My research spans statistical inference after data transformations, privacy-preserving synthetic data, differential privacy, machine learning methods for social science, and election forecasting. Publications appear in venues including ICLR, ACM PODS, PNAS, the Harvard Data Science Review, Political Analysis, and Political Science Research and Methods.
Working Papers#
- Learning from Small Data: Tabular Foundation Models for Political Science. Working paper, PDF. Software: tabfound.
- Diversity Profiles for Party Systems. Under review. Software: partyscape.
- When Privacy Measures Fail: A Theoretical Critique of Empirical Disclosure Risk Assessment for Synthetic Data. With Jonathan Latner and Jörg Drechsler. Under review.
- Learning by Solving: Building Government Data Literacy One Problem at a Time. With Malte Schierholz, Julia von Bartenwerffer, Jacob Beck, Felix Henninger, Leonhard Kestel, Michael König, Heidi Seibold, Wiebke Weber, and Frauke Kreuter. Under review.
Published#
Evaluates TabPFN as the prediction step in Supreme Court forecasting, double machine learning and multiple imputation, with nothing to tune.
Benchmarks that find GANs uncompetitive for tabular synthetic data often compare a default implementation, not the method.
Three ways to build prior knowledge about the data into GANs for private tabular synthetic data, with differential privacy.
Introduces GANs to R users and presents the RGAN package, which runs on torch and makes GAN design choices easy to vary.
Examining what privacy guarantees synthetic data can satisfy even without formal guarantees during synthesizer training.
Perceived threat raises approval of the incumbent leader during a crisis, while anxiety lowers it. Evidence from German panel data.
Examining researcher variability in computational reproduction across 85 independent teams.
Guidelines for deriving meaningful interpretations from regression results with log-transformed dependent variables.
Ready-made synthetic data generators are not plug-and-play: preprocessing matters as much as tuning, and so does knowing the generator.
Algorithms for continually releasing differentially private synthetic data from longitudinal data collections.