Multi-period Causal Impact Analysis
Multi-period Bayesian Causal Impact Analysis · Also known as: multi-period CausalImpact, staggered causal impact, repeated-period causal impact, multi-wave CausalImpact
Multi-period Causal Impact Analysis extends the Bayesian structural time-series framework of Brodersen et al. (2015) to settings where an intervention occurs across multiple distinct periods, is applied at staggered times to different units, or where researchers wish to evaluate cumulative and period-specific effects within a single unified model. It builds a synthetic counterfactual from control covariates and projects it across each intervention window to quantify causal effects.
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 multi-period Causal Impact when your intervention occurs at multiple points in time, is rolled out to different units at staggered dates, or has distinct implementation phases (e.g., pre-launch, launch, post-launch follow-up). The method requires a continuous outcome time series with a sufficient pre-intervention baseline (ideally at least as long as the longest post-period), and at least one unaffected control covariate series. It excels when the control covariates closely track the outcome in the pre-period and when the goal is to detect period-specific changes in effect magnitude. Do not use it when there is no clean pre-intervention baseline, when all available control series are themselves affected by the treatment, when sample sizes are very small (fewer than ~30 pre-period observations per window), or when the outcome is purely cross-sectional with no temporal structure.
Strengths & limitations
- Handles staggered and multi-wave interventions within a single coherent Bayesian framework, producing both window-specific and cumulative effect estimates.
- Builds a data-driven synthetic counterfactual from control covariates via spike-and-slab variable selection, avoiding the need to manually specify a comparison group.
- Full posterior inference yields calibrated credible intervals that honestly reflect model uncertainty, unlike frequentist point-estimate approaches.
- Accommodates complex temporal patterns — trend, seasonality, autocorrelation — that would violate simpler interrupted time-series assumptions.
- Period-specific effect estimates allow researchers to diagnose whether treatment effects grow, decay, or reverse across successive intervention windows.
- Requires a sufficiently long and stable pre-intervention series; short baselines lead to poorly identified counterfactuals and wide credible intervals.
- Quality of the counterfactual depends critically on the availability of valid control covariates that are correlated with the outcome but unaffected by treatment; if none exist, the method degenerates to a simple trend extrapolation.
- Computationally more demanding than standard CausalImpact because MCMC or state-space estimation is run for each period or cohort.
- Assumes no interference between units (SUTVA); if treatment in one period or cohort spills over to others, the counterfactual is contaminated.
- Staggered designs require careful cohort-specific baseline definition; if windows overlap or pre-periods are short, period effects are confounded.
Frequently asked
How does multi-period CausalImpact differ from running standard CausalImpact separately for each period?
Running separate models for each window ignores that the pre-period baseline and model parameters should be shared or consistently specified. A joint multi-period specification ensures that the same counterfactual model governs all windows, making period-to-period comparisons of effect sizes coherent and reducing the risk of overfitting each window independently.
What constitutes a valid control covariate in this framework?
A control covariate must be correlated with the outcome in the pre-intervention period and must not itself be affected by the intervention. Good examples include concurrent metrics from untreated markets, regions, or product lines. If all plausible covariates are contaminated, the method reverts to a local-trend extrapolation, which is far weaker.
How long should the pre-intervention baseline be?
As a practical rule, the pre-period should be at least as long as the longest single post-intervention window, and ideally two to three times as long. Short baselines (fewer than ~30 observations) make it hard for the BSTS model to reliably estimate trend and seasonality parameters, leading to wide and uninformative credible intervals.
Can I apply this method when units receive the intervention at different calendar times?
Yes. In staggered rollout designs, each cohort's own pre-intervention observations define its baseline. Effects are estimated separately per cohort and then aggregated — typically as a weighted average — to produce an overall average treatment effect across the full sample period.
What if my outcome series is seasonal?
Seasonality can be incorporated directly into the BSTS state-space model as a trigonometric or dummy-variable seasonal component. Explicitly modelling seasonality prevents it from being absorbed into the estimated treatment effect, which would happen if the intervention coincided with a predictable seasonal upturn or downturn.
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 ↗
- Bojinov, I., & Shephard, N. (2019). Time series experiments and causal estimands: exact randomization tests and trading. Journal of the American Statistical Association, 114(528), 1665-1682. DOI: 10.1080/01621459.2018.1527225 ↗
How to cite this page
ScholarGate. (2026, June 3). Multi-period Bayesian Causal Impact Analysis. ScholarGate. https://scholargate.app/en/causal-inference/multi-period-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 Causal Impact AnalysisCausal inference↔ compare
- Causal Impact AnalysisCausal inference↔ compare
- Difference-in-DifferencesEconometrics↔ compare
- Interrupted Time SeriesCausal inference↔ compare
- Multi-period Interrupted Time SeriesCausal inference↔ compare
- Synthetic Control MethodCausal inference↔ compare