Related Experiment Video
Updated: Jun 2, 2026
05:39
A Fine Motor Task to Study Joint Kinematics in a Preclinical Model of Neurodegenerative Disease
Published on: June 13, 2025
Do complex models increase prediction of complex behaviours? Predicting driving ability in people with brain
Carrie R H Innes1, Dominic Lee, Chen Chen
1Van der Veer Institute for Parkinson's and Brain Research , Christchurch , New Zealand. carrie.innes@vanderveer.org.nz
Quarterly Journal of Experimental Psychology (2006)
|May 13, 2011
Summary
Complex kernel-based models show high initial accuracy for predicting driving ability but fail to generalize to new data, unlike simpler methods like discriminant analysis and logistic regression.
Area of Science:
- Neuroscience
- Machine Learning
- Behavioral Science
Background:
- Predicting complex behaviors like driving ability is challenging with simple models.
- More complex models, such as kernel methods, were hypothesized to improve prediction accuracy.
Purpose of the Study:
- To evaluate six modeling approaches for predicting driving ability in individuals with brain disorders.
- To compare the performance of traditional models (DA, BLR, NCRA) against kernel methods (SVM, PK, KP).
Main Methods:
- Assessed 501 individuals with brain disorders using computerized sensory-motor and cognitive tests (SMCTests™).
- Applied discriminant analysis (DA), binary logistic regression (BLR), nonlinear causal resource analysis (NCRA), support vector machine (SVM), product kernel density (PK), and kernel product density (KP).
- Utilized cross-validation to estimate prediction accuracy on independent data.
Main Results:
- Kernel methods (SVM, PK) achieved high initial classification accuracy (99.6%, 99.8%).
- Cross-validation revealed significantly lower accuracy for kernel methods (71-76%) compared to DA (74-75%) and BLR (75-76%).
- Overfitting was identified as a likely cause for the kernel methods' poor generalization.
Conclusions:
- Kernel-based models, while effective for initial data classification, do not improve prediction accuracy on independent data due to overfitting.
- Simpler, traditional models like discriminant analysis and binary logistic regression offer more reliable predictions for driving ability in this population.
- The study highlights the importance of cross-validation for assessing the true predictive power of models in complex behavioral tasks.
