Are Linear Regression Models White Box and Interpretable?

"Simple" models like linear regression are not automatically interpretable: Despite widespread assumptions that linear regression is a "white box" model, the paper argues this perception is inaccurate. Linear regression models face significant interpretability challenges that are often overlooked, suggesting the common distinction between "simple" and "complex" models oversimplifies explainability concerns.

Multiple hidden complexities undermine linear regression interpretability: The paper identifies eight specific challenges affecting linear regression's interpretability, including linearity assumptions, multicollinearity (when features are correlated with each other), normalization choices, and uncertainty quantification—issues that can obscure rather than clarify what the model is actually doing.

Fairness and feature contribution are not straightforward in linear models: Even when a linear regression model shows that a feature has a positive or negative effect on outcomes, this doesn't guarantee the explanation is fair or that the feature's true contribution is understood, particularly when dealing with real-world data complexities.

Simple models deserve the same explainability scrutiny as complex ones: The paper recommends that organizations should apply the same rigorous explainability standards and XAI tools to linear regression models as they do to deep neural networks, rather than assuming simplicity equals transparency.

Responsible AI requires questioning default assumptions: For practitioners implementing AI systems, this research suggests that relying on conventional wisdom about model interpretability—without critically examining actual explainability challenges—can create false confidence in understanding how models make decisions.