Causal inference asks whether a defined action produced an observed result. Performance can improve after an intervention because of seasonality, customer mix or another simultaneous change, so timing alone is weak evidence. Marketways designs an experiment or credible comparison that estimates what would otherwise have happened. This helps the client expand actions that work and stop assigning value to convenient coincidence.
The decision this method supports
We use Causal Inference & Experimentation to help clients answer: What would probably have happened without the intervention?
How the method works
Causal inference asks whether an intervention changed an outcome. The difficult part is constructing a credible account of what would probably have happened without the intervention. Randomised experiments can create that comparison directly, while quasi-experimental methods use timing, thresholds or comparison groups when randomisation is impractical.
A business example
A company introduces a new retention offer in one customer group and observes lower churn. Causal analysis examines whether the fall resulted from the offer or from differences in customer mix, seasonality or wider market conditions.
How the client uses the result
Causal methods help leaders decide whether an action produced the result that followed and what might have happened without it. This prevents an attractive KPI movement from being credited to the wrong intervention. The business can expand an action that works, revise one that does not and examine indirect effects before declaring success.
What we deliver
We produce estimates, comparisons, intervals, model results and sensitivity tests. A manager should be able to see the size of the result, the uncertainty around it, the assumptions required and which interpretation the evidence does not support.
Limits and complementary methods
Causal analysis answers a narrower question about the effect of a defined intervention and often requires stronger research design. A predictive model may anticipate outcomes more accurately without explaining what will happen if management changes something.
Selected methods and techniques
We select from these established methods according to the decision, evidence and operating conditions.
- A/B testing / experiments: Randomly assign people, sites or other comparable cases to different versions of a change and compare the outcomes. Random assignment helps separate the effect of the change from pre-existing differences between the groups.
- Before-after analysis: Compare outcomes before and after a change; without a credible comparator or design, the difference cannot establish the change's causal effect.
- Causal analysis: Estimate whether a specified change causes a defined operational outcome and by how much, compared with a stated alternative. Define the intervention, comparator, eligible units, outcome, timing and assumptions before choosing a randomised experiment or a credible quasi-experimental or observational design. Check pre-intervention comparability, confounding, spillovers, missing outcomes, uncertainty and sensitivity to the assumptions. This method estimates the effect of changing something. Root-cause analysis instead investigates mechanisms behind an observed problem, and regression alone usually estimates association rather than a causal effect.
- Causal inference: Estimate whether an action or change actually affects an outcome, rather than simply occurring alongside it. Define the change, what would happen without it and the assumptions that make the comparison credible.
- Causal rival analysis: Set out credible competing explanations for an observed outcome, derive what each would predict and identify evidence that can distinguish among them. Give priority to observations that conflict with a rival rather than evidence compatible with every explanation. This challenges causal interpretation; it does not establish causation unless the design and assumptions support it.
- Causal-loop diagramming: Map how variables in a system influence one another through reinforcing and balancing feedback loops, including delays. It explains how system structure can produce behaviour over time, such as growth, oscillation or unintended consequences. Diagrams are qualitative hypotheses; they do not quantify effects.
- Counterfactual analysis: Ask what would happen, or would have happened, under a different specified condition. A counterfactual is that alternative situation; claiming it shows a causal effect requires a credible explanation of how the change affects the outcome.
- Counterfactual fairness testing: Use an explicit cause-and-effect model to ask whether a prediction for the same person would change if a protected characteristic, such as sex, had been different. The model must account for how other characteristics would change too; simply editing one data field is not enough.
- Difference-in-differences: Estimate an intervention’s effect by subtracting the change in a comparison group from the change in the treated group. The key assumption is that, without the intervention, the groups would have followed parallel trends; similar earlier trends help assess but do not prove that assumption.
- Event-study analysis: Estimate how an outcome changes across periods before and after an intervention relative to an omitted baseline period, ordinarily using comparison units and a panel or difference-in-differences design. It can reveal anticipatory effects, dynamic effects and the plausibility of pre-treatment trend assumptions. Its interpretation depends on the identification assumptions and the estimator used, especially when adoption timing differs between units.
- Experimental / quasi-experimental design review: Check whether a study can credibly separate the effect of a change from other influences. Review randomised experiments and quasi-experiments, which use comparisons without full random assignment, for differences between groups and problems in carrying out the study.
- Interrupted time-series analysis: Estimate whether an outcome’s level or trend changed after an intervention using repeated observations before and after the intervention. Model the pre-intervention trend, immediate level change, post-intervention trend and relevant autocorrelation or seasonality. A change at the intervention point is not necessarily caused by the intervention if other events occurred at the same time; a credible comparison series strengthens attribution.
Parent method family
Statistics & Econometrics explains how this method connects to adjacent methods and relevant services.
Related service families
These service families contain business questions supported by this method. Service pages link to the wider method family so readers can understand the complete analytical approach.