Dynamic Entropy Balancing
Dynamic Entropy Balancing for Longitudinal Causal Inference · Also known as: DEB, longitudinal entropy balancing, entropy balancing with time-varying treatment, sequential entropy balancing
Dynamic Entropy Balancing extends the entropy balancing reweighting approach to settings with time-varying treatments in panel or longitudinal data. It constructs unit weights at each time period such that the covariate distributions of treated and comparison units are balanced on specified moments, adjusting sequentially for prior treatment history and time-varying confounders to estimate the causal effect of treatment sequences on outcomes.
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 Entropy Balancing when you have panel or longitudinal data with a treatment that varies over time and time-varying covariates that are affected by past treatment (intermediate confounders). It is well suited when sequential ignorability is defensible and when you want to avoid specifying a parametric propensity model for each period. It works for binary and multi-valued time-varying treatments. Do not use it when treatment is assigned only once; standard entropy balancing suffices. Also avoid it when sequential ignorability is implausible — for instance, when unmeasured time-varying confounders are likely — as no reweighting approach can correct for unobserved confounding.
Strengths & limitations
- Directly balances covariate moments without estimating propensity score models, reducing sensitivity to model misspecification at each period.
- Handles time-varying treatments and confounders rigorously by conditioning on the full treatment history at each step.
- Produces exact balance on all specified moments by construction, unlike propensity-based methods that achieve balance only approximately.
- Transparent diagnostics: standardized mean differences after weighting provide a clear, period-by-period check of covariate balance.
- Flexible enough to balance any user-specified moments, including interactions and higher-order terms of the covariates.
- Relies on the sequential ignorability assumption (no unmeasured time-varying confounders), which cannot be tested from data alone.
- Weights can become extreme when treatment groups differ substantially on covariates or when positivity is weak, inflating variance and reducing effective sample size.
- The number of balance constraints grows with the number of periods and covariates; over-specification can lead to an infeasible optimization problem.
- Software implementations for the dynamic case are less standardized than for static entropy balancing, requiring more careful manual setup.
- Inference typically requires bootstrapping, which increases computational cost substantially for large panels.
Frequently asked
How does dynamic entropy balancing differ from a marginal structural model?
Both handle time-varying treatments by constructing inverse-probability-type weights. Marginal structural models estimate propensity scores parametrically at each period and weight by their inverses. Dynamic entropy balancing replaces propensity score estimation with direct moment-matching optimization, eliminating the need to correctly specify a parametric treatment model at each time period, though both rely on sequential ignorability.
What is sequential ignorability and when is it reasonable?
Sequential ignorability requires that, at each period, treatment assignment is independent of potential outcomes given observed treatment history and measured time-varying covariates. It is reasonable when all important confounders are measured and updated over time. It fails if important unmeasured factors — such as unrecorded health shocks or unobserved firm characteristics — influence both treatment and outcomes.
How do I choose which covariate moments to balance?
Start with means of all time-varying covariates and any lagged outcomes included as controls. Add variance and cross-product constraints for covariates suspected to have non-linear relationships with treatment or outcome. Increasing constraints improves balance but can make weights more extreme and the optimization harder to solve.
Can dynamic entropy balancing handle continuous or multi-valued treatments?
Yes. The moment-balancing framework does not require binary treatment. For multi-valued or continuous treatments, the comparison is between units following different treatment trajectories, and balance constraints can be set up to equate covariate distributions across trajectory groups at each period.
What sample size is needed?
The effective sample size after weighting — not the raw n — determines precision. With many time periods or highly imbalanced groups, extreme weights can shrink the effective sample substantially. Each treatment-history group should have at least 30 to 50 effective observations after weighting to support reliable inference.
Sources
- Hainmueller, J. (2012). Entropy Balancing for Causal Effects: A Multivariate Reweighting Method to Produce Balanced Samples in Observational Studies. Political Analysis, 20(1), 25-46. DOI: 10.1093/pan/mpr025 ↗
- Blackwell, M., & Glynn, A. N. (2018). How to Make Causal Inferences with Time-Series Cross-Sectional Data under Selection on Observables. American Political Science Review, 112(4), 1067-1082. DOI: 10.1017/S0003055418000357 ↗
How to cite this page
ScholarGate. (2026, June 3). Dynamic Entropy Balancing for Longitudinal Causal Inference. ScholarGate. https://scholargate.app/en/causal-inference/dynamic-entropy-balancing
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.
- Dynamic Inverse Probability WeightingCausal inference↔ compare
- Dynamic Propensity Score MatchingCausal inference↔ compare
- Entropy BalancingCausal inference↔ compare
- Inverse Probability WeightingCausal inference↔ compare
- Marginal Structural ModelCausal inference↔ compare
- Propensity Score WeightingCausal inference↔ compare