Causal interpretation rules for encoding and decoding models in neuroimaging.

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Causal terminology is frequently used when interpreting encoding and decoding models in neuroimaging.
• The validity of these causal statements often lacks empirical support.
• Distinguishing between encoding and decoding models alone is insufficient for warranted causal claims.

Purpose of the Study:

• To investigate which causal statements derived from neuroimaging models are empirically supported.
• To clarify the role of experimental paradigms (stimulus-based vs. response-based) in causal interpretations.
• To explore how combining encoding and decoding models can yield deeper causal insights.

Main Methods:

• Theoretical analysis of encoding and decoding models in different experimental paradigms.
• Examination of feature relevance in stimulus-based and response-based settings.
• Integration of encoding and decoding models trained on identical neuroimaging datasets.

Main Results:

• Encoding models in stimulus-based paradigms support unambiguous causal interpretations.
• Relevant features in encoding and decoding models have distinct meanings across paradigms.
• Combining encoding and decoding models provides causal insights beyond individual model capabilities.

Conclusions:

• The interpretation of causal relationships from neuroimaging models necessitates careful consideration of model type and experimental design.
• Stimulus-based encoding models offer the most direct path to causal inference.
• Hybrid approaches combining encoding and decoding models enhance the understanding of neural causal mechanisms.