Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Design-Robust Event-Study Estimation under Staggered Adoption Diagnostics, Sensitivity, and Orthogonalisation

Craig S Wright

TL;DR

This paper develops a design-first framework for event-study and DiD estimands under staggered adoption with heterogeneous effects. It provides exact probability limits for TWFE-based event studies, introduces computable design diagnostics for negative weights and horizon contamination, and offers sensitivity-robust inference under restricted violations of parallel trends using a calibrated deviation class. The approach emphasizes convex, transparent aggregation across cohorts and event horizons, and it integrates covariate orthogonalisation via Riesz representers to enable high-dimensional nuisance learning with valid inference. Monte Carlo designs and a replicable Bank Deregulation Shock application illustrate the practical performance of robust estimators, diagnostics, and sensitivity reporting. Overall, the method improves interpretability, reliability, and robustness of dynamic treatment effect estimation in settings where timing variation and heterogeneity are intrinsic to policy and market dynamics.

Abstract

This paper develops a design-first econometric framework for event-study and difference-in-differences estimands under staggered adoption with heterogeneous effects, emphasising (i) exact probability limits for conventional two-way fixed effects event-study regressions, (ii) computable design diagnostics that quantify contamination and negative-weight risk, and (iii) sensitivity-robust inference that remains uniformly valid under restricted violations of parallel trends. The approach is accompanied by orthogonal score constructions that reduce bias from high-dimensional nuisance estimation when conditioning on covariates. Theoretical results and Monte Carlo experiments jointly deliver a self-contained methodology paper suitable for finance and econometrics applications where timing variation is intrinsic to policy, regulation, and market-structure changes.

Design-Robust Event-Study Estimation under Staggered Adoption Diagnostics, Sensitivity, and Orthogonalisation

TL;DR

This paper develops a design-first framework for event-study and DiD estimands under staggered adoption with heterogeneous effects. It provides exact probability limits for TWFE-based event studies, introduces computable design diagnostics for negative weights and horizon contamination, and offers sensitivity-robust inference under restricted violations of parallel trends using a calibrated deviation class. The approach emphasizes convex, transparent aggregation across cohorts and event horizons, and it integrates covariate orthogonalisation via Riesz representers to enable high-dimensional nuisance learning with valid inference. Monte Carlo designs and a replicable Bank Deregulation Shock application illustrate the practical performance of robust estimators, diagnostics, and sensitivity reporting. Overall, the method improves interpretability, reliability, and robustness of dynamic treatment effect estimation in settings where timing variation and heterogeneity are intrinsic to policy and market dynamics.

Abstract

This paper develops a design-first econometric framework for event-study and difference-in-differences estimands under staggered adoption with heterogeneous effects, emphasising (i) exact probability limits for conventional two-way fixed effects event-study regressions, (ii) computable design diagnostics that quantify contamination and negative-weight risk, and (iii) sensitivity-robust inference that remains uniformly valid under restricted violations of parallel trends. The approach is accompanied by orthogonal score constructions that reduce bias from high-dimensional nuisance estimation when conditioning on covariates. Theoretical results and Monte Carlo experiments jointly deliver a self-contained methodology paper suitable for finance and econometrics applications where timing variation is intrinsic to policy, regulation, and market-structure changes.
Paper Structure (123 sections, 2 theorems, 193 equations, 21 figures, 12 tables)

This paper contains 123 sections, 2 theorems, 193 equations, 21 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Framework and estimands
  3. Panel structure, timing, and the design measure space
  4. Potential outcomes, dynamic effects, and economically interpretable aggregation
  5. Identifying assumptions, operators, and sensitivity-compatible primitives
  6. Conventional TWFE event-study regressions
  7. Specification
  8. Probability limit and contamination decomposition
  9. A computable representation
  10. Heterogeneity-robust estimation and transparent aggregation
  11. Cohort-time and group-time estimators
  12. Event-time aggregation
  13. Comparison with imputation estimators
  14. Design diagnostics
  15. Negative-weight risk index
  16. ...and 108 more sections

Key Result

Proposition 1

Maintain absorbing adoption in eq:absorbing, Assumptions ass:noanticip and ass:pt, and the TWFE event-study specification eq:twfe-es with event window $\mathcal{K}$ and baseline $k_0$. Suppose there exist at least two cohorts $g_1\neq g_2$ with $\mathbb{P}(G_i=g_j)>0$ and at least two calendar times then there exists at least one $k\in\mathcal{K}\setminus\{k_0\}$ such that the TWFE probability lim

Figures (21)

  • Figure 1: Cohort--time geometry under staggered adoption. The adoption schedule induces cohort--time cells $(g,t)$ and associated event times $k=t-g$. Event-time indicators select diagonals in this diagram. After residualising by unit and time fixed effects, the resulting regressors typically change sign and are not mutually orthogonal, which is the design mechanism behind negative weights and cross-horizon contamination in TWFE event studies.
  • Figure 2: Sensitivity region in $(B,\Gamma)$ space. Shaded set indicates parameter pairs for which the target conclusion remains unchanged under restriction class $\mathcal{R}$.
  • Figure 3: Cohort timing used in every Monte Carlo run: $T=12$ and $A\in\{4,6,8,10,\infty\}$.
  • Figure 4: Violation grid for $(B,\Gamma)$ shown separately for each $\Delta(\mathcal{R})\in\{0,0.25,0.50,1.00\}$.
  • Figure 5: Monte Carlo execution loop used for each of the $3\times 64$ design cells.
  • ...and 16 more figures

Theorems & Definitions (6)

  • Definition 1: Negative-weight mass and cross-horizon contamination
  • Proposition 1: Design-driven contamination
  • proof : Proof sketch
  • Definition 2: Negative-weight mass
  • Definition 3: Cross-horizon contamination
  • Theorem 1: Uniformly valid confidence regions