September 18, 2026 (10:00 am-12:00 pm)
Gross Hall 270
Erin Hartman
Associate Professor of Political Science
University of California, Berkeley
Erin Hartman’s research bridges statistics and the social sciences, developing methods that help researchers answer complex causal questions while strengthening collaboration across disciplines. Her work focuses on causal inference, survey design and analysis, external validity of experiments, falsification testing, and survey weighting.
TALK: Decomposing Changes Across Surveys
When different polls produce different results, how can we tell whether those differences reflect real changes in public opinion or simply differences in the surveys themselves?
In this talk, Hartman will present a formal framework for breaking down differences across surveys into three components: statistical noise, compositional differences, and true public opinion change.
The approach uses an M-estimation framework to simultaneously estimate parameters and variances while accounting for nuisance parameters and correlated estimators. The talk also introduces a sensitivity analysis to evaluate how results may change when they depend on untestable identifying assumptions.
After validating the method through simulations, the framework is applied to four polls from the 2024 U.S. presidential election. The results indicate that summer events, including the incumbent’s debate performance and the subsequent candidate transition, led to a statistically significant shift of undecided voters toward the Democratic nominee.
The research offers a more rigorous way to distinguish genuine shifts in voter sentiment from differences caused by survey composition and statistical variation.
Interested in polling, public opinion, elections, survey research, or quantitative methods?
The Speaker Series is sponsored by Duke’s Social Science Research Institute and held jointly with UNC.
SHORT COURSE: Innovative Methods for Survey Weighting
September 17, 2026 (1:00-5:30 pm)
*includes a 30 minute break*
Gross Hall 330
This workshop will cover the foundations of survey weighting together with several modern, data-driven techniques.
We’ll begin with survey sampling designs and design-based inference, including Horvitz–Thompson and Hájek estimators and the design weights that underlie them. We then examine survey nonresponse, coverage error, and nonprobability samples, along with standard approaches including inverse propensity weighting, post-stratification, linear weighting (GREG), and raking.
The workshop also covers practical considerations such as how weighting affects precision, what to weight on, when to trim, and how to assess balance. The workshop also focuses on advanced weighting techniques, including multilevel calibration weighting and kernel population weighting. These data-driven approaches reduce researcher degrees of freedom and relax functional-form assumptions.
Throughout, participants get hands-on experience through analytical and code-based worksheets in R.
[Register for the Short Course]