With the availability of machine learning (ML) models, predictive modeling finally brings psychology in a position to become a full-fledged natural science. In predictive modeling, we search for an unknown function f(x) that allows for making generalizable predictions for an untrained test sample. While the term overfitting was much of a "philosophical" question in psychology for a long time, in part addressed by model evaluation metrics such as AIC/BIC, now overfitting can be simply quantified by the difference in explained variance between the training and validation/test samples. Now we can guess models that allow to describe the complexity of us human beings.

Many researchers counter-argue that ML models are so complex that they can hardly be understood. I argue that, for instance, penalized regression models can easily be understood in terms of simple linear main or interaction effects. And for nonlinear effects, an algorithmic model is always superior to mere verbal theorizing. It may take some work to understand such complex effects and their interactions, but any random forest should be reducible to a simple decision tree with only a minor amount of lost explained variance. And a simple decision tree can be applied by anyone! Machine learning takes time, of course, and I hope that one day, psychology will get his first "Einstein" finding a closed-form expression for the initial guesses for f(x). A well-known ML principle is "no food for free", though, so we need Phd students good in math. But at least for undergraduates, ML now is quite simple, because JMP Pro is now freely downloadable for anyone with a university email. We use this software in our teaching and it is so simple to learn ML. So you may start training your first neural network model, today.

Author: Markus Hofmann