Difference-in-differences vs. synthetic control: choosing the causal estimator for a geo test

Difference-in-differences vs. synthetic control: choosing the causal estimator for a geo test

You ran a geo experiment — switched a channel on in some markets, off in others. Now which method extracts the causal lift?

Difference-in-differences (DiD)
DiD compares the before-after change in treated markets to the before-after change in control markets, subtracting out shared trends. Its load-bearing assumption is parallel trends: absent treatment, test and control would have moved together. If your treated markets were already growing faster, DiD mistakes that for ad effect.

Synthetic control
Synthetic control builds a weighted blend of untreated markets that closely tracks the treated market's pre-period trajectory, then projects that synthetic counterfactual forward. It relaxes the strict parallel-trends demand by matching the pre-trend explicitly, which is why it shines when you have one or few treated units and a rich donor pool.

The tradeoff
DiD is simple, transparent, and works with many treated units, but it fails silently when trends diverge. Synthetic control handles a single treated market elegantly and makes the counterfactual visible, but it can overfit the pre-period and gives messy inference (placebo tests, not clean p-values).

Bottom line for practitioners: With many treated and control markets and credible parallel pre-trends, DiD is the cleaner, more powerful choice. With one or a handful of treated markets — a single test city, a country launch — synthetic control is the right tool because it constructs a bespoke counterfactual. Always plot the pre-period fit first; if it's poor, neither estimate is trustworthy.