Identifying Selection Bias from Observational Data

David Kaltenpoth, Jilles Vreeken

Identifying Selection Bias from Observational Data

Selection bias—where observed data represents only a specific sub-population—is a critical challenge in machine learning, often leading to spurious correlations and biased models. While standard correction methods typically require a "clean" reference dataset, this research demonstrates that selection bias can often be identified from a single observed dataset by detecting broken symmetries and exploiting the mathematical properties of probability distributions.

The Problem of Identifiability

In many real-world scenarios, we deal with "selection effects" where data is pruned based on certain criteria. A classic example is Berkson’s Paradox, where a study might falsely suggest a negative correlation between two independent variables because the sample was restricted to a specific group (e.g., hospital patients). This paper addresses the identifiability problem: determining the conditions under which we can mathematically prove a single dataset has been biased and subsequently recover the original distribution and selection boundary.

Selection bias represented as a graphical model (a) and visual examples of linear selection boundaries (b, c).

Methodological Frameworks

The authors introduce two primary approaches for detecting bias without a reference population:

  1. EXP (Parametric Approach): This method assumes the underlying data follows an exponential family distribution (e.g., Gaussian or Poisson). The authors prove that applying a linear selection rule (e.g., $a^\top X > 0$) results in a distribution that no longer belongs to the exponential family. By maximizing the log-likelihood of a model that explicitly accounts for a selection boundary, they can estimate both the original parameters and the hidden selection criteria.
  2. INV (Invariance-based Approach): This non-parametric method relies on the inherent invariances (symmetries) of many natural distributions. For example, a centered Gaussian is rotationally symmetric; selection bias "chops off" part of the distribution, breaking this symmetry. Using Kernel Mean Embeddings, the algorithm identifies an orthogonal transformation that should have been an invariance and uses the resulting "bad" points to reconstruct the selection boundary via a Support Vector Machine (SVM).

Experimental Results

The methods were evaluated on synthetic data, the Palmer Penguins dataset, and the Open Exoplanet Catalogue. Key findings include:

  • Boundary Recovery: Both EXP and INV achieved high cosine similarity (often > 0.8) between discovered and true selection boundaries.
  • Variable Identification: The algorithms accurately identified which specific variables were used to filter the data, as measured by Jaccard similarity.
  • Robustness: While EXP performs well on Gaussian data, INV is significantly more robust for heavy-tailed distributions like the $t$-distribution, where parametric assumptions often fail.
Variable identification performance (Jaccard similarity) for different methods across normal and t-distributed data.

Conclusion and Future Directions

The core insight of this work is that symmetry serves as a "smoking gun" for selection bias. The internal structure of biased data contains sufficient clues to suggest what the unbiased version should have looked like. While currently focused on linear selection boundaries, future research aims to utilize Variational Autoencoders (VAEs) or Generative Adversarial Networks (GANs) to identify more complex, non-linear selection biases.

Summary updated

BibTeX
@inproceedings{kaltenpoth2023identifying,
  author = {David Kaltenpoth and Jilles Vreeken},
  title = {{Identifying Selection Bias from Observational Data}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i7.25987},
}
Figure