Zubizarreta, JoséShen, Zhu2026-06-0920262026-05-142026Shen, Zhu. 2026. When, Where, and Who? New Methods for Causal Inference with Panel, Geographic, and Multisite Data. Doctoral Dissertation, Harvard University Graduate School of Arts and Sciences.32701806https://dash.harvard.edu/handle/1/42740352This dissertation develops new methods for causal inference with panel, geographic, and multisite data. It studies three related questions: how to interpret and estimate causal effects in panel settings with staggered interventions, how to decompose geographic differences in outcomes into place and population components, and who should be targeted under treatment effect heterogeneity in multisite experiments. Across these settings, the dissertation emphasizes transparent methods for estimation, decomposition, and policy learning that make the underlying causal comparisons and identifying assumptions explicit. The first chapter studies event studies with staggered treatment adoption. It develops an experimental perspective on event studies, with a focus on information borrowing from modeling assumptions. It proposes robust weighting estimators that increasingly use more information across units and time periods, justified by increasingly stronger assumptions on the treatment assignment and potential outcomes mechanisms. It also provides a novel closed-form decomposition of the classical dynamic two-way fixed effects (TWFE) regression estimator, as well as a class of recently proposed estimators for event studies, revealing in finite samples the hypothetical experiment that these estimators approximate. In addition, the chapter develops diagnostics for event studies, including covariate balance, sign reversal, effective sample size, and the contribution of each observation to the analysis. The second chapter studies geographic differences in health care utilization. Using linked Medicare claims for beneficiaries who move between Hospital Referral Regions (HRRs), it develops a causal decomposition of regional utilization differences into place and population components. A large and growing literature in health care, labor economics, and public finance uses mover designs to separate place effects from population sorting. These designs typically rely on fixed effects regressions that impose strong homogeneity assumptions about post-move dynamics across origins, destinations, and time since move, and are difficult to interpret causally when these assumptions fail. The chapter instead identifies path-specific mover effects, indexed by origin, destination, and time since move, and aggregates them into counterfactuals that place the same population under different regional environments. It derives identifying assumptions, proposes an estimator, and develops inference that accounts for the aggregation step. The third chapter studies policy learning in multisite randomized experiments under budget constraints. Motivated by the Baby's First Years (BFY) study, it develops an interpretable policy learning framework that addresses three central challenges: limited treatment capacity, cross-site heterogeneity and deployment uncertainty, and the need for decision rules that are suitable for implementation by program administrators. The chapter establishes finite-sample regret guarantees for the resulting policies under a general deployment population. Applying the framework to BFY, it uncovers substantial treatment effect heterogeneity that conventional average-effect analyses miss: targeted allocation improves three of seven age-3 and age-4 developmental outcomes by 13-22 percent relative to randomized allocation under the same budget. These findings shift the policy question from whether unconditional cash transfers improve child development on average to how limited transfer budgets can be allocated to maximize developmental gains. Taken together, these chapters develop new methods for causal inference that clarify when treatment effects emerge, where differences arise, and who stands to benefit, making causal heterogeneity more interpretable and actionable for policy evaluation and decision-making.application/pdfencausal inferencehealth policyheterogeneous treatment effectsmover designspanel datapolicy learningBiostatisticsThesis or Dissertation2026-06-090000-0003-4564-5438