Model-assisted inference for dynamic causal effects in staggered rollout cluster randomized experiments

Staggered rollout cluster randomized experiments (SR-CREs) involve sequential treatment adoption across clusters, requiring analysis methods that address a general class of dynamic causal effects, anticipation, and non-ignorable cluster-period sizes. Without imposing any outcome modeling assumptions, we study regression estimators using individual data, cluster-period averages, and scaled cluster-period totals, with and without covariate adjustment from a design-based perspective. We establish consistency and asymptotic normality of each estimator under a randomization-based framework and prove that the associated variance estimators are asymptotically conservative in the Löwner ordering. Furthermore, we conduct a unified efficiency comparison of the estimators and provide recommendations. We highlight the efficiency advantage of using estimators based on scaled cluster-period totals with covariate adjustment over their counterparts using individual-level data and cluster-period averages. Our results rigorously justify linear regression estimators as model-assisted methods to address an entire class of dynamic causal effects in SR-CREs.

💡 Research Summary

This paper addresses the methodological challenges inherent in staggered rollout cluster randomized experiments (SR‑CREs), where clusters adopt a treatment at different calendar periods. Traditional analyses of such designs often assume a single level of observation, no anticipation of treatment, static treatment effects, and ignore the potentially non‑ignorable variation in cluster‑period sizes. The authors propose a comprehensive framework that simultaneously accommodates anticipation, dynamic causal effects, and heterogeneous cluster‑period sizes.

First, they formalize a finite‑population causal model. Under a cluster‑level SUTVA, each individual’s potential outcome depends only on its own cluster’s adoption time, ruling out between‑cluster interference. Treatment adoption times are randomized according to pre‑specified allocation counts, providing a pure randomization‑based source of inference.

The central causal quantity is the Dynamic Weighted Average Treatment Effect (DWATE), denoted τ_j(a,a′), which measures the weighted difference in potential outcomes at calendar period j between clusters that adopt the treatment at times a and a′. Weights can be defined at the individual level (w_ijk) or at the cluster‑period level (w_ij·), allowing the estimand to represent either an individual‑average effect or a cluster‑average effect. The authors show three mathematically equivalent representations of τ_j(a,a′): (1) a weighted sum over individual outcomes, (2) a weighted sum over cluster‑period averages, and (3) a simple average of scaled cluster‑period totals. This equivalence motivates three families of regression estimators.

The three estimators are:

1. Individual‑level regression – a linear regression of individual outcomes on treatment‑adoption indicators (and optional covariates).
2. Cluster‑period‑average regression – a regression using the mean outcome within each cluster‑period as the dependent variable.
3. Scaled cluster‑period‑total regression – a regression on the total outcome in each cluster‑period after scaling by the cluster‑period weight.

All estimators are “model‑assisted”: they remain consistent for the DWATE even if the linear model is misspecified, because consistency is derived from the randomization design rather than model correctness.

The asymptotic analysis assumes the number of clusters I → ∞ while the number of rollout periods J remains fixed, and cluster‑period sizes N_ij are bounded. Under mild balance conditions on the cluster weights (no single cluster dominates), each estimator is shown to be √I‑consistent and asymptotically normal.

For variance estimation, the authors adopt two design‑based robust methods: the Liang‑Zeger cluster‑robust (CR) variance estimator for the individual‑level regression, and the White‑Huber heteroskedasticity‑consistent (HC) estimator for the cluster‑average and total regressions. They prove that both estimators are asymptotically conservative in the Löwner ordering, meaning the estimated covariance matrix dominates the true asymptotic covariance matrix. This guarantees that confidence intervals constructed from these variance estimates will have at least nominal coverage.

A unified efficiency comparison reveals that, when covariates are included, the scaled cluster‑period‑total regression attains the smallest asymptotic variance among the three families. The efficiency gain stems from two sources: (i) collapsing individual observations into a single weighted total per cluster‑period eliminates within‑cluster sampling variability, and (ii) covariate adjustment removes additional noise at the cluster‑period level. Consequently, the total‑based estimator dominates the average‑based estimator, which in turn dominates the individual‑level estimator, provided the weighting scheme satisfies the balance assumptions.

Simulation studies explore a range of realistic scenarios: varying degrees of cluster‑period size heterogeneity, presence or absence of covariate effects, and different magnitudes of anticipation. Results confirm the theoretical findings: the HC variance estimator is indeed conservative, and the total‑based estimator consistently yields lower mean‑squared error and tighter confidence intervals than the alternatives.

An empirical illustration uses a stepped‑wedge public‑health intervention where schools adopt a nutrition program at different times. The authors estimate several DWATEs, including (a) the anticipation effect of delaying adoption, (b) the duration effect of longer exposure, and (c) the overall average treatment effect across all adoption times. Covariate‑adjusted total‑based regressions produce the most precise estimates, highlighting the practical advantage of the proposed method.

In conclusion, the paper provides a rigorous, design‑based justification for using linear regression as a model‑assisted tool in SR‑CREs. By explicitly incorporating anticipation, dynamic effects, and non‑ignorable cluster‑period sizes, and by establishing consistency, asymptotic normality, and conservative variance estimation, the authors deliver a complete methodological toolkit. The efficiency hierarchy they uncover offers clear guidance: practitioners should prefer covariate‑adjusted regressions on scaled cluster‑period totals whenever the data structure permits. This work substantially advances causal inference for staggered rollout experiments across fields such as public policy, medicine, and education.