Research map / Federated learning and privacy / Differential privacy

Differentially private learners for heterogeneous treatment effects

ICLR 2025 · Federated learning and privacy › Differential privacy · approach: Private data analysis frameworks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DP-CATE, an output-perturbation framework that makes Neyman-orthogonal (R-learner/DR-learner) CATE meta-learners differentially private, for finitely many CATE queries and for releasing the full CATE function via a Gaussian process calibrated in an RKHS.

Problem

Estimating conditional average treatment effects (CATE) from sensitive observational patient data while guaranteeing differential privacy, without violating causal assumptions or biasing estimates.

Approach

Calibrates Gaussian output-perturbation noise using the gross-error sensitivity (norm of the second-stage influence function) of Neyman-orthogonal two-stage meta-learners, and for functional release adds a calibrated Gaussian process noise to an RKHS (Gaussian kernel) CATE estimator.

Key results

• On synthetic Dataset 2, DP-CATE PEHE decreases with increasing privacy budget and converges toward the non-private learner's PEHE.
• On MIMIC-III (d=1312 queries) and TCGA (d=2659 queries), the precision loss from DP relative to the non-private learner is described as fairly small.
• Excess empirical risk bounds of O(n^-2 ln n) for finite queries and O(n^-2) for functional queries, both o(n^-1).

Measured effects reported in the paper

Surprising findings

• The DR-learner needs larger calibration noise because the influence function includes 1/(pi(1-pi)), so the authors recommend the R-learner under DP.

Limitations

• No suitable baseline exists, so evaluation only shows how error changes with privacy budget rather than comparing to competing DP CATE methods.
• Functional release requires Gaussian kernel RKHS second stage and convex Lipschitz loss; iterative querying cost grows with number of queries.
• Uncertainty quantification and fairness under privacy are left unexplored.

Experimental setup

Datasets: MIMIC-III TCGA synthetic CATE datasets

Benchmarks: synthetic CATE datasets (Dataset 1, Dataset 2) MIMIC-III TCGA

Metrics: PEHE

Baselines: none (no suitable DP CATE baseline; non-private learner used as reference)

Builds on: R-learner DR-learner smooth sensitivity output perturbation DP-SGD

Closest papers by meaning

• Privacy Preserving Adaptive Experiment Design (ICML 2024 · similarity 0.88)
• Meta-Learners for Partially-Identified Treatment Effects Across Multiple Environments (ICML 2024 · similarity 0.87)
• Estimating Heterogeneous Treatment Effects by Combining Weak Instruments and Observational Data (NeurIPS 2024 · similarity 0.86)
• Learning Representations of Instruments for Partial Identification of Treatment Effects (ICML 2025 · similarity 0.86)
• A Non-parametric Direct Learning Approach to Heterogeneous Treatment Effect Estimation under Unmeasured Confounding (NeurIPS 2024 · similarity 0.86)
• Quantifying Aleatoric Uncertainty of the Treatment Effect: A Novel Orthogonal Learner (NeurIPS 2024 · similarity 0.85)