modeldependence

Model dependence, sometimes written as modeldependence, is the dependence of conclusions or predictions on the choice of model or modeling approach. It describes how sensitive results are to the assumptions inherent to a given model, including its structure, parameters, priors, and data transformations. If several models explain the data comparably well but diverge in their predictions for new cases, the results exhibit model dependence.

Model dependence appears in statistics, econometrics, physics, climate science, machine learning, and beyond. In statistics, parameter estimates can change with different likelihoods or priors. In physics and cosmology, inferred quantities often rely on the chosen theoretical framework. In ML, performance and explanations can depend on architecture, feature engineering, and training procedures, leading to different conclusions from the same data.

Unchecked model dependence can lead to overconfidence, biased policies, or poor generalization if the true generative process differs from all considered models. It is a form of model uncertainty and is distinct from data uncertainty.

Mitigation strategies include sensitivity analysis—systematically varying models or assumptions to gauge impact; model averaging or ensemble methods that pool across models; and explicit reporting of results across a set of plausible models. Information criteria, cross-validation, and Bayesian model comparison help assess relative plausibility. Robust or nonparametric methods can reduce reliance on strong assumptions. Transparent documentation of modeling choices is essential.

Related concepts include epistemic uncertainty, identifiability, and model misspecification. Model dependence is often addressed alongside data uncertainty to provide a complete picture of confidence in conclusions.