Untersuchung von Geschlechterunterschieden in PISA mittels Neuronaler Netze

Research on gender differences in academic achievement typically focuses on differences in summary performance. There is reason to believe this view masks some of the ways gender is related to performance on scholastic tasks. To investigate this, we used an Artificial Neural Network (ANN) to predict gender from responses to individual items. Because ANNs excel at detecting complex and non-linear patterns, we took the network's accuracy as a measure of the extent of gender-related differences contained in the responses. For a comparison with the traditional method of understanding gender differences through aggregate performance differences, we also developed a model predicting gender based on aggregate performances. Contrary to the findings of Loesche (2019), who applied a similar method to a large-scale assessment of German third-graders, the comparison indicates that aggregate performances provide a reasonable approximation of the overall extent to which gender is associated with item responses. However, this observation does not hold for performances in individual domains, particularly science, or items. An exploratory analysis suggests that this contradiction may be explained by the network’s ability to infer multiple competencies from individual tasks.