Attribution, incrementality or MMM: which method for which decision?
Attribution, experiment and MMM do not answer the same question. The selector offers a transparent teaching orientation; it replaces neither the framing of the estimand, nor the audit of the data, nor a scientific review.
Choose one option per line. The result is educational orientation, not scientific validation.
Marketing Mix Modeling + Triangulation
Tied at 8 points: weigh the methods against the cost of error, feasibility and the level of evidence required.
| Method | Score | Status | Use |
|---|---|---|---|
| Marketing Mix Modeling | 8 points | tied orientation | to validate |
| Triangulation | 8 points | tied orientation | to validate |
| Experiment | 4 points | alternative | to validate |
| Attribution | 2 points | alternative | to validate |
Educational orientation · the final decision is human
In depth
Understand the mechanism before using the tool.
Measuring marketing means estimating what would have happened without the action. The observed metric is not yet the causal effect.The method depends on the decision, the grain, the horizon and the counterfactual available. Attribution, experiment and aggregate model answer different questions and must be triangulated.
Define population, intervention, comparator, outcome and horizon before choosing the method: this is the estimand the analysis must actually identify.
Align the exposure period, the effect lag and the outcome window; guard against anticipation, contamination and carry-over effects.
Build the counterfactual explicitly: randomisation where possible, otherwise identification assumptions and verifiable diagnostics.
Report the interval, the power, the residual bias, the sensitivity and the economically relevant magnitude, not statistical significance alone.
A measurement becomes decisional when its interval separates the options and the plausible effect exceeds the cost and the risk of acting.
Points of vigilance
The distinctions that prevent a misreading.
It allocates an observed outcome according to a rule; it does not necessarily identify what marketing created.
It compares the exposed outcome with the counterfactual outcome, controlling for systematic differences.
Internal validity concerns the cause within the study; external validity concerns the transfer to other contexts.
Protocol
From framing to validation.
Define population, intervention, comparison, outcome and horizon before choosing the method.
Randomise where possible; otherwise justify the identification assumptions and the diagnostics.
Report the interval, the power, the sensitivity and the economically relevant magnitude.
Compare experiment, aggregate model, operational data and knowledge of the mechanism.
An observed ROAS can be far above the incremental ROAS
Situation. A campaign is credited with 1,000 conversions at €100 of margin, for €40,000 of spend. A test estimates that only 35% are incremental.
Calculation. Observed margin ROAS: 100,000 ÷ 40,000 = 2.5. Incremental ROAS: 35,000 ÷ 40,000 = 0.875.
Decision. Attribution describes where the conversions were recorded; the test changes the economic decision.
Adversarial review
Five questions to settle before presenting a recommendation.
- Is the unit actionable?
Check that it matches the level at which a decision can be executed and tracked. A broader average can reverse the reading by hiding loss-making segments, channels or periods.
- Does the horizon cover the mechanism?
Recompute the conclusion over a short window and a long one. Immediate costs, retention, saturation and delayed effects do not materialise at the same time.
- Does the comparator really represent the absence of action?
Explain what would have happened without the decision. A historical trend or an attribution of contacts is not enough if selection, seasonality or simultaneous changes remain active.
- Which assumption tips the choice?
Identify the parameter that changes the ranking, publish its plausible range and show the prudent scenario. A recommendation without sensitivity hides its main risk.
- Who stops or revises the decision?
Name the owner, the observed signal, the threshold and the review date. Without a stopping rule, a favourable pilot easily becomes a permanent, unexamined commitment.
Acceptance conditions
What must be true before acting.
- Write down the causal quantity sought before the analysis plan.
- Report statistical uncertainty and the threshold of economic relevance.
- Never mix attributed and incremental conversions in the same ratio.
Limits
What the analysis does not prove.
- An underpowered experiment does not prove the absence of an effect.
- Contamination between groups reduces the measurable difference.
- A local, short result does not automatically transfer to every market and horizon.
References
Works used to frame the method.
- Rubin, D. B. (1974), “Estimating Causal Effects of Treatments in Randomized and Nonrandomized Studies”.
- Holland, P. W. (1986), “Statistics and Causal Inference”, Journal of the American Statistical Association.
- Lewis, R. A. & Rao, J. M. (2015), Quarterly Journal of Economics.
To explore next
Continue along the same reasoning.
Which method produces evidence suited to the decision?
Back to the theme →Related topicMarketing KPIs: measurement contract and worked caseDefine one KPI per decision, formula, source, window and guardrail, with an economic cascade and a downloadable template.
Read →Related topicMarketing causality: bias, counterfactual and diagnosisUnderstand selection bias, reverse causality, omitted variables and mediation before attributing a marketing effect.
Read →Related topicMarketing incrementality: methods and choice of counterfactualCritical comparison of incrementality methods, selection criteria, lift calculation and rules of use.
Read →Related topicMarketing geo test: design, power and biasBuild a geo test with matched markets, a minimum detectable effect, spillover control and a dataset.
Read →Next themeMarketing Mix ModelingConnect data, model, validation, allocation and governance.
Explore →