Panel Discussion
summary
Panel covers three practical research threads: a classroom RCT that implements local differential privacy and analyzes how participant cheating breaks identification, including IRB tradeoffs and attempts to weight datasets with different epsilons; a CP‑DAG program that treats known bucketed graphs as inputs, gives asymptotic-efficient estimators and reports simulation checks when graph learning is noisy; and empirical/validation notes on synthetic‑control assumptions (convex‑span vs linear‑span) and cross‑validation to check reconstruction accuracy.
AI-written summary · about
chapters (18)
- 0:00Intervention and independence
- 4:00Data fusion and sensitivity
- 5:36Defining participant behavior
- 7:53Experimental design and scope
- 9:32Incorporating domain knowledge
- 11:37Model accuracy and privacy
- 13:26Simulations and study design
- 15:44Covariate adjustment
- 17:56Practical trial constraints
- 19:15Rationale for clinical trials
- 20:37Contextualizing research goals
- 22:41Statistical validity
- 24:20Graph learning approaches
- 27:18Placebo tests for validation
- 28:12Methodological heuristics
- 29:56Clinical trial value
- 30:36Future directions
- 32:08Closing remarks
related talks
01h 00mPANEL: Statistical Theory, Privacy and Data Analysis
58mLarge-Scale Private Learning, Part III
26mAaron Roth, Associate Professor, University of Pennsylvania, NYC 2017
43mPanel: Privacy Design Patterns for AI Systems: Threats and Protections
10mCommunicating Differential Privacy Guarantees to Data Subjects
09mIntegrating Differential Privacy and Contextual Integrity
19mPractical Considerations for Differential Privacy
29mPrivacy-preserving methods: Building secure projects