Causal Inference With Corrupted Data: Measurement Error, Missing Values, Discretization, And...
summary
Census releases add differential privacy noise and plan wage discretization (rounding to 50¢), which simulations show can shrink coverage from 95% down to roughly 70% and distort trend estimates. The project presents an end-to-end pipeline that combines matrix-completion–style cleaning, sample splitting to preserve row independence, and inference that claims root-n consistency, Gaussian approximation and semi-parametric efficiency under an approximate low-rank covariance assumption.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 0:20Motivation
- 0:44Differential Privacy
- 1:41The 2010 Census
- 2:30Discretization
- 3:52Standard Practice
- 4:36Solution
- 5:16Challenges
- 6:38Project Summary
related talks
01h 03mStatistical Inference and Privacy, Part II
01h 07mStatistical Inference and Privacy, Part I
44mKeep it Clean: Why Bad Data Ruins Projects and How to Fix it
01h 43mPerspective
01h 11mStatistical Limits of Causal Inference
27mYou've got trust issues, we've got solutions: Differential Privacy
57mA New Perspective on High-Dimensional Causal Inference
01h 02mUnmeasured Confounding and More Recent Developments/Challenges in Causal Discovery