Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Differentially Private Learners for Heterogeneous Treatment Effects

Maresa Schröder, Valentyn Melnychuk, Stefan Feuerriegel

TL;DR

The paper tackles privately estimating conditional average treatment effects from observational data by introducing DP-CATE, a Neyman-orthogonal, model-agnostic framework. It develops two output-perturbation schemes: a finite-queries variant that perturbs a vector of CATE estimates using influence-function-based noise calibration, and a functional-queries variant that privately releases the entire CATE function via Gaussian-process noise in an RKHS. The authors establish theoretical guarantees, including quasi-oracle efficiency, under $(\varepsilon,\delta)$-DP, and demonstrate robustness across synthetic and real-world medical datasets, with performance converging to non-private baselines as privacy budgets grow. This framework enables privacy-preserving, subgroup- or function-level CATE analysis in settings like personalized medicine, while remaining compatible with common meta-learners such as the R-learner and DR-learner. Overall, DP-CATE provides a flexible, principled approach to balancing data privacy with actionable individualized treatment effect estimation in observational data.

Abstract

Patient data is widely used to estimate heterogeneous treatment effects and thus understand the effectiveness and safety of drugs. Yet, patient data includes highly sensitive information that must be kept private. In this work, we aim to estimate the conditional average treatment effect (CATE) from observational data under differential privacy. Specifically, we present DP-CATE, a novel framework for CATE estimation that is Neyman-orthogonal and further ensures differential privacy of the estimates. Our framework is highly general: it applies to any two-stage CATE meta-learner with a Neyman-orthogonal loss function, and any machine learning model can be used for nuisance estimation. We further provide an extension of our DP-CATE, where we employ RKHS regression to release the complete CATE function while ensuring differential privacy. We demonstrate our DP-CATE across various experiments using synthetic and real-world datasets. To the best of our knowledge, we are the first to provide a framework for CATE estimation that is Neyman-orthogonal and differentially private.

Differentially Private Learners for Heterogeneous Treatment Effects

TL;DR

The paper tackles privately estimating conditional average treatment effects from observational data by introducing DP-CATE, a Neyman-orthogonal, model-agnostic framework. It develops two output-perturbation schemes: a finite-queries variant that perturbs a vector of CATE estimates using influence-function-based noise calibration, and a functional-queries variant that privately releases the entire CATE function via Gaussian-process noise in an RKHS. The authors establish theoretical guarantees, including quasi-oracle efficiency, under -DP, and demonstrate robustness across synthetic and real-world medical datasets, with performance converging to non-private baselines as privacy budgets grow. This framework enables privacy-preserving, subgroup- or function-level CATE analysis in settings like personalized medicine, while remaining compatible with common meta-learners such as the R-learner and DR-learner. Overall, DP-CATE provides a flexible, principled approach to balancing data privacy with actionable individualized treatment effect estimation in observational data.

Abstract

Patient data is widely used to estimate heterogeneous treatment effects and thus understand the effectiveness and safety of drugs. Yet, patient data includes highly sensitive information that must be kept private. In this work, we aim to estimate the conditional average treatment effect (CATE) from observational data under differential privacy. Specifically, we present DP-CATE, a novel framework for CATE estimation that is Neyman-orthogonal and further ensures differential privacy of the estimates. Our framework is highly general: it applies to any two-stage CATE meta-learner with a Neyman-orthogonal loss function, and any machine learning model can be used for nuisance estimation. We further provide an extension of our DP-CATE, where we employ RKHS regression to release the complete CATE function while ensuring differential privacy. We demonstrate our DP-CATE across various experiments using synthetic and real-world datasets. To the best of our knowledge, we are the first to provide a framework for CATE estimation that is Neyman-orthogonal and differentially private.

Paper Structure

This paper contains 31 sections, 16 theorems, 66 equations, 9 figures, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related work
  3. Problem formulation
  4. Neyman-orthogonal meta-learners for CATE estimation
  5. Differential privacy
  6. Problem statement
  7. Our framework: DP-CATE
  8. DP-CATE for a fixed number of queries
  9. DP-CATE for complete CATE functions
  10. Experiments
  11. Evaluation on synthetic datasets
  12. Evaluation on medical datasets
  13. Discussion
  14. Additional background material
  15. Differential privacy
  16. ...and 16 more sections

Key Result

Theorem 1

Let $z:= (a,x,y)$ define a data sample following the joint distribution $\mathcal{Z}$ and ${\hat{\eta}_\mathrm{DP} = (\hat{\mu}_\mathrm{DP},\hat{\pi}_\mathrm{DP})}$ the nuisance functions estimated on dataset $\tilde{D}$ in a $(\varepsilon/2, \delta/2)$-differentially private manner of choice.As we where $\gamma(T, D)$ is a sample gross-error sensitivity with $T(D)$ instead of $T(P)$ in Eq. (eq:g

Figures (9)

  • Figure 1: Setting: CATE estimation under DP. Only the trusted data curator can access the data, while published CATE estimates do not allow private information about individuals to be inferred.
  • Figure 2: Comparison of relevant literature.
  • Figure 3: Dataset 1 (finite queries). Predictions under different base learners and privacy budgets.
  • Figure 4: Dataset 1 (functional queries). Predictions under different privacy budgets.
  • Figure 5: Dataset 2. Prediction errors under different privacy budgets of DP-CATE (finite) on KR and NN and DP-CATE (functional) over 10 runs. Plots centered at the PEHE of the original learner.
  • ...and 4 more figures

Theorems & Definitions (34)

  • Definition 1: Differential privacy Dwork.2009
  • Definition 2: Gaussian noise privacy mechanism Dwork.2014
  • Definition 3
  • Theorem 1: for finite queries
  • proof
  • Corollary 1: Parametric second-stage regression
  • proof
  • Corollary 2: Non-parametric second-stage regression
  • proof
  • Theorem 2: Neyman-orthogonality and quasi-oracle efficiency of
  • ...and 24 more