Synthetic Control for Interventions That Cannot Be A/B Tested

Market-level campaigns, country launches, rebrands and major product changes are hard to evaluate. A/B testing is often impossible, because audiences cannot be split into independent groups, or too expensive. Teams then fall back on weak correlations or simple pre/post comparisons, which say little about the true effect of a large investment.

Synthetic control replaces the missing control group with a constructed one. It takes a donor pool of untreated units, such as regions, customer segments or product lines, and finds the weighted combination that best reproduces the treated unit's outcome levels, trends and relevant covariates before the intervention. After the intervention, that weighted combination serves as the counterfactual: an estimate of what the treated unit would have done without the change.

Its main advantage over difference-in-differences is the assumption it avoids. DiD relies on parallel trends between treated and control groups, which often fails in practice. Synthetic control instead optimises for pre-treatment fit, and with a long enough pre-period it can also absorb unobserved confounders that move the outcome over time. That makes it effective at separating a genuine effect from seasonality and market-wide fluctuations.

Inference works differently from a standard test. The usual approach is a placebo test: rerun the method with each donor unit treated as if it had received the intervention, and check whether the real unit's post-intervention gap stands out from the placebo gaps.

Implementation requires consistent historical time series for every unit, which a well-modelled GA4 BigQuery export can provide. For the modelling, R's Synth package and Python's pysyncon are common choices, and CausalImpact offers a related Bayesian approach. Data completeness matters as much as the method: if tracking coverage differs between units, for instance because of uneven consent rates, the synthetic control inherits that bias.

Synthetic control is not universal. It needs a stable pre-period, an adequate donor pool and careful interpretation. For interventions that cannot be randomised, however, it replaces intuition with a defensible estimate.