Documentation menu

Causal methods

The methods available in the Causal Lab — geo experiments, difference-in-differences, synthetic control, double ML, Bayesian impact and pre/post.

Causal methods

The Causal Lab offers several methods. Each suits a different situation; choosing the right one is most of the work.

The methods

Geo experiment — split regions into treatment and control, change spend in one, and measure the difference. Ideal for channel or campaign incrementality.
Difference-in-differences (DiD) — compare the change in a treated group against the change in an untreated group over the same period, cancelling out shared trends.
Synthetic control — build a weighted "synthetic" version of the treated unit from untreated ones, and compare against it. Good when you have one treated region and many candidates for the comparison.
Double machine learning (Double ML) — estimate an effect while controlling for many confounding variables.
Bayesian structural impact — model the counterfactual time series and measure the gap after an intervention, with credible intervals.
Pre/post — the simplest before-and-after comparison; useful as a quick read but the weakest at ruling out other causes.

Choosing a method

Favour designs with a genuine control (geo, DiD, synthetic control) over pre/post when you can, because they guard against mistaking a trend for an effect. The Lab guides you toward a method that fits your data.

Reading the result

Every method returns an estimated effect with uncertainty. Treat a tight, robust estimate as actionable and a wide, noisy one as suggestive — and report the uncertainty alongside the number.

Documentation menu

Causal methods

The methods available in the Causal Lab — geo experiments, difference-in-differences, synthetic control, double ML, Bayesian impact and pre/post.

Causal methods

The Causal Lab offers several methods. Each suits a different situation; choosing the right one is most of the work.

The methods

Geo experiment — split regions into treatment and control, change spend in one, and measure the difference. Ideal for channel or campaign incrementality.
Difference-in-differences (DiD) — compare the change in a treated group against the change in an untreated group over the same period, cancelling out shared trends.
Synthetic control — build a weighted "synthetic" version of the treated unit from untreated ones, and compare against it. Good when you have one treated region and many candidates for the comparison.
Double machine learning (Double ML) — estimate an effect while controlling for many confounding variables.
Bayesian structural impact — model the counterfactual time series and measure the gap after an intervention, with credible intervals.
Pre/post — the simplest before-and-after comparison; useful as a quick read but the weakest at ruling out other causes.

Choosing a method

Reading the result

Every method returns an estimated effect with uncertainty. Treat a tight, robust estimate as actionable and a wide, noisy one as suggestive — and report the uncertainty alongside the number.