Epistemic frontiers and the distinction between causality, information, and predictability in pattern recognition.

Area of Science:

• Machine Learning
• Causal Inference
• Statistical Modeling

Background:

• High predictive accuracy is often mistaken for causal understanding.
• Machine learning models can exploit spurious correlations and confounding factors.
• Post-hoc explanations may accurately reflect model behavior but mislead about the real phenomenon.

Purpose of the Study:

• To propose a framework separating causal relations, population-level statistical dependence, and protocol-dependent predictive effects.
• To clarify why predictive success and feature attributions do not automatically imply mechanistic interpretations.
• To provide a principled reference for true signal using conditional mutual information.

Main Methods:

• Developed a framework to distinguish three layers of evidence: causal relations, population dependence, and finite-sample effects.
• Utilized controlled simulations to test the framework's efficacy.
• Analyzed the impact of bootstrap resampling and SHAP (SHapley Additive exPlanations) on feature importance.

Main Results:

• Bootstrap resampling can amplify chance correlations, leading to false positives.
• SHAP can assign high importance to confounded variables, despite being faithful to the model.
• Conditional mutual information serves as a principled measure for population predictive value.

Conclusions:

• Predictive success and feature importance are protocol-bounded evidence, not direct indicators of causality.
• Interpreting model results requires reporting the protocol, robustness checks, and intended inferential scope.
• Distinguishing between predictive effects and causal understanding is crucial for reliable scientific conclusions.