This project studies how to perform reliable statistical inference when data access is constrained by privacy requirements. The central challenge is that privacy-preserving mechanisms deliberately perturb the data or the inferential procedure, while scientific users still need valid uncertainty quantification, hypothesis tests, confidence sets, and predictive guarantees.
The work connects robust statistics, asymptotic theory, and modern privacy frameworks. Early work on perturbed M-estimation investigates how robust estimators behave under privacy-motivated perturbations. More recent work studies differentially private inference for categorical data, equivalence testing under privacy constraints, conformal prediction with private quantile search, and Bayesian approximations for privatized data.
The broader aim is to build tools that make privacy constraints part of the statistical model rather than an afterthought. This includes understanding the inferential cost of privacy, designing procedures that retain useful operating characteristics after privatization, and developing methods that are suitable for sensitive data settings in public health, social science, and administrative data.
Selected related publications include:
- Perturbed M-Estimation: A Further Investigation of Robust Statistics for Differential Privacy
- Differentially Private Conformal Prediction via Quantile Binary Search
- Fiducial Matching: Differentially Private Inference for Categorical Data
- Equivalence Testing Under Privacy Constraints
- Large-Sample Bayesian Approximations for Privatized Data
- Data of the Defense and the Defense of Data
