What is a control variable?

Sales move for reasons that have nothing to do with marketing: a price cut, a 20%-off week, Black Friday, a new store, a rate decision for a bank. If the model does not know about them, it will credit the sales spike to whatever marketing happened to run that week. A control variable tells the model about such a factor so the effect is explained by it rather than by a channel.

Every variable added takes explanatory power away from the media channels and has to be forecast too, so the rule at Odins is to use few. The obvious ones with a direct effect go in: price and promotions, distribution where it changed, and special days such as Black Friday. Competitor activity is usually left in the baseline or treated as a trend. Known one-off shocks, a weather event or an outage, are flagged in the history as noise rather than forecast.

The test for a candidate is whether it can be known ahead. A variable that explains the past but cannot be planned, such as a brand's rank on a comparison site next quarter, improves the fit and ruins the forecast. The model looks forward, so its controls have to be knowable forward.

A model with many controls fits history beautifully and explains nothing about the marketing, because the controls soak up the effect. A model with few controls and well-set priors explains the media and lets the baseline trend absorb the rest. Which controls a model uses is visible in Odins, and the list is decided together with the customer during the build.

• Baseline (base sales): The sales that would have happened without paid marketing. Everything the model credits to a channel is measured above it.
• Seasonality: The yearly pattern in demand that the model separates from marketing, so a December spike is not credited to December's ads.
• Prior: What the model believes about a channel before it sees the data. The part of MMM that decides how much the data can teach.
• Channel contribution: What each channel added to sales above the baseline over a period. The decomposition behind every ROAS figure.

Go one level deeper: Price promotions vs advertising in MMM. How the model works. Or go back to all terms in the glossary.