Baseline or Base in MMM is NOT “Unexplained Sales”
Recently I came across a interesting post in which the OP said the following “Baseline is revenue + whatever revenue the model was simply incapable of attributing to ads, which can happen for many reasons, such as complex interactions and behaviors not included in the model.”
Other misconceptions is “the baseline is a residual of everything the model’s media variables couldn’t explain”.
Technically the above are not accurate.
The base / intercept is the mean value of Y (your KPI) when you turn all your independent variables to zero. The main point is that the variables are all included and accounted for. So the sentence “behaviors not included in the model” is wrong.
The behaviors not included in the model points to endogeneity via omitted variable bias (OVB) issue.
So the base is not “revenue + whatever revenue the model was simply incapable of attributing to ads”. The incapability to attribute is actually the ‘Error’ of your model, this is manifested in the way of MAPE, NRMSE etc.
Technically it will be wrong to club Base and Error because on one hand you are calculating sales when all the media + marketing + macroeconomic effects are set to zero; and on the other you have real inability of the model to explain the observed sales.
The two are not the same .
The very word ‘baseline’ mean bare minimum. Hence Intercept is the only measure that fits this description.
📌 Base Implies counterfactual and Incrementality
I don’t agree with Bayesian modeling paradigm of Meridian but I would highly recommend anybody to read their documentation. It is very comprehensive, accurate and most of all honest about a lot of shortcomings of both MMM and Bayesian methods.
The meridian document states -“The baseline is the expected outcome in the counterfactual scenario where all treatment variables are set to their baseline values. For paid and organic media, the baseline values are zero […]. Estimating the baseline allows one to understand what would have happened if they did not engage in paid media, organic media, or other non-media treatments.”
This take is accurate and practical. In fact we had penned an article “MMM is the biggest incrementality test one can do” way before. (Link in comments).
📌 Practical Suggestions
“Expected outcome when treatment variables are set to baseline values”
Actually that’s fine as an interpretation layer and I concur with this.
Statistically, this still maps back to the intercept concept, possibly with practical constraints (e.g. some variables can’t literally be zero). Again I have written about this in my article (link in comments).
Setting variables to minimum or realistic values = practical modeling choice
It does NOT redefine what an intercept is.
“Baseline” or “Base” in MMM is just Intercept and has a strict statistical meaning.
If we start diluting that meaning, then anything can be called anything. Which is how bad measurement decisions creep in.
