On this page
- Core mechanism
- Synthetic control vs matched market tests (MMTs)
- Synthetic control vs Difference-in-Differences (DiD)
- Synthetic control vs geo lift testing
- Donor selection and control quality
- Comparison with platform-native lift studies
- Geo testing implementation with synthetic control weights
- Tools and platforms
- Research: SCM vs panel-aware Double Machine Learning (DML)
- Key terms
Synthetic Control Method
Synthetic Control Method
The Synthetic Control Method (SCM) is a causal inference technique that estimates what would have happened in the absence of an intervention by constructing a weighted combination of untreated comparison units — the "synthetic control" — whose pre-intervention trajectory closely matches the treated unit's history. The gap between the treated unit's observed post-intervention outcome and the synthetic control's predicted trajectory is taken as the causal effect estimate. This page records what sources report; it does not advise.
Core mechanism
Per Haus.io (retrieved 2026-11-05): synthetic control builds a counterfactual by creating a weighted blend of several untreated regions, rather than relying on a single matched market or on pure randomisation. "Instead of cherry-picking one similar region, a handful of markets make up your control group."
Per Recast ("Deep-Dive: Synthetic Controls", retrieved 2026-11-05): weights are chosen to minimise the pre-treatment difference between the treatment group and synthetic control — "a convex combination summing to one." The optimal weights are obtained via constrained regression over the pre-intervention period.
Per MetricGate (retrieved 2026-11-05): the causal effect estimate at each post-period time point is the gap between the treated unit's observed outcome and the synthetic control's predicted outcome. This delivers a time-series of treatment effects, not a single average.
Per MetricGate (retrieved 2026-11-05): inference relies on placebo tests — applying the same procedure to each untreated unit and comparing the treated unit's observed effect to the resulting distribution — rather than standard p-values, because SCM is designed for settings with few treated units (sometimes only one). A common diagnostic is the post/pre RMSPE ratio: a ratio above roughly 2–5 is suggested as indicating the effect is large relative to pre-period noise.
Synthetic control vs matched market tests (MMTs)
Per Haus.io (retrieved 2026-11-05): matched market tests (MMTs) pair one treated region with a single similar control region. Any outlier event in that single control region adds noise across the entire experiment. Synthetic control mitigates this by distributing the counterfactual across multiple donor regions.
Per Haus.io (vendor simulation, retrieved 2026-11-05): using a representative customer's geo-based holdout data across 210 DMAs over a six-month pre-treatment period, Haus's synthetic control approach was described as "4X more precise" than matched market tests (measured by variance in estimated lift).
Note on vendor confidence: the "4X precision" claim and the illustrative scenario are both Haus-reported. No independent peer-reviewed validation was found in sources retrieved.
Synthetic control vs Difference-in-Differences (DiD)
Per MetricGate (retrieved 2026-11-05): synthetic control does not require the parallel trends assumption that Difference-in-Differences depends on, because it explicitly models pre-period levels and trends through the weighting process.
Per MetricGate (retrieved 2026-11-05): DiD becomes difficult or infeasible when only one or a very small number of units are treated; synthetic control is designed precisely for this scenario.
Per MetricGate (retrieved 2026-11-05): large discrepancies between DiD and synthetic control estimates suggest the parallel trends assumption required by DiD is violated; MetricGate recommends running both as a robustness check.
Per Statsig (retrieved 2026-11-05): Microsoft data scientists are described as often pairing SCM with DiD analysis, treating convergent results as increased confidence in the estimate.
Per a practitioner guide (jgreathouse9.github.io, retrieved 2026-11-05): DiD or Generalised SCM is recommended when more than ten units are treated, reserving classic synthetic control for situations with a limited number of treated units.
Synthetic DiD: per jgreathouse9.github.io (retrieved 2026-11-05), Arkhangelsky et al. (2021) described a hybrid "Synthetic DiD" approach combining the advantages of both methods.
Synthetic control vs geo lift testing
Per MetricGate (retrieved 2026-11-05): Geo-Lift Testing randomises treatment across many regions, providing experimental lift estimates without relying on matching assumptions, whereas synthetic control requires pre-period matching and donor selection when randomisation across many regions is not possible.
Per MetricGate (retrieved 2026-11-05): synthetic control works when only one market is treated, unlike geo lift tests that need multiple treated regions to support randomisation.
Donor selection and control quality
Per Statsig (retrieved 2026-11-05): control selection is described as "the number one reason synthetic control analyses fail," with Measured's geo-testing research cited as finding that random market selection can badly distort results.
Per Augmented SCM (Statsig, retrieved 2026-11-05): when pre-period fit is poor, practitioners move to Augmented SCM; Bayesian Structural Time Series (BSTS) or parametric models are recommended when donor units are not credible or the pre-period is short.
Comparison with platform-native lift studies
Per SegmentStream (retrieved 2026-11-05): in-platform lift studies (Meta, Google, TikTok, Snap) are critiqued because test groups receive better tracking (pixels, click IDs) than control groups (server-side CAPI only), the methodology is opaque, and the platform selling the ads also grades them.
Per SegmentStream (retrieved 2026-11-05): geo holdouts run typically 2–4 weeks and cannot capture brand effects that compound over months.
Geo testing implementation with synthetic control weights
Per SegmentStream (retrieved 2026-11-05): in a geo holdout using synthetic control weights, the same pre-fit coefficients (e.g. California ×0.64, Ohio ×1.19, Nevada ×2.08) are applied during evaluation, rather than comparing raw absolute sales volumes. The key safeguard is Minimum Detectable Effect (MDE) calculated upfront from region count, volume per region, and baseline variance.
Per Google Developers — Meridian GeoX documentation (retrieved 2026-11-05): Meridian GeoX recommends three geo test strategies:
- Heavy-up: increase spend in treatment geos — measures response to additional advertising
- Go-dark: stop ads in treatment geos — measures contribution of existing advertising
- Holdback: serve ads only in treatment geos, hold out control geos — measures new-market launch lift
Per Google Developers — Meridian GeoX documentation (retrieved 2026-11-05): recommended total timeline is 7–14 weeks (up to 2 weeks pretest planning, 3–10 weeks active test, at least 2 weeks cool-down), and the test should cover at least one full purchase cycle. (as-of 2026-11-05)
Tools and platforms
Per Recast (retrieved 2026-11-05): Meta's open-source GeoLift package and Recast's own GeoLift tooling are cited as tools for geo-based experimentation using synthetic control approaches.
Per Haus.io (retrieved 2026-11-05): Haus uses a "layered model approach" under its Causal Intelligence (cAI) system: for each geo-test it generates multiple models with different assumptions, checks each on an out-of-sample placebo period, and uses AI to weight the models by out-of-sample predictive accuracy. Haus also runs hundreds to thousands of placebo tests on historical data for each experiment. (as-of 2026-11-05, vendor self-description)
Per business.google.com (retrieved 2026-11-05): Google says Meridian (the broader open-source ecosystem including GeoX) passed one million GitHub downloads in its first 18 months. (as-of 2026-11-05, self-reported Google figure)
Research: SCM vs panel-aware Double Machine Learning (DML)
The same paper cites Uber, Airbnb, and LinkedIn as having each deployed causal inference for geo-level marketing measurement: Uber built a spline-regularised learner for causal marketing lift; Airbnb uses causal inference to estimate guest-demand elasticities; LinkedIn proposed an asymmetric-budget-split design for nation-wide lift measurement.
Key terms
| Term | Meaning |
|---|---|
| Donor pool | The set of untreated units from which synthetic control weights are drawn |
| Convex combination | Weighted sum where all weights are non-negative and sum to one |
| Placebo test | Running the SCM procedure on untreated units to generate a null distribution for inference |
| RMSPE ratio | Root Mean Square Prediction Error post-period / pre-period — used to gauge effect significance |
| Parallel trends | The assumption required by DiD that treated and control units would have followed the same trajectory absent treatment — SCM does not require this |
| Augmented SCM | Extension that adds a bias-correction term when pre-period fit is imperfect |
| Synthetic DiD | Hybrid of SCM and DiD proposed by Arkhangelsky et al. (2021) |
| MDE | Minimum Detectable Effect — the smallest real effect the experiment design can reliably detect |