Methodology

This section holds worked methodological notes. Each note takes a question that clients or analysts ask of an MMM, shows why the default additive PanelMMM does or does not answer it, and sets out the estimation designs that would answer it, ranked by the evidence each can support.

The notes do not document library features. Where a note uses AMMM3 code, the code runs on the released surface described in Model Specification. Where a design needs something AMMM3 does not build, the note says so and does not imply a roadmap.

Read the Agency Workflow first. Its claim classes (data description, conditional model result, predictive evidence, causal evidence, decision recommendation) are used throughout this section to state what each design can and cannot support.

Notes

  • Halo effects between revenue channels How to turn a request to “study halo” into one of four distinct estimands, what AMMM3 estimates by construction, and which designs answer each estimand.

  • Price and promotion in an MMM What adding price or promotion to control_columns commits you to, why both are frequently endogenous, when promotion is a mediator or collider rather than a confounder, and why the budget optimiser cannot set price.

  • Always-on and low-variance channels Why a channel that never moves returns a prior-shaped coefficient that the estimability gate does not catch, which screens to run before fitting, which designs change the answer, and why tightening the prior is refused.