Synthetic Control
Consensus definition
Introduced by Abadie and Gardeazabal (2003) and formalised by Abadie, Diamond and Hainmueller (2010)1, the synthetic control method constructs a counterfactual for a single treated unit as a convex combination of untreated donor-pool units. Weights are chosen to minimise the gap between the treated unit and the weighted composite across the pre-treatment period — outcomes and relevant covariates alike. A good pre-period fit is a necessary condition for credibility: it is the visible evidence that the synthetic unit would have tracked the treated one absent the intervention2. After the intervention, the period-by-period gap between actual and synthetic outcomes is the estimated effect; inference typically uses placebo permutations across donor units to build a null distribution1. Its transparency — weights and fit are directly inspectable — drove broad uptake across economics and social science2.
rhinegold operator refinement
Rhinegold's reframe: when you run one intervention on one brand, page or domain, there is no natural partner identical to your treated unit. A synthetic control solves this by drawing on a donor pool of untreated pages, domains or keyword clusters and finding the data-driven blend that best matches the pre-intervention trajectory — not a hand-picked lookalike. The resulting synthetic series is the counterfactual you can defend to a client, and the post-intervention gap is the causal estimate.
Operational use
Collect weekly or monthly outcomes (clicks, impressions, citation share, share of voice) for the treated unit and a pool of untreated candidates over a meaningful pre-period; optimise donor weights to minimise pre-period fit error; then plot the post-intervention gap and run placebo permutations across donors to judge whether the gap is unusual. The post/pre RMSPE ratio quantifies credibility.
Measurement boundary
SCM needs a sufficiently long pre-period — bias falls as it grows2 — and a donor pool genuinely unaffected by the intervention (no spillover). If the treated unit lies outside the donor pool's range, interpolation fails and bias can be large; very large pools raise overfitting risk. It yields a per-period gap, not a single coefficient with a standard error, so communication needs care.
Distinct from
From the control group: a conventional control is fixed by design or matching; SCM builds its comparison algorithmically from a weighted blend of many units after observing the pre-period. From difference-in-differences: DiD asserts parallel pre-trends as an untestable assumption; SCM makes the pre-period fit explicit and optimised — a visible diagnostic instead of an assertion. From the counterfactual: that is the general concept; SCM is one concrete way to construct it when no clean single-unit comparison exists.
Common mistakes
- Selecting the donor pool after inspecting post-intervention outcomes — cherry-picking donors invalidates the method and inflates the apparent effect.
- Accepting a poor pre-period fit and reading the post-period gap as causal — if the synthetic unit did not track the treated unit before, the divergence is uninformative.
- Using too few pre-periods, so weights fit noise rather than signal.
- Ignoring spillover: if the intervention indirectly moves donor units, they become contaminated controls.