Marketing causality: identifying the path that manufactures false performance.
Directly verifiable evidence
Primary sources to check the key claims.
- Holland (1986), Statistics and Causal Inference (opens in a new tab)
Definition of the causal problem and the counterfactual.
Evidence status
What was not reproduced.
The source's training was re-run and reproduces its predictive performance (holdout R² of 0.843). The graph of relationships between channels does not come out of it: it comes back undifferentiated, whereas the paper presents one whose links would suggest plausible causal pathways. This dossier therefore validates no causal conclusion, and that measurement supports with a figure what the page states: predicting well is not having established a structure.
Key distinctions
Three conditions before calculating.
An explicit estimand
Average effect, effect on the treated, effect of a budget tranche or long-run effect are not the same quantity. The design cannot be chosen before the population, the action and the horizon have been defined.
Open paths
A variable common to the action and to the outcome creates a non-causal path. Controlling for it can help, but controlling for a mediator or a collider can also introduce a new bias.
Design before model
Randomisation, an eligibility rule, or credible geographic or temporal variation create the comparison. A more complex model does not automatically repair a missing counterfactual.
Method
Building a causal claim proportionate to the evidence
- 01Define the intervention
State who receives what, when, at what dose and relative to which absence of action; a composite campaign must be decomposed if its elements can vary.
- 02Draw the causes
Map prior variables, targeting criteria, latent demand, competition, mediators and outcome; the graph serves to decide what to measure and what not to control.
- 03Choose the variation
Prefer randomisation; failing that, document the assumption that makes a variation credibly quasi-experimental and look for placebo tests or pre-trends.
- 04Bound the use
Report the interval, the sensitivity to confounders and the population covered; local evidence must not be turned into a universal return.
The map of bias paths
The graph separates the effect sought, reverse causality, selection and common variables. Each path calls for a different design or diagnostic.
An apparent return of €18 falls to €7 after identification
A historical regression attributes €18 of sales to every thousand euros of media. Yet budgets are increased before high-demand weeks and during promotions.
The decomposition estimates €6 of bias linked to forecast demand, €3 to promotions and €2 to season, that is €11. A comparable geographic test measures an effect of €7, consistent with 18 − 11.
The usable causal effect is €7 within the scope of the test. The €18 return remains a descriptive association and must not guide the allocation.
Acceptance conditions
What must be true to act.
- 01
Population, intervention, comparison, outcome and horizon are defined.
- 02
The common causes of the action and the outcome are mapped.
- 03
The design creates independent variation or defends a testable assumption.
- 04
The final claim stays limited to the population and the range observed.
Limits
What this analysis does not prove.
- A causal graph formalises assumptions; it does not make them true.
- Effects can vary by segment, dose, period and interaction with other levers.
- Sensitivity analyses reduce overconfidence but do not reveal an unmeasured confounder.
References
Works cited.
- Pearl, J. (2009), Causality: Models, Reasoning, and Inference.
- Hernán, M. A. & Robins, J. M. (2020), Causal Inference: What If.
- Shadish, W. R., Cook, T. D. & Campbell, D. T. (2002), Experimental and Quasi-Experimental Designs.