What is Multicollinearity, and Why Should You Care?
I’ll admit, when multicollinearity won our poll for the next hot topic to be discussed on ‘MMMorning! – with Will Marks’, I was surprised. But I’m glad it did, because this isn’t just a technical quirk for analysts — it’s a fundamental challenge for anyone who wants to measure media impact accurately. And as I said on the show, “Multicollinearity is hiding in plain sight, and it is very easy to identify through some simple analysis.”
To help explain, I brought in Nate Tomasetti, Staff Data Scientist at Mutinex. Nate’s job is to keep improving our Market Mix Models (MMM) so they deliver the best possible results for our clients. And as he put it, “Multicollinearity means some of the data looks the same — we’re not saying it has the same values all the time, but if they’re going up at the same time and down at the same time, it can cause lots of problems in the model.”
Think of it like five people pushing a car at once. The car moves, but you can’t tell who’s doing the heavy lifting. If your media channels are always “pushing” together — spending and pausing in sync — your model can’t separate out what’s really driving results.
Bayesian Priors Can Only Help so Much
Modern MMMs use Bayesian approaches, which help by setting “priors” — essentially, guardrails for what’s plausible. This stops the model from making wild claims, like “TV did 0% of the work and digital did 100%.”
But as Nate points out, priors can only do so much: “If the data is really multicollinear, then a model can’t really say which channel is doing more than the other based on the data because they’re all working at the exact same time.”
And that’s a problem. If the data can’t tell the model what’s happening, the priors take over — and if your priors are too strong or too biased (intentionally or not), your results will be skewed.
This is where generalised models show their strength. Rather than relying on hand-tuned assumptions, they draw on patterns from across businesses and industries to form realistic expectations, whilst also keeping these priors weak enough to let the data lead, not dictate.
Injecting Variation
Here’s the bottom line: models are math, not magic. They can’t break apart data that has no variation. And marketers, we have more power than we think to fix this. Nate summed it up perfectly when he said, “The only real way to break apart multicollinearity is to get rid of multicollinearity — and we can do this by adding more variation, not just in the amounts that we’re spending on each channel but the timing.”
So how do you solve Multicollinearity?
- Seek it:
Work with your analytics partners to identify where your channels are moving in lockstep. This is often hiding in plain sight, and a simple correlation analysis can reveal it. - Break it:
- Stagger Channel Activations: Does every channel need to launch in week one? Try offsetting TV and digital by a week or two.
- Vary Budgets: Don’t spend exactly the same amount every campaign. Ramp up, then down. This creates the variation models need to learn.
- Test it:
Challenge your model to predict the outcome from a period it has not seen. The strongest models earn trust not just by explaining the past, but by proving they can predict what is to come.
The Final Word: Plan with Measurement in Mind
In summary, if you want accurate measurement, you need to plan for it. Don’t just assume your current approach is optimal — if you’re not measuring accurately, you can’t know. Treat your media plan as a laboratory: inject change, break up those patterns, and let your models do their best work.
Watch the MMMorning! episode here.
[Will Marks is Head of Marketing Science at Mutinex]