A good prediction can hide an unstable split
If two channels repeatedly rise and fall together, a model can use their combined signal to predict sales while struggling to separate their contributions. Regularisation may stabilise the coefficients, but it does not create independent evidence about which channel caused the outcome.
- Inspect joint campaign timing
- Test split stability
- Choose the decision level
- Design separating variation
The multicollinearity guide covers the statistical problem. The practical question here is whether the business really needs separate channel decisions, and what evidence would support them.
The response you infer depends on transformation choices and what the data can separate. The how marketing mix models work learning path connects this idea to the assumptions, practical checks and next steps.
To work through the connected decisions, read Robyn model selection: review the Pareto candidates and Paid search MMM: separate demand capture from demand creation.
Work through a joint campaign
Imagine television and online video are always launched in the same weeks at a fixed two-to-one spend ratio. A model can assign different coefficients to them while preserving almost the same combined predicted contribution. A precise-looking ROI table may conceal that substitution between coefficients.
Inspect the raw campaign calendar and transformed inputs. Adstock and saturation can change the degree of correlation, so a single correlation matrix on raw spend is not the whole diagnosis. Compare coefficient and allocation stability across plausible specifications.
Choose among grouping, uncertainty and testing
Grouping the channels can be sensible if they are bought and managed as one package and the decision is whether to increase that package. It changes the estimand: the result concerns the combined intervention, not each component. Document the grouping rule and preserve the underlying series.
If the business must choose between the channels, grouping alone does not answer the question. Report the uncertainty and consider a design that varies one component while holding the other sufficiently stable. The testing guide helps turn that information gap into a testable decision.
Avoid a mechanically convenient answer
Do not drop a channel merely because its coefficient has an inconvenient sign or because a variance statistic crosses a generic threshold. Investigate control selection, outcome definitions and the plausible causal structure. A forced positive coefficient can still allocate credit incorrectly.
Compare a grouped model, plausible separate-channel models and any relevant experimental evidence on the same commercial decision. The ridge regression guide explains the role of shrinkage without treating it as a causal cure.
The review should state what level of effect is supported, which budget movements remain ambiguous and what new variation would help. A credible combined estimate with a clear test plan can be more useful than a fragile table of separate returns.
Sources and further practice
This article’s worked example is illustrative, not a measured client result. Use the linked methodology to check the assumptions before applying it.