Adding 68 cortical thickness values did not reduce observed Raven prediction error in PIOP2 after ridge selection in PIOP1

We trained models in 208 PIOP1 participants and tested them in 224 PIOP2 participants. The reference model used age, recorded sex, and education category. The added model used 68 cortical thickness values. We selected its ridge penalty only in PIOP1. Reference error was 4.789449 Raven points. Error with thickness was 4.789478 points. Gain was -0.000029 points. Its approximate 95% bootstrap interval was [-0.352095, 0.017988]. The 5000 bootstrap draws refit both models and selected the penalty again. All copies of one person stayed in the same validation fold. With the fitted models fixed, the gain interval was [-0.000151, 0.000090]. The interval from refitted models was 1534 times wider. Reverse prediction, head-size adjustment, and global thickness checks also gave no clear gain. The source file and paper give different PIOP1 education proportions. We used the labels in the source file. The package contains fixed source tables, all model fits, code, and checks with a separate solver. The result applies to this model procedure and these complete records. It does not show that cortical thickness has no relation to reasoning.

cortical thickness · Raven reasoning · prediction · ridge regression · AOMIC · bootstrap uncertainty