Toward Stable Predictions of Apprentices’ Training Success

Abstract

Mechanical (statistical) predictions have proven to be useful in personnel selection. However, such predictions require the use of an algorithm to aggregate different predictor scores. The identification of such an algorithm requires analyzing predictor and criterion data obtained from previous applicants. The present manuscript compared predictions made by two different statistical methods: artificial neural networks (ANNs) and multiple regression analysis. Therefore, three consecutive cohorts of apprentices (n = 322, 217, and 118) were examined. Algorithms were derived from one cohort and applied to more recent cohorts. It was shown that ANNs outperformed linear predictions in a cross-validation within the cohorts. However, applying trained ANNs to other samples resulted in a predictive power which was worse than most of the linear predictions. Thus, we conclude that ANNs should only be used as selection algorithm if their validity in different cohorts has been confirmed.