A relationship between spend and sales is causal only when it compares what happened with what would have happened wi…
A relationship between spend and sales is causal only when it compares what happened with what would have happened without the decision, for the same units and horizon. Demand often drives spend, promotions follow weak periods and channels target customers already near purchase. These paths create attributed performance that was not caused.
Decision and method.
Authorise a recommendation only when major confounding paths are blocked, measured or explicitly left uncertain. Define population, intervention, comparison, outcome and horizon before choosing a design. Map prior variables, targeting criteria, latent demand, competition, mediators and outcome. Prefer randomisation; otherwise document the assumption that makes quasi-experimental variation credible and seek placebo tests or pre-trends. Report the interval, sensitivity to confounders and covered population: local proof is not universal return.
Worked example.
A historical regression attributes €18 in sales to every €1,000 of media, but budgets rise before high-demand weeks and during promotions. Decomposition estimates €6 demand-forecast bias, €3 promotions and €2 seasonality, or €11. A comparable geographic test measures €7, consistent with €18 − €11. The usable causal effect is €7 within the test scope; €18 is descriptive association and must not guide allocation.
Checks and limits.
Average effect, effect on exposed people, a budget tranche and long-term effect are different quantities. A common cause creates a non-causal path; controlling a mediator or collider can introduce bias. A causal graph formalises hypotheses but does not make them true; effects vary by segment, dose, period and interaction, while sensitivity analysis cannot reveal an unmeasured confounder.
Resources and sources.
Download the causal-bias register. Related: estimate what would not have happened, create geographic variation, choose a measurement method. Sources: Pearl (2009); Hernán & Robins (2020); Shadish, Cook & Campbell (2002).