Identification and Estimation of Staggered Difference-in-Differences with Network Spillovers

Abstract: This paper develops a difference-in-differences framework for staggered policy adoption when units can be affected by other units' adoption. For each treated cohort and event time, the framework separates the effect of own adoption, the spillover effect generated by other adopters, and the total effect under the realized rollout. Identification uses a prespecified summary of spillover exposure and parallel trends comparisons among units with the same exposure at the baseline and target dates. Spillover effects are learned from never-treated units and evaluated for treated cohorts under the exposure distribution they face. We construct estimators for these effects and an inference procedure that allows for spatial dependence. Monte Carlo simulations illustrate that standard DID estimators that ignore spillovers can miss the total effect, whereas the proposed estimators have small bias for these effects and the associated confidence intervals have coverage close to the nominal level. In an empirical study of the Community Health Centers rollout, estimated spillovers account for a substantial share of the effect on older-adult mortality.

• The paper introduces three causal effects (DSE, CSE, DTE) using explicit exposure mapping to handle network spillovers in staggered policy adoption.
• It employs saturated long-difference comparisons and spillover transport methods to overcome biases typical in standard DID estimates.
• Empirical results, including Monte Carlo simulations, reveal that ignoring spillovers underestimates policy impacts, stressing the need for interference-aware methodologies.

Identification and Estimation in Staggered Difference-in-Differences with Network Spillovers

Overview

This paper ("Identification and Estimation of Staggered Difference-in-Differences with Network Spillovers" (2605.15119)) develops a difference-in-differences (DID) econometric framework for scenarios with staggered policy adoption and network-dependent spillovers. The standard DID assumption of no interference—where each unit's outcome depends only on its own treatment—is replaced with a more general modeling of spillover exposures that can result from neighboring units' treatment status. The paper formalizes identification, estimation, and inference for three distinct causal estimands within this context: own-adoption switching effect (DSE), spillover effect absent own adoption (CSE), and the total effect of realized rollout (DTE).

Conceptual Framework

Exposure Mapping and Potential Outcomes

The setup allows unit $i$'s outcome at time $t$ to depend on a vector of adoption times $G$ for all units, accommodating arbitrary network spillovers and treatment timing. Spillover exposure is mapped via a prespecified function (e.g., geographic or network-based weights, temporal kernel), yielding a discrete exposure state $H_{it}$ used for causal modeling. The exposure mapping is central to both identification and empirical implementation, as it reduces the full network structure to interpretable exposure states.

Causal Effects

• Dynamic Switching Effect (DSE): Captures the effect of switching one's own adoption status, holding exposure from neighbors constant at the realized state.
• Control-State Spillover Effect (CSE): Measures the effect of exposure from others when own adoption status is left untreated, averaged over the exposure faced by the cohort.
• Dynamic Total Effect (DTE): The sum of DSE and CSE, representing the aggregate effect of the realized staggered rollout, i.e., moving from pure control to the empirical treatment-exposure regime.

This decomposition is nontrivial under interference: standard DID contrasts conflate direct, indirect (spillover), and baseline contamination effects.

Identification Strategy

The identification of DSE uses parallel trends among never-treated units with matched baseline and target exposure states, recovering the counterfactual untreated trend for treated cohorts. CSE identification relies on within-never-treated comparisons across exposure states, combined with assumptions that allow transportability of these spillover estimates to the treated cohort's exposure distribution.

The paper formalizes conditions for identification:

• Parallel Trends (conditional on exposure states)
• No-Anticipation
• Cross-Cohort Transportability

Notably, the pure direct effect under zero exposure is generally not point-identified without additional assumptions (e.g., no interaction between treatment and exposure).

Estimation and Inference

Estimators

• DSE is estimated via saturated long-difference comparisons within exposure-covariate strata.
• CSE is estimated by fitting spillover responses in the never-treated source sample, transported to treated cohorts' exposure distributions.
• DTE is the sum of DSE and CSE on the joint admissible support.

A crucial aspect is support overlap—retained cells must have sufficient representation in both treated and comparison groups for valid estimation. The approach is implementable via closed-form formulas and saturated regressions.

Inference

Inference leverages stacked estimating equations, yielding joint influence rows for DSE, CSE, and DTE. Spatial heteroskedasticity- and autocorrelation-consistent (HAC) covariance estimators are applied, permitting spatial/network dependence in outcomes. Asymptotics are based on conditional finite-population sequences, ensuring valid inference under realistic spatial or network dependence structures.

Numerical Results

Monte Carlo simulations probe estimator properties under various data-generating processes, including additive and interaction designs. Standard DID estimators that ignore spillovers exhibit substantial bias and low coverage relative to the DTE target. The proposed estimators demonstrate negligible bias and nominal coverage (close to 95%), confirming robustness under exposure-induced interference.

Event-time aggregates are carefully constructed using common support to avoid extrapolation beyond the empirical data's representation.

Empirical Application: Community Health Centers

The empirical study revisits Bailey and Goodman-Bacon (2015)'s analysis of CHC rollout and county-level mortality, using spatial exposure within 50 miles as the mapping. The decomposition reveals that standard DID estimators understate the mortality reduction—estimated DTE exceeds standard DID by more than 50% at several event times. Spillover effects (CSE) contribute approximately 40% of the estimated total effect, indicating that indirect benefits to nearby counties are significant. Never-treated counties are themselves often exposed, further suggesting contamination in standard comparison groups.

Implications and Extensions

The methodological contributions provide a rigorous statistical toolkit for staggered DID designs with interference, relevant to empirical settings where policy diffusion, peer effects, and spatial spillovers are substantial. Practically, accurate policy evaluation in these contexts demands explicit modeling of exposure and spillovers; otherwise, causal estimates of policy efficacy may be seriously misestimated.

Future directions include:

• Estimating (rather than prespecifying) the exposure mapping: This could enable more data-driven characterizations of network effects.
• Diagnostics and sensitivity for baseline contamination: In staggered designs, pre-treatment exposure may ensue for later-adopting cohorts, requiring careful attention.
• Modeling endogenous timing and exposure: Treatment adoption may itself be network-dependent, introducing additional complexities in the causal analysis.

Conclusion

This research systematically addresses identification, estimation, and inference for DID designs with staggered adoption and network spillovers, clarifying the causal structure and proposing estimators with proven robustness to interference. It demonstrates that spillover-aware estimands materially differ from those obtainable by conventional DID, both theoretically and empirically. The approach is extendable to varied settings featuring network effects, and its adoption will yield more credible and interpretable causal inferences in the presence of interference.