Three-quarter Sibling Regression for Denoising Observational Data

Shiv Shankar, Daniel Sheldon, Tao Sun, John Pickering, Thomas G. Dietterich

Abstract

Many ecological studies and conservation policies are based on field observations of species, which can be affected by systematic variability introduced by the observation process. A recently introduced causal modeling technique called 'half-sibling regression' can detect and correct for systematic errors in measurements of multiple independent random variables. However, it will remove intrinsic variability if the variables are dependent, and therefore does not apply to many situations, including modeling of species counts that are controlled by common causes. We present a technique called 'three-quarter sibling regression' to partially overcome this limitation. It can filter the effect of systematic noise when the latent variables have observed common causes. We provide theoretical justification of this approach, demonstrate its effectiveness on synthetic data, and show that it reduces systematic detection variability due to moon brightness in moth surveys.

Introduction

In fields such as ecology, astronomy, and public health, researchers frequently rely on observational data plagued by systematic measurement errors. For instance, in moth surveys using light traps, lunar brightness acts as a systematic noise source, suppressing counts across all species simultaneously. Distinguishing this shared noise from true biological signals is a critical challenge.

A previously proposed solution, half-sibling regression (HSR), assumes that observed variables are independent except for shared noise. In ecological contexts, however, species often share common environmental drivers (e.g., seasonality), making them naturally dependent. Applying HSR in these scenarios often leads to "throwing the baby out with the bathwater"—the algorithm mistakenly removes the genuine biological signal because it is correlated across species.

Methodology: Three-Quarter Sibling Regression

To address the limitations of HSR, we introduce Three-quarter Sibling Regression (3QS). This approach leverages known "process covariates"—environmental factors like day of the year—to disentangle shared signals from shared noise.

The causal distinction is illustrated in the sibling analogy:

  • Half-siblings (HSR): Share only a hidden noise source ($N$).
  • Three-quarter siblings (3QS): Share both a hidden noise source ($N$) and a known environmental source ($X$).
Causal structures for (a) half-sibling regression and (b) three-quarter sibling regression.

The 3QS method operates via a two-step residual process:

  1. Account for Knowns: A regression model predicts species counts based on known environmental factors ($X$). The resulting residual contains both the shared noise ($N$) and the intrinsic biological variation.
  2. Filter Noise: By analyzing the residuals of auxiliary "sibling" species, 3QS identifies common unexplained fluctuations. These are attributed to shared noise and subtracted from the target species' count.

Under an additive noise model, we prove that 3QS is guaranteed to improve accuracy over raw measurements and can theoretically achieve perfect signal reconstruction if the noise is fully captured by auxiliary species.

Evaluation and Results

We evaluated 3QS using both synthetic simulations and real-world moth observation data from the "Discover Life" project.

Synthetic Data Performance

In controlled simulations, HSR failed to distinguish natural correlation from noise, resulting in high Mean Squared Error (MSE). In contrast, 3QS demonstrated superior performance, with error rates decreasing significantly as the number of auxiliary "sibling" species ($n$) increased.

Comparison of 3QS and HS performance on synthetic data across varying species counts (c) and noise levels (d).

Denoising Real-World Moth Counts

Applying 3QS to eight years of moth survey data revealed its practical utility. Raw moth counts typically show a negative correlation with moon brightness. 3QS reduced this artificial bias by an average of 7%, leading to count distributions that more accurately reflect true biological variation. Furthermore, 3QS outperformed standard Generalized Additive Models (GAMs) and HSR in predicting abundance on "gold-standard" moonless nights.

Discussion and Future Work

While 3QS provides a robust framework for denoising dependent observational data, it assumes independence between the known environment ($X$) and hidden noise ($N$). Future research could explore relaxing this assumption and integrating Bayesian methods to provide uncertainty estimates for denoised outputs. Despite these limitations, 3QS represents a significant step forward in causal modeling for complex, noisy biological systems.

Summary updated

Abstract

Many ecological studies and conservation policies are based on field observations of species, which can be affected by systematic variability introduced by the observation process. A recently introduced causal modeling technique called 'half-sibling regression' can detect and correct for systematic errors in measurements of multiple independent random variables. However, it will remove intrinsic variability if the variables are dependent, and therefore does not apply to many situations, including modeling of species counts that are controlled by common causes. We present a technique called 'three-quarter sibling regression' to partially overcome this limitation. It can filter the effect of systematic noise when the latent variables have observed common causes. We provide theoretical justification of this approach, demonstrate its effectiveness on synthetic data, and show that it reduces systematic detection variability due to moon brightness in moth surveys.
BibTeX
@inproceedings{shankar2019three,
  author = {Shiv Shankar and Daniel Sheldon and Tao Sun and John Pickering and Thomas G. Dietterich},
  title = {{Three-quarter Sibling Regression for Denoising Observational Data}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  url = {https://openreview.net/forum?id=eLhp9R_j1W},
}
Figure