Comparing methods addressing multi-collinearity when developing\n prediction models

Abstract

Clinical prediction models are developed widely across medical disciplines.\nWhen predictors in such models are highly collinear, unexpected or spurious\npredictor-outcome associations may occur, thereby potentially reducing\nface-validity and explainability of the prediction model. Collinearity can be\ndealt with by exclusion of collinear predictors, but when there is no a priori\nmotivation (besides collinearity) to include or exclude specific predictors,\nsuch an approach is arbitrary and possibly inappropriate. We compare different\nmethods to address collinearity, including shrinkage, dimensionality reduction,\nand constrained optimization. The effectiveness of these methods is illustrated\nvia simulations. In the conducted simulations, no effect of collinearity was\nobserved on predictive outcomes. However, a negative effect of collinearity on\nthe stability of predictor selection was found, affecting all compared methods,\nbut in particular methods that perform strong predictor selection (e.g.,\nLasso).}\n

Ähnliche Arbeiten