Inputs and Variables

An MMM is only as good as its inputs. They fall into three buckets: media variables (the levers you control), control variables (drivers you don't), and the dependent variable (what you predict).

Getting these right, at the correct granularity and cadence, is most of the work in a credible model.

Choose a clean outcome: units sold, revenue, or new customers. Be consistent about whether it's gross or net, and whether returns are excluded.

Model at the granularity you can act on, usually national or regional weekly. Daily data adds noise; monthly hides dynamics.

You can feed media as spend or as exposure (impressions, GRPs, clicks). Impressions decouple delivery from cost, which matters when CPMs swing.

A common pattern: model on impressions for the response curve, then convert back to spend for ROI. Always keep units consistent within a channel.

Advertising effects persist after the flight ends. Adstock models this decay so this week's sales reflect prior weeks' media.

The decay rate lambda controls memory: 0.0 means no carryover, 0.8 means a long tail. TV usually has higher adstock than search.

Doubling spend rarely doubles sales. Saturation transforms capture diminishing returns so the model doesn't extrapolate linearly forever.

Hill or log transforms bend the response curve, giving each channel a realistic ceiling and an efficient operating range.

Controls absorb non-media drivers so they aren't wrongly credited to ads. Key ones: price, promotions, distribution/availability, competitor activity, and macroeconomic indices.

Omitting a strong control biases media coefficients, often inflating whichever channel happened to run during a promo.

Many products have strong seasonal patterns. Capture them with month/week dummies, Fourier terms, or a seasonal index, plus a trend term for underlying growth or decline.

Without these, seasonal sales lifts leak into media variables that happened to run in peak season.

MMM separates a base (what you'd sell with zero new media: brand equity, distribution, price) from incremental contributions driven by media.

A healthy model often shows a large stable base. If media "explains" almost everything, the controls are probably too weak.

When channels flight together, the model can't tell their effects apart and coefficients become unstable, sometimes flipping sign.

Diagnose with correlation matrices and VIF. Mitigate by varying spend over time, adding geo variation, or using Bayesian priors to stabilize estimates.

Align all series to the same calendar and time zone. Fill gaps deliberately, not with silent zeros that mean "no spend" versus "no data".

Deflate monetary series for inflation if the window is long, and document every transformation so results are reproducible.

The final design matrix layers raw controls plus transformed media. Each media column has been adstocked then saturated before entering the regression.

This pipeline order matters: carryover first (when the effect lands), saturation second (how much it lands).

Identify the right transform for the right phenomenon.

MMM inputs are the dependent variable, media variables (transformed by adstock then saturation), and control variables (price, promo, distribution, seasonality, macro).

Strong controls protect against over-crediting media, and clean, consistently-scaled, well-documented series are the foundation of a trustworthy model.