Bayesian Causal Impact Analysis
Bayesian Causal Impact Analysis via Structural Time Series · Also known as: CausalImpact, Bayesian structural time series causal inference, BSTS causal impact, Bayesian intervention analysis
Bayesian Causal Impact Analysis uses a Bayesian structural time series (BSTS) model to estimate the causal effect of an intervention on a time series outcome. Developed by Brodersen and colleagues at Google in 2015, it builds a probabilistic counterfactual — what the series would have looked like without the intervention — from pre-intervention data and optional control covariates, then compares it with the observed post-intervention values to produce a fully Bayesian posterior over the causal effect.
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.
When to use it
Use Bayesian Causal Impact when you have a single treated time series with a clearly defined intervention date, a meaningful pre-intervention history (ideally at least 60 observations), and access to control covariates (other series unaffected by the intervention) that co-move with the outcome. It is particularly well-suited to digital marketing, product launches, policy roll-outs, or any setting where randomisation is impossible and you need a probabilistic effect estimate. Do not use it when the pre-intervention period is very short (fewer than roughly 30 time points), when the intervention is gradual with no clear start date, or when no valid control series or historical baseline exists to anchor the counterfactual — in such cases interrupted time series or difference-in-differences may be more appropriate.
Strengths & limitations
- Produces a full posterior distribution over the causal effect, providing genuine uncertainty quantification including credible intervals and posterior probabilities.
- Automatically handles trend, seasonality, and structural breaks through the flexible BSTS state-space formulation.
- Incorporates control covariates via spike-and-slab regression, performing automatic variable selection and reducing confounding from unobserved common factors.
- Yields intuitive visual output: observed series, counterfactual, and pointwise and cumulative impact plots with uncertainty bands.
- Does not require a control group of units, only control time series, making it applicable where parallel groups are unavailable.
- Scales to high-dimensional predictor sets without manual covariate selection.
- Requires a sufficiently long and stable pre-intervention period; short or non-stationary pre-periods can produce unreliable counterfactuals.
- The counterfactual quality depends heavily on the quality and number of control covariates; without good controls the Bayesian model essentially extrapolates a trend, which may be misleading.
- Cannot handle cases where the intervention affects the control series (spillover), violating the exclusion restriction for the covariates.
- Computationally more intensive than classical interrupted time series methods, though manageable for most applied data sizes.
- Assumes the pre-intervention model structure (trend, seasonality) is the correct one; model misspecification can bias the counterfactual.
Frequently asked
What makes this Bayesian rather than a standard interrupted time series?
Standard interrupted time series fits a regression with a structural break and reports frequentist confidence intervals. Bayesian Causal Impact places priors on all model parameters, propagates uncertainty through a state-space model, and delivers a full posterior distribution over every quantity — counterfactual trajectories, pointwise effects, and cumulative impact — providing richer and more honest uncertainty quantification.
How many pre-intervention time points do I need?
The BSTS model needs enough pre-intervention data to learn the trend, seasonal pattern, and covariate relationships reliably. In practice, at least 60 pre-intervention observations are recommended for stable estimates; fewer than 30 typically produces wide and unreliable credible intervals.
Do I need control series (covariates)?
They are not strictly required, but they are highly recommended. Without control covariates, the counterfactual is based only on the outcome's own history and may miss confounders. Good control series are variables that correlate with the outcome in the pre-period but were not affected by the intervention.
What does the posterior probability of a causal effect mean?
It is the probability, given the model and data, that the intervention had a positive (or negative) effect. A value of 0.97 means that 97% of the posterior mass falls on the side of a positive effect. It is not a classical p-value, but it plays an analogous communicative role and is often easier to interpret.
When should I prefer difference-in-differences over this method?
Prefer DiD when you have multiple treated and control units observed over a relatively small number of time periods and parallel-trends can be validated with pre-treatment data. Bayesian Causal Impact is preferable when you have a single treated unit with a long time series and good control covariates but no natural parallel comparison group.
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 ↗
- Scott, S. L., & Varian, H. R. (2014). Predicting the present with Bayesian structural time series. International Journal of Mathematical Modelling and Numerical Optimisation, 5(1-2), 4-23. DOI: 10.1504/IJMMNO.2014.059942 ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Causal Impact Analysis via Structural Time Series. ScholarGate. https://scholargate.app/en/causal-inference/bayesian-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.
- Causal Impact AnalysisCausal inference↔ compare
- Difference-in-DifferencesEconometrics↔ compare
- Interrupted Time SeriesCausal inference↔ compare
- Synthetic Control MethodCausal inference↔ compare