Handling collinearity in Picture-Word Interference studies: Balancing theoretical relevance and model robustness
We investigated the Picture-Word Interference Effect by parametrically manipulating semantic similarity while controlling for lexical association and orthographic/phonological similarity. We also aimed to control for lexical-semantic variables at both the picture and word levels, picture-specific visual properties, and stimulus-independent confounders. However, collinearity is a major challenge in psycholinguistics as linguistic variables are highly interrelated. Additionally, with predictors for both pictures and words (e.g., word frequency for the picture name versus the distractor word), it becomes unclear which are more critical for performance modulation, complicating predictor selection. This affects model building, interpretability, and robustness. We seek guidance on detecting and mitigating collinearity while ensuring theoretically relevant controls. Should we use dimensionality reduction, residualization, or penalized regression? Furthermore, we welcome insights on structuring mixed-effects models to balance complexity and convergence. Feedback from experts in statistical modeling and psycholinguistics would be invaluable in refining our approach.