Private Causal Inference

Matt J. Kusner, Yu Sun, Karthik Sridharan, Kilian Q. WeinbergerAISTATS 2016 · OpenReview

Abstract: Causal inference deals with identifying which random variables ”cause” or control other random variables. Recent advances on the topic of causal inference based on tools from statistical estimation have enabled researchers to uncover these relationships in complex datasets. However, these datasets—often containing sensitive medical or financial information—pose significant privacy risks. This work introduces the first formal framework for Private Causal Inference, merging the predictive power of Additive Noise Models with the rigorous guarantees of Differential Privacy.

Introduction

In the era of big data, understanding "what causes what" is the ultimate goal for researchers in fields ranging from epidemiology to economics. While correlation is easily found, causal discovery allows us to understand the underlying mechanisms of a system. However, standard causal discovery algorithms are inherently "privacy-leaking." Because these models often analyze subtle patterns in noise and residuals, they are vulnerable to reconstruction attacks that could reveal the presence of specific individuals in a dataset.

This research, originally presented at AISTATS 2016, addresses this gap by providing a framework to identify causal relationships while maintaining Differential Privacy (DP). The core challenge lies in the fact that causal inference relies on sensitive statistical measures that are highly susceptible to single-point changes in a dataset. By adding calibrated noise to these measures, the authors demonstrate that causal truths can be uncovered without sacrificing individual anonymity.

Methodology

The authors focus on the Additive Noise Model (ANM) framework. In an ANM, if $X$ causes $Y$, then $Y$ can be expressed as a function of $X$ plus some independent noise ($Y = f(X) + N$). Crucially, this independence typically does not hold in the reverse direction ($X \to Y$). To privatize this process, the researchers applied DP to the "dependence scores" used to determine the causal direction.

Sensitivity Analysis and Robust Scoring

To implement DP, one must determine the "global sensitivity" of a metric—how much a single data point can change the result. The authors mathematically derived sensitivity bounds for common dependence measures:

  • Spearman’s $\rho$: $30/m$
  • Kendall’s $\tau$: $4/m$
  • HSIC (Hilbert-Schmidt Independence Criterion)

Recognizing that standard variance is often too sensitive to outliers for effective privatization, they also introduced a robust dependence score based on the Interquartile Range (IQR).

Training vs. Test Privacy

The framework splits the problem into two distinct phases. For the test set, global sensitivity is used to add Laplace noise. For the training set (where the regression model is built), they utilize a "Propose-Test-Release" framework. This ensures that if a dataset is too unstable—meaning one person's data could significantly alter the model—the algorithm refuses to provide an answer rather than risking a privacy leak.

Experimental Results and Analysis

The methods were evaluated using the Cause-Effect Pairs Challenge, a standard benchmark of real-world variable pairs with known causal directions (e.g., Altitude $\to$ Temperature).

Performance of various private dependence measures (Spearman's $\rho$, Kendall's $\tau$, HSIC, and IQR) as a function of the privacy budget $\epsilon$.

As shown in the comparison above, the private versions of HSIC and rank-based scores (Spearman and Kendall) maintain high accuracy in identifying the correct causal direction, even with strict privacy budgets (low $\epsilon$). The rank-based correlations proved particularly effective because their inherent stability makes them more compatible with DP noise.

The Regularization "Hump"

A key finding in the training phase was the relationship between regularization ($\lambda$) and inference accuracy.

The "hump" effect: Probability of correct inference as a function of the regularization parameter $\lambda$.

There is a clear "sweet spot" for $\lambda$. If the regularization is too small, the model becomes overly sensitive to individual data points, triggering privacy protections that may lead to "null" results. If $\lambda$ is too large, the model becomes too "blurry" to distinguish the independent noise required to identify the causal direction.

Discussion and Future Work

The research highlights that causal inference is often robust to privacy noise because the "margin" between the correct and incorrect causal directions is frequently large enough to withstand calibrated noise. Furthermore, as dataset size ($m$) increases, the required noise per data point effectively shrinks, making private causal inference highly viable for large-scale data analysis.

However, limitations remain. This work primarily addresses bivariate pairs ($X, Y$). Extending these guarantees to complex Directed Acyclic Graphs (DAGs) with many variables and potential hidden confounders is a significant area for future research. Additionally, the tendency of the algorithm to return "null" results on unstable data, while necessary for privacy, presents a challenge for researchers working with smaller, noisier datasets.

Summary updated

Abstract

Causal inference deals with identifying which random variables ”cause” or control other random variables. Recent advances on the topic of causal inference based on tools from statistical estimation...
BibTeX
@inproceedings{kusner2016private,
  author = {Matt J. Kusner and Yu Sun and Karthik Sridharan and Kilian Q. Weinberger},
  title = {{Private Causal Inference}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2016},
  url = {https://openreview.net/forum?id=3D0UC7o-0IN},
}
Figure