Can Your MMM Support a Causal iROAS Claim?

Marketing mix modeling becomes dangerous at the exact moment a descriptive output is promoted into a causal instruction: “move the next dollar to this channel.”

An MMM may summarize historical relationships, forecast outcomes and encode saturation or carryover. Those are useful jobs. None automatically proves what would happen under a different spend policy. A causal iROAS claim needs a defined intervention and an argument that the observed data identify its counterfactual.

Start with the claim, not the model

“What was Facebook ROAS?” is too ambiguous. Which spend move? Over what range? In which markets? Against which outcome? At what delay? Under which other budget commitments?

A useful causal target is narrower: the expected change in a defined business outcome from a feasible local shift in one channel’s planned spend, under an explicit operating regime. This is not a semantic preference. It determines what variation, controls and timing the analysis must support.

A 2026 SSRN paper by Dara Dehghan formalizes this discipline for a focal-channel local-shift effect under planned-budget regimes. Its contribution is explicitly an identification and reporting framework rather than a new estimator. It is a recent working paper, so the right use is as a rigorous release standard to evaluate—not as settled universal proof.

The six-gate release standard

1. 01

   Decision target

   Name the channel, feasible spend shift, outcome and time horizon. A generic contribution number is not an intervention.

2. 02

   Decision cadence

   The modeled cadence must represent when spend can change and when outcomes can respond, including carryover and reporting delay.

3. 03

   Budget assignment

   Document why spend moved: approved plans, pacing, contracts, inventory constraints and demand forecasts can all confound the observed relationship.

4. 04

   Independent variation

   The data must contain enough supported variation to distinguish the focal shift from co-moving channels. Extrapolation is not identification.

5. 05

   Locked buys and interference

   Scope commitments, spillovers and downstream demand capture. If the intervention cannot be isolated, narrow the unit, market or claim.

6. 06

   Validation and sensitivity

   Report uncertainty, sensitivity, placebo logic and external calibration. Predictive fit alone does not validate the counterfactual.

Why holdout fit is not enough

Two models can predict the observed range equally well and imply different actions outside it. Dew, Padilla and Shchetkina show that nonlinear response and time-varying effectiveness can be difficult to distinguish in ordinary MMM data, especially when spend is autocorrelated. The competing explanations may fit similarly while recommending different allocations.

This is the gap between prediction and intervention. Cross-validation asks whether a model predicts withheld observations from a similar process. A budget decision asks what will happen after changing that process. The second question needs overlap, assignment logic and counterfactual assumptions that the first test does not establish.

Calibration helps—but does not erase identification

Incrementality experiments can contribute external information. Google researchers describe a Bayesian calibration method that parameterizes ROAS so experimental results can inform priors; their simulations report reduced bias and posterior uncertainty. That is a useful bridge between periodic experiments and an always-on planning model.

But calibration is not a universal causal stamp. The experiment must be relevant to the modeled channel, market, outcome, spend range and period. A lift result from one campaign cannot silently validate every historical coefficient or every future budget level.

Use three release states

This makes model governance operational. The team does not debate whether “MMM is causal” in the abstract. It decides whether a specific claim passes a specific release gate.