User-Guided Interpretable Models: Rashomon Effect, Interaction, and Computation

When moving from a simple univariate regression to a multivariate one, practitioners are often confronted with a frustrating phenomenon: the coefficient sign of a variable can flip, so that it disagrees with its (univariate) relationship with the outcome.An effect that is positive in isolation can appear negative in the context of other variables, or vice versa.This instability undermines the interpretability of the model, especially in high-stakes domains like health care where domain knowledge suggests a clear directional relationship.In "Univariate-Guided Sparse Regression," Chatterjee et al. ( 2025) tackle this problem.They propose a method, uniLasso, that preserves the signs of univariate coefficients while leveraging the predictive strength of a sparse multivariate model.Their approach uses leave-one-datapoint-out univariate slopes as guides within an adaptive lasso framework.The result is a model that is sparser and more accurate than the standard lasso but also, by design, more aligned with the foundational insights gleaned from single-variable analysis.This commentary discusses their work in the context of a broader challenge: aligning models more generally with human intuition while keeping them sparse enough so that people can easily understand them.