Panel Data Causal Impact Analysis
Also known as: Panel CausalImpact, multi-unit causal impact, panel BSTS causal inference, panel structural time-series causal analysis
Panel data causal impact analysis extends the Bayesian structural time-series approach of Brodersen et al. (2015) to multi-unit panel settings, estimating the counterfactual for several treated units simultaneously using control units as a donor pool. It produces credible intervals for the causal effect at each post-intervention time point, aggregated across units and periods.
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 panel data causal impact analysis when you have an intervention that simultaneously affects multiple treated units, panel time-series data with a clear pre- and post-intervention window, and a set of untreated control units serving as a donor pool. It is particularly valuable when the number of treated units is moderate (2-30) and no single unit alone provides enough data for reliable inference. It is not appropriate when there is no clean intervention date, when the control pool is small or poorly matched, when the pre-period is very short (fewer than 20 time points), or when units have highly heterogeneous dynamics that a shared model would distort.
Strengths & limitations
- Handles multiple treated units simultaneously, pooling statistical power and producing unit-specific as well as aggregate effect estimates.
- Automatic covariate selection via spike-and-slab priors reduces the risk of overfitting when many control units are available.
- Full Bayesian posterior yields genuine credible intervals rather than asymptotic approximations, reflecting uncertainty about the counterfactual.
- Accommodates local trends and seasonal patterns through structural time-series components, making it robust to non-stationary outcomes.
- Produces intuitive visualizations: observed vs. counterfactual trajectories and cumulative impact plots for each unit.
- Requires a sufficient pre-intervention period (typically at least 20 time points) to fit reliable state-space components and covariate weights.
- Computationally demanding: MCMC fitting over many units and time points can be slow without tuned samplers.
- The counterfactual validity depends on the quality of the control donor pool; poorly matched control units produce misleading intervals.
- Like all causal impact methods, it cannot handle spillovers from treated to control units, which violate the stable unit treatment value assumption.
Frequently asked
How does this differ from running CausalImpact separately for each treated unit?
Running separate models ignores correlations across treated units and discards shared information in the pre-period. The panel version pools covariate weights and state-space parameters, yielding more stable estimates — especially when individual pre-periods are short — and naturally aggregates effects across units with correctly propagated uncertainty.
How many pre-intervention time points are needed?
As a practical rule, at least 20 pre-intervention time points per unit are recommended so that the BSTS model can reliably learn trend and seasonal structure. Shorter pre-periods lead to wide, uninformative credible intervals.
What makes a good donor pool?
Control units should be plausibly unaffected by the intervention, share similar pre-intervention dynamics with the treated units, and be drawn from the same population. A poor donor pool — units with structurally different trends — undermines the counterfactual regardless of how sophisticated the model is.
How should I validate the causal claim?
Conduct placebo tests by artificially shifting the intervention date into the pre-period and checking that the model detects no spurious effect. Also apply the model to known-null control units and confirm near-zero estimated effects.
Can I use this with an unbalanced panel?
Yes, but missing observations complicate the MCMC and may require unit-specific imputation or a careful handling of missing-at-random assumptions within the state-space framework.
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 ↗
- Abadie, A., Diamond, A., & Hainmueller, J. (2010). Synthetic control methods for comparative case studies: Estimating the effect of California's tobacco control program. Journal of the American Statistical Association, 105(490), 493-505. DOI: 10.1198/jasa.2009.ap08746 ↗
How to cite this page
ScholarGate. (2026, June 3). Panel Data Causal Impact Analysis. ScholarGate. https://scholargate.app/en/causal-inference/panel-data-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
- Panel Data Difference-in-DifferencesCausal inference↔ compare
- Panel Data Interrupted Time SeriesCausal inference↔ compare
- Panel Data Synthetic Control MethodCausal inference↔ compare
- Synthetic Control MethodCausal inference↔ compare