Causal Impact Analysis
Bayesian Structural Time-Series Causal Impact Analysis · Also known as: CausalImpact, BSTS causal inference, Bayesian causal impact, counterfactual time-series analysis
Causal Impact Analysis, introduced by Brodersen et al. (2015) at Google, uses Bayesian structural time-series models to estimate what would have happened to an outcome had an intervention never occurred. By constructing a probabilistic counterfactual from pre-treatment data and control covariates, it quantifies point-in-time and cumulative treatment effects with full posterior uncertainty intervals.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
+13 more
When to use it
Use Causal Impact when you have a single treated unit observed as a time series, a clearly defined intervention date, and at least one unaffected control covariate series (a related metric that was not exposed to the treatment). It is especially well-suited for digital marketing, product launches, A/B tests without a held-out control group, and policy evaluations where the outcome is a time series. The pre-intervention window should be sufficiently long (typically at least three times the post-intervention window) to fit a reliable model. Do not use it when the intervention affected all available control series (contamination), when the pre-intervention series is too short to calibrate the model, or when the assumption that controls are unaffected by the treatment cannot be defended.
Strengths & limitations
- Provides full posterior uncertainty quantification for pointwise and cumulative effects, rather than single-point estimates.
- Automatically constructs a synthetic counterfactual from multiple control covariates without requiring a matched control group of units.
- Handles complex time-series structure including local trends and seasonality via the BSTS framework.
- Applicable to single treated units — situations where Difference-in-Differences or matching methods require many treated observations.
- Spike-and-slab prior on regression coefficients performs implicit variable selection among candidate controls.
- Requires that control covariate series are genuinely unaffected by the intervention; if controls are contaminated, the counterfactual will absorb part of the treatment effect.
- A short pre-intervention window yields a poorly calibrated BSTS model, producing wide and unreliable credible intervals.
- The method applies to a single treated unit or aggregate series; it does not directly generalise to panels with multiple treated units at different times.
- Model specification choices (trend type, seasonality, choice of covariates) can materially influence results and require substantive judgment.
Frequently asked
How is Causal Impact different from Difference-in-Differences?
DiD compares average changes across a treatment group and a control group of units. Causal Impact works on a single treated time series and constructs a synthetic counterfactual from control covariate series using a Bayesian state-space model. Causal Impact also provides full posterior distributions rather than point estimates, and naturally accommodates trend and seasonality.
How long should the pre-intervention period be?
A common guideline is for the pre-intervention window to be at least three times as long as the post-intervention window. Longer pre-intervention periods allow the BSTS model to learn the time-series structure and the relationship with controls more reliably. Very short pre-intervention windows (fewer than 30 time points) are generally insufficient.
What if I have no control covariates?
The method can run with a local-level trend model and no covariates, but the counterfactual will then be based solely on the pre-treatment outcome trajectory extrapolated forward. Without controls, any post-treatment shock that is unrelated to the intervention can masquerade as a treatment effect, making the assumption of a stable trend especially demanding.
What does the posterior probability reported by CausalImpact mean?
It is the posterior probability that the cumulative effect has the sign consistent with the direction of impact (e.g., positive). A high posterior probability, together with a credible interval that excludes zero, indicates evidence of a genuine effect. It is not a classical p-value and should not be treated as one.
Can Causal Impact handle seasonal data?
Yes. The BSTS model supports weekly, monthly, or custom seasonal components that are estimated from pre-intervention data and projected forward in the counterfactual. If your time series shows strong seasonality, specifying the correct seasonal period is important for a well-calibrated counterfactual.
Sources
- Brodersen, K. H., Gallusser, F., Koehler, J., Remy, N., & Scott, S. L. (2015). Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics, 9(1), 247-274. DOI: 10.1214/14-AOAS788 ↗
- CausalImpact. Wikipedia. link ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Structural Time-Series Causal Impact Analysis. ScholarGate. https://scholargate.app/en/causal-inference/causal-impact-analysis
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Bayesian Structural Time SeriesBayesian↔ compare
- Difference-in-DifferencesEconometrics↔ compare
- Interrupted Time SeriesCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare
- Synthetic Control MethodCausal inference↔ compare