Multi-period Doubly Robust Estimation
Multi-period Doubly Robust Causal Effect Estimator · Also known as: longitudinal DR estimation, multi-period DR, multi-wave doubly robust, sequential doubly robust estimation
Multi-period doubly robust (DR) estimation extends the classic doubly robust approach to longitudinal settings with multiple treatment periods and time points. It combines an outcome regression model and a propensity score model for each period, retaining consistency of the causal effect estimate as long as at least one of the two models is correctly specified at every time point.
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 doubly robust estimation when you have longitudinal or panel data with multiple time periods, units receiving treatment at different times (staggered adoption), and a rich set of time-varying covariates that may confound treatment timing. It is well-suited when you cannot confidently specify either the selection process or the outcome trajectory but can specify at least one of them reasonably. Avoid it when you have very few time periods (a standard two-period DiD may suffice), when sample sizes per period are very small (fewer than 50 units per group), or when treatment is perfectly predicted by observed covariates (positivity violation).
Strengths & limitations
- Consistent whenever either the propensity score model or the outcome regression is correctly specified at each period, providing a substantial robustness advantage over methods relying on a single model.
- Accommodates staggered treatment adoption and heterogeneous treatment timing naturally, computing group-time specific causal effects.
- Semiparametrically efficient under correct specification of both models, achieving the lowest possible asymptotic variance in that class.
- Flexible: modern implementations allow machine-learning estimators such as random forests or LASSO to fit the nuisance models, enabling high-dimensional covariate adjustment.
- Naturally extends to aggregating heterogeneous effects across cohorts and time horizons using weighted averages.
- Requires correct specification of at least one nuisance model at every time point; if both are misspecified even slightly across several periods, bias can accumulate.
- Computationally more demanding than single-period DR or standard DiD, particularly with many periods and machine-learning nuisance models requiring cross-fitting.
- Inference is complex: naive standard errors that ignore nuisance-model estimation uncertainty are invalid; influence-function or bootstrap-based variance estimation is required.
- The positivity assumption must hold at every period; near-violations cause propensity weights to explode and inflate variance dramatically.
- Interpretation of aggregated effect estimates requires care when treatment timing is highly heterogeneous across units.
Frequently asked
How does multi-period DR differ from a standard doubly robust estimator?
The standard DR estimator handles a single treatment period and a single post-outcome. The multi-period extension applies the double-robustness logic sequentially across all time points, accounting for time-varying covariates and staggered treatment adoption that a one-shot DR estimator cannot address.
Does the double-robustness guarantee still hold with machine-learning nuisance models?
Yes, provided cross-fitting (sample splitting) is used: estimate the nuisance models on a held-out fold and evaluate predictions on the complementary fold. Without cross-fitting, overfitting in flexible models invalidates the double-robustness property.
What is the positivity assumption and why does it matter?
Positivity requires that every unit has a strictly positive probability of being treated or untreated at each period given its covariates. Violations cause propensity weights to explode, making estimates highly unstable. Check empirically by inspecting the distribution of estimated propensity scores for near-zero or near-one values.
How do I aggregate group-time treatment effects into a single summary?
Common aggregations weight group-time effects by cohort size to produce an overall ATT, or by exposure length to highlight dynamic effects. The csdid and did R packages implement these aggregations directly from Callaway and Sant'Anna's framework.
When should I prefer multi-period DR over a simpler stacked DiD?
Choose multi-period DR when you have rich time-varying confounders you want to adjust for beyond fixed effects, when treatment timing is not as-good-as-random, or when you need the robustness guarantee that one misspecified model does not invalidate the whole analysis.
Sources
- Bang, H., & Robins, J. M. (2005). Doubly robust estimation in missing data and causal inference models. Biometrics, 61(4), 962-973. DOI: 10.1111/j.1541-0420.2005.00377.x ↗
- Callaway, B., & Sant'Anna, P. H. C. (2021). Difference-in-differences with multiple time periods. Journal of Econometrics, 225(2), 200-230. DOI: 10.1016/j.jeconom.2020.12.001 ↗
How to cite this page
ScholarGate. (2026, June 3). Multi-period Doubly Robust Causal Effect Estimator. ScholarGate. https://scholargate.app/en/causal-inference/multi-period-doubly-robust-estimation
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.
- Difference-in-DifferencesEconometrics↔ compare
- Doubly Robust EstimationCausal inference↔ compare
- Dynamic Difference-in-DifferencesCausal inference↔ compare
- Inverse Probability WeightingCausal inference↔ compare
- Marginal Structural ModelCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare