clean dataset, it could be all you need. It requires much less setup and tuning to get to your results.
Bayesian methodology: Flexibility, probability, continuous
The Bayesian approach treats model parameters as probability distributions rather than fixed values. Rather than a single answer, the model holds a range of possibilities, with some possibilities more probable than others. These distributions are called priors. We aim for priors to be weakly informative, meaning they nudge the model towards a reasonable starting point, while staying flexible enough to move where the data tells it to.
When the model runs, it constantly updates the prior using the data it is given. The result is an updated probability distribution, called the posterior. The better defined the prior, the more accurate the estimate of the parameter. The data still influences the outcome either way, but a good prior gives it less room to produce an implausible result.
A well-defined prior is one thing, but incorporating external evidence is where Bayesian modelling really comes into its own.
Calibration considerations for Bayesian
When you conduct an incrementality test, you are learning something concrete about a specific channel. So, why wouldn’t your model learn from it too?
That’s exactly what calibration allows. Instead of your Marketing Mix Model and your experiments sitting in separate reports that occasionally contradict each other, the test results feed directly into the prior, and the model updates accordingly. It fills in the gaps for all the weeks you weren’t running a test.
The result is a model that gets smarter over time. Every test you run, every result you feed in, makes the next iteration more informed. That’s something a Frequentist model simply cannot do.
Limiting uncertainty with Bayesian
Frequentist models give you a single number. Bayesian models don’t just give you an estimate; they tell you how much to trust that estimate.
Take two channels sitting on your media plan:
| Average ROAS | Credible Interval | |
|---|---|---|
| Channel A | 3.0 | 1.0 – 5.0 |
| Channel B | 2.8 | 2.7 – 2.9 |
Channel A looks better on paper. But Channel B is the safer investment. Its credible interval is tighter. You know roughly what you’re going to get. Channel A’s range is far wider, meaning the model is far less certain. It might be a ROAS of 5. It also might be burning money. Which channel would you choose?
As the world of online privacy grows, attribution measurement becomes increasingly unreliable. This is why Marketing Mix Modeling is emerging as the primary methodology to measure your marketing efforts. At Impression, we specialise in building robust, bespoke MMMs, tailored to each client. If you’d like to explore what that could look like for your business, get in touch.
