A quick-commerce company runs a large outdoor and TV campaign in Hyderabad only. Eight weeks later, orders in Hyderabad are up. The obvious question is how much of that rise the campaign caused. You cannot randomise a city, and there is only one Hyderabad. Comparing it with Bengaluru alone is fragile: Bengaluru has a different customer mix, different competitors and its own local events. Synthetic control offers a better comparison by constructing one.
The core idea
Instead of picking a single comparison city, you build a weighted blend of several untreated cities (the “donor pool”) whose combined history tracks the treated city as closely as possible before the intervention. That blend is the synthetic Hyderabad. After the campaign starts, the synthetic version continues on its path, showing what Hyderabad would plausibly have looked like without the campaign. The gap between real and synthetic Hyderabad is the estimated effect.
The method was developed by Alberto Abadie and co-authors. Its best-known applications are Abadie and Gardeazabal’s study of the economic cost of conflict in the Basque Country (2003) and Abadie, Diamond and Hainmueller’s study of California’s 1988 tobacco control programme (2010), where a weighted mix of other US states served as “synthetic California”.
A small worked example
Suppose the optimisation finds that the best blend is 50% Pune, 30% Ahmedabad and 20% Jaipur. The weights are non-negative and sum to one, so the synthetic city is a genuine mixture, not an extrapolation. Weekly orders, as an index:
| Period | Pune | Ahmedabad | Jaipur | Synthetic Hyderabad | Actual Hyderabad |
|---|---|---|---|---|---|
| Pre, month 1 | 100 | 90 | 70 | 91.0 | 91 |
| Pre, month 2 | 104 | 95 | 72 | 94.9 | 95 |
| Pre, month 3 | 108 | 99 | 75 | 98.7 | 99 |
| Post campaign | 112 | 102 | 78 | 102.2 | 109 |
Before the campaign, synthetic Hyderabad matches the real one closely. Afterwards, real Hyderabad reaches 109 while the synthetic version reaches 102.2. The estimated campaign effect is about 6.8 index points, roughly 6.7% above what would otherwise have happened. Notice that all cities grew; a simple before-and-after in Hyderabad would have credited all that general growth to the campaign.
In a real analysis, the pre-period would cover many more weeks, and the weights would be chosen by matching not only past outcomes but also predictors such as population, order frequency, average basket size and store density.
How you know whether to believe it
With a single treated unit, there is no conventional standard error. Instead, synthetic control relies on a set of checks.
Pre-period fit. If the synthetic city cannot track the real one before treatment, there is no reason to trust it afterwards. A poor fit usually means the treated unit is unusual (the largest or fastest-growing city, say) and no mix of smaller units can reproduce it.
Placebo tests. Pretend each donor city received the campaign, build a synthetic version for it from the remaining cities, and measure its “effect”. If Hyderabad’s gap is larger than nearly all the placebo gaps, the result is unlikely to be chance. If several untreated cities show similar gaps, the Hyderabad result is not convincing.
Leave-one-out. Drop each donor city in turn and rebuild. If removing Pune makes the effect vanish, the conclusion depends on one city and deserves caution.
Backdating. Pretend the campaign started a month earlier than it did. A credible synthetic control should show no gap during that fake treatment month.
Where it goes wrong
- Contaminated donors. If the TV campaign also reached Pune through national channels, or if Hyderabad’s success drew stock and riders away from neighbouring cities, the donors were affected too.
- Simultaneous local shocks. A competitor exiting Hyderabad in the same month, a city-specific festival or a rain disruption elsewhere will be attributed to the campaign.
- Too short a pre-period. A few data points can be matched by almost any blend, giving a false sense of fit. More history makes the match meaningful.
- Too many donors. A large donor pool full of dissimilar units can overfit noise. Restrict the pool to units that are plausibly comparable.
When to reach for it
Synthetic control is suited to situations with one or a few treated units, a reasonable number of untreated comparison units, and a long, stable history: a regional launch, a state-level regulation, a pricing change in one country, a new warehouse serving one zone. It is more transparent than many alternatives because you can show the weights and the before-and-after chart, and a non-technical audience can see the gap opening up.
It complements difference-in-differences rather than replacing it. DiD uses a simple average of comparison units and assumes parallel trends; synthetic control chooses the weights to make the pre-period match, which often makes the parallel-trends story more believable. Several geo-experiment tools used in marketing measurement draw on similar ideas. Where you can plan ahead, the strongest design is to randomise which regions get the campaign; where you cannot, a well-checked synthetic control is a respectable second best.
Written by the RP Strativa Team. The company, cities, weights and figures are illustrative.
Leave a Reply