MMM: Were the right variables included in the model?
Understand if the foundations are right.
Multicollinearity happens when two or more channels move together so closely that the model can’t tell which one is driving sales.
If TV and OOH advertising always run at the same time, the model is forced to make a judgement call. It may rely on tiny timing differences or simply allocate credit based on statistical convenience.
For OOH, this is critical. As a supporting channel, it often overlaps with lead media. If overlap is high, ROI can swing dramatically based on modelling choices, including sometimes appearing as with zero attribution levels.
There are standard checks you can ask your modeller to run to check for issues with multicollinearity.
- Correlation matrix: A correlation matrix shows how closely channels move together over time, with scores above 0.7 indicating the model may struggle to separate their individual effects.
- Variance inflation factor (VIF): VIF measures how much overlap with other variables makes a channel harder to isolate, with scores above 5 suggesting its results may be unreliable. It’s almost a mini-MMM, assessing the media channel itself rather than the sales.
Have correlation and VIF tests been run to confirm channels can be separated reliably?
YES!
Multicollinearity isn’t an issue.
Good - correlation levels are within safe limits and VIF scores are under control. That means channels have enough separation in timing and weight for the model to measure each independently.
You’re less likely to see dramatic swings in ROI caused by statistical overlap.
NO!
There are signs of multicollinearity.
If channels consistently run together, the model may struggle to untangle their effects. Small modelling decisions can then cause large shifts in attribution.
In these situations, ROI figures can become surprisingly fragile.
To understand how to diagnose and manage overlap: