The Department of Biomedical Data Science invites you to attend a seminar in person on Thursday, November 12 from 12:00-1:00pm at DHMC, Auditorium H with José Zubizarreta, PhD, Daniel C. Tosteson Professor of Health Care Policy at Harvard Medical School & Professor of Biostatistics at Harvard T.H. Chan School of Public Health. In observational studies, a fitted model may yield an estimate of a treatment effect without making explicit the comparison on which that estimate rests. Which observations are compared, and how much does each contribute? Which features of a randomized experiment, if any, does the comparison reproduce? These questions arise in spatial regressions with Gaussian process errors, two-way fixed effects analyses of staggered policy adoption, and mixed effects meta-analyses of trials. In this lecture, we derive exact, closed-form weighting representations of these estimators. In each case, the weights solve a common optimization problem that minimizes their dispersion subject to balance requirements. These finite-sample identities make explicit the comparisons implicit in the models, providing a basis for examining their causal interpretation and reconsidering their model specification. (i) For spatial data, specifying the error structure of a linear model is equivalent to including a set of hidden covariates under a particular form of regularization. (ii) For panel data, the weights reveal how two-way fixed effects and related estimators borrow information across units and time periods, clarifying the assumptions about the dynamics of treatment effects needed to justify that borrowing. (iii) For meta-analytic data, the weights identify studies and observations that are poorly aligned with the covariate distribution of the target population, helping to assess when their exclusion is prudent. These representations give rise to diagnostics that open the analysis to scrutiny, and they bring methods that appear unrelated onto common ground.
"The Experiment Inside the Model: Exact Weighting Representations for
Spatial, Panel, and Meta-Analytic Studies”
Thursday, November 12 from 12:00-1:00pm ET
José R. Zubizarreta, PhD
Daniel C. Tosteson
Professor of Health Care Policy at
Harvard Medical School &
Professor of Biostatistics at
Harvard T.H. Chan School of Public Health
The presentation will take place both in person at DHMC, Auditorium H (and via Zoom).
Please write to biomedical.data.science@dartmouth.edu to request the Zoom login details.
José R. Zubizarreta, PhD, is the Daniel C. Tosteson Professor of Health Care Policy at Harvard Medical School, Professor of Biostatistics at the Harvard T.H. Chan School of Public Health, and a Faculty Affiliate in the Department of Statistics at Harvard University. His research focuses on developing statistical methods for causal inference to advance health care and public policy. Dr. Zubizarreta earned his PhD in Statistics from the Wharton School of the University of Pennsylvania, an MA in Economics from Universidad de Chile, and an undergraduate degree in Engineering from Pontificia Universidad Católica de Chile. A former Fulbright Scholar, Dr. Zubizarreta has received research support from the Alfred P. Sloan Foundation, the Warren Alpert Foundation, and other major research organizations. He serves as an Associate Editor for several leading journals, including the Annals of Applied Statistics, the Harvard Data Science Review, and the Journal of the American Statistical Association. Dr. Zubizarreta is an elected Fellow of the American Statistical Association (ASA). His honors include the Kenneth Rothman, William Cochran, and Tom Ten Have Memorial Awards, as well as the ASA’s Mid-Career Excellence Award in Health Policy Statistics. Dr. Zubizarreta has also received the Myrto Lefkopoulou Distinguished Lectureship at Harvard and has been selected to deliver the Wayne A. Fuller Distinguished Lecture at Iowa State University.