Dynamic Counterfactual Impact Evaluation
Also known as: dynamic CIE, dynamic treatment evaluation, time-varying counterfactual analysis, longitudinal counterfactual evaluation
Dynamic Counterfactual Impact Evaluation (dynamic CIE) extends standard counterfactual program evaluation to settings where treatment is assigned sequentially across multiple periods. Rather than comparing a single treated versus untreated state, it estimates the causal effect of entire treatment trajectories or regimes, accounting for how intermediate outcomes and time-varying covariates feed back into subsequent treatment decisions.
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 dynamic CIE when treatment is time-varying or delivered in multiple doses across periods, and when intermediate outcomes or covariates affected by past treatment influence future treatment selection. Typical contexts include labor market programs with repeated participation, clinical trials with adaptive dosing, and social interventions with re-enrollment. It requires longitudinal data with rich time-varying covariates measured at each period before treatment assignment. Do not use it when treatment is a single point-in-time decision with no feedback loop — standard (static) counterfactual evaluation is sufficient there, and the extra complexity of dynamic methods is unwarranted.
Strengths & limitations
- Handles time-varying treatment and time-varying confounding, which static methods (standard DiD, propensity score matching) cannot address without bias.
- Identifies causal effects of entire treatment regimes or sequences, answering policy-relevant questions about optimal timing and dosage.
- G-computation and marginal structural model frameworks provide consistent estimation under correct model specification.
- Explicitly models feedback between past treatment and future covariates, making the causal assumptions transparent and testable.
- Compatible with a range of outcome types (continuous, binary, time-to-event) via appropriate weighting or outcome models.
- Requires rich longitudinal data with time-varying covariates measured before each treatment decision — often unavailable in administrative datasets.
- Consistency of IPW estimators depends on correct specification of the propensity model at every period; misspecification compounds across periods.
- Defining treatment regimes requires substantive judgment; the choice of regimes to compare shapes all conclusions.
- Sample sizes needed for reliable estimation grow with the number of periods and the complexity of regimes.
- The sequential ignorability (no unmeasured time-varying confounding) assumption is strong and cannot be tested directly.
Frequently asked
How does dynamic CIE differ from standard counterfactual impact evaluation?
Standard CIE compares a treated group to an untreated counterfactual at a single point in time. Dynamic CIE compares entire sequences of treatment decisions across multiple periods, and it explicitly adjusts for the feedback between past treatment, intermediate outcomes, and future treatment selection — a complication that static methods ignore and that leads to bias when treatment is time-varying.
What is sequential ignorability, and why does it matter?
Sequential ignorability (also called sequential conditional independence) requires that at each period, treatment assignment is independent of future potential outcomes given all observed history up to that point. It is the dynamic analogue of the ignorability assumption in propensity score methods. If unmeasured factors affect both current treatment and future outcomes, the assumption fails and estimates are biased.
Can I use this method with administrative panel data?
Yes, but with caution. Administrative data often lack the rich set of time-varying covariates needed at each decision point. If key time-varying confounders (e.g., health status, job search intensity) are not recorded, positivity and sequential ignorability may be violated. Enrich the data with linked surveys or auxiliary records where possible.
What should I do if propensity score weights are very large?
Large weights signal near-violations of the positivity assumption — some covariate histories are very rare under one treatment regime. Stabilized weights (dividing by the marginal probability of the observed treatment) reduce variance. Trimming extreme weights (e.g., capping at the 99th percentile) is also common, though it introduces some bias in exchange for reduced variance.
Is dynamic CIE the same as a marginal structural model?
Not exactly. A marginal structural model (MSM) is one specific estimation framework within the dynamic CIE family — it specifies the marginal mean of the potential outcome under each regime and is fitted with stabilized inverse probability weights. G-computation is an alternative estimation approach that models the outcome directly at each period. Dynamic CIE is the broader evaluation paradigm; MSMs and g-computation are the main tools for implementing it.
Sources
- Robins, J. M. (1986). A new approach to causal inference in mortality studies with a sustained exposure period — application to control of the healthy worker survivor effect. Mathematical Modelling, 7(9-12), 1393-1512. DOI: 10.1016/0270-0255(86)90088-6 ↗
- Lechner, M. (2009). Sequential causal models for the evaluation of labor market programs. Journal of Business and Economic Statistics, 27(1), 71-83. DOI: 10.1198/jbes.2009.0006 ↗
How to cite this page
ScholarGate. (2026, June 3). Dynamic Counterfactual Impact Evaluation. ScholarGate. https://scholargate.app/en/causal-inference/dynamic-counterfactual-impact-evaluation
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.
- Counterfactual Impact EvaluationCausal inference↔ compare
- Difference-in-DifferencesEconometrics↔ compare
- Dynamic Difference-in-DifferencesCausal inference↔ compare
- Marginal Structural ModelCausal inference↔ compare
- Panel Event StudyCausal inference↔ compare
- Propensity Score WeightingCausal inference↔ compare