Statistical Inference and Privacy, Part I
summary
Differential privacy adds explicit noise (Laplace around a truncated mean) so standard estimators give biased, anti-conservative p‑values and confidence intervals even at large n; variance from the mechanism grows with the data range B so small B induces truncation bias and large B inflates intervals (examples show intervals becoming an order of magnitude wider). Remedies model the mechanism as a nuisance — using a marginal likelihood or measurement‑error surrogates and MCMC/variational/bootstrap approximations — yielding nearly optimal DP confidence intervals under specified loss.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 0:35Outline
- 1:47Feinberg Problem
- 6:10Data Privacy
- 9:27Physical Inference
- 12:05Differential Privacy
- 15:25Differential Privacy Interpretation
- 29:11Sampling Distributions
- 32:19Adding Noise
- 40:26Variance of Noise
- 43:31Interval Estimation
- 47:15Marginal likelihood
- 47:55surrogate
- 48:30sample sizes
- 51:52loss function
- 54:22differentially private
- 55:32design
- 58:41data quality
related talks
01h 00mPANEL: Statistical Theory, Privacy and Data Analysis
47mFinite Sample Differentially Private Confidence Intervals
47mDifferential Privacy from a Statistical Perspective–Obtaining Valid Inferences...
01h 14mTutorial: Privacy, Part II
48mRobust Statistics and Bias Correction: Ideas for Differential Privacy and Utility
31mWhen Existing Techniques in Linear Regression Preserve Differential Privacy
01h 00mDifferential Privacy: Fundamentals to Forefront I
38mDifferentially Private Inference for Binomial Data