Causal Discovery using In-Distribution Knockoff Interventions

This direction is merged into another direction: non-stationary causal discovery. Its last version stays readable and citable here.

Version 1 of 17 papers17 conceptsfrozen 2026-10-04

To understand how one thing causes another, scientists often perform interventions: they change one variable and observe what happens to the rest of the system. In the world of machine learning, we simulate these interventions by perturbing input data and watching how a model's predictions change.

However, deep learning models are notoriously sensitive to "out-of-distribution" (OOD) data. If you intervene on a variable by replacing it with random noise or a constant value, you create a data point that looks nothing like what the model saw during training. The model’s performance might crash not because the variable was causally important, but simply because the model is confused by an unrealistic input.

Recent research by Maha Shadaydeh, Joachim Denzler, and their colleagues addresses this by using "Knockoff" variables to perform interventions that stay within the model's comfort zone.

The Problem of the Broken Manifold

When we train a deep learning model—such as the DeepAR probabilistic forecasting tool used in Causal Discovery using Model Invariance through Knockoff Interventions—the model learns the "manifold" of the data. It understands not just the individual variables, but the complex, nonlinear correlations between them.

If we want to test if variable $X$ causes $Y$, a traditional approach might be to "break" the connection by replacing $X$ with Gaussian noise. But as Shadaydeh and Wasim Ahmad point out in Deep Learning-based Group Causal Inference in Multivariate Time-series, these synthetic perturbations often fall outside the training distribution. This shift leads to unreliable predictions, making it impossible to tell if a change in the model's residual error is a genuine causal signal or just an artifact of OOD degradation.

Knockoffs as Statistical Stunt Doubles

The solution is to use Knockoffs. A knockoff is a synthetic "twin" of a variable. It is designed to be a statistical ringer: it preserves the original data's distribution and its correlation structure with all other variables, but it is conditionally independent of the target variable we are trying to predict.

Think of a knockoff as a stunt double. To the model, the knockoff looks exactly like the original actor (the real variable). It fits perfectly into the scene (the data manifold). However, because the knockoff doesn't actually contain the unique causal information of the original variable, we can use it to run a controlled experiment:

  1. Train a model on the original data.
  2. Replace a variable (or a group of variables) with its knockoff.
  3. Observe the prediction error.

If the model’s error stays roughly the same, the model is "invariant" to that change, suggesting the variable wasn't a primary cause. If the error spikes, it’s a strong sign of a causal link. Because the knockoff is in-distribution, we can trust that the error spike is due to the loss of causal information, not model confusion.

Handling Hidden Causes and Groups

This framework is particularly powerful for complex time-series data where variables aren't just individual points but sequences with hidden influences.

In Sequential Causal Effect Variational Autoencoder: Time Series Causal Link Estimation under Hidden Confounding, Maha Shadaydeh and Violeta Teodora Trifunov introduced SCEVAE. This method uses knockoffs to generate "faithful counterfactuals"—scenarios that are physically and statistically plausible even when there are hidden variables (confounders) affecting the system. By keeping the counterfactuals within the learned data manifold, SCEVAE can estimate causal intensity more accurately than methods that rely on random noise.

Furthermore, the research extends to group causal inference. In many systems, like the human brain or the global climate, variables act in functional blocks. The gCDMI framework, detailed in Deep Learning-based Group Causal Inference in Multivariate Time-series, generates knockoffs for entire groups of variables. This allows researchers to untangle whether a specific region of the brain or a specific atmospheric pattern causes a downstream effect, while maintaining the intricate covariance structure that defines those groups.

Why It Works: Model Invariance

The core intuition behind this entire line of research is model invariance. A truly causal relationship should be robust. By using knockoffs, these researchers have created a way to probe that robustness without breaking the underlying machinery of the deep learning models.

By ensuring that interventions are "realistic" (in-distribution), they transform deep learning models from "black-box" predictors into reliable tools for scientific discovery, capable of identifying the hidden levers that drive complex systems.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_discovery_using_in_distribution_k_e97e8d34,
  title = {Causal Discovery using In-Distribution Knockoff Interventions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-using-in-distribution-knockoff-interventions/rdr_e97e8d34a61745bdace5b2ca6752182f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_discovery_using_in_distribution_k_e97e8d34_v1,
  title = {Causal Discovery using In-Distribution Knockoff Interventions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-using-in-distribution-knockoff-interventions/rdr_e97e8d34a61745bdace5b2ca6752182f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}