Researchers have long tried to identify factors that could explain why programming is easier for some than the others or that can be used to predict programming performance. The motivation behind most studies has been identifying students who are at risk to fail and improving passing rates on introductory courses as these have a direct impact on retention rates. Various potential factors have been identified, and these include factors related to students' background, programming behavior or psychological and cognitive characteristics. However, the results have been inconsistent.
This thesis replicates some of these previous studies in a new context, and pairwise analyses of various factors and performance are performed. We have data collected from 3 different cohorts of an introductory Java programming course that contains a large number of exercises and where personal assistance is available. In addition, this thesis contributes to the topic by modeling the dependencies between several of these factors. This is done by learning a Bayesian network from the data. We will then evaluate these networks by trying to predict whether students will pass or fail the course. The focus is on factors related to students' background and psychological and cognitive characteristics.
No clear predictors were identified in this study. We were able to find weak correlations between some of the factors and programming performance. However, in general, the correlations we found were smaller than in previous studies or nonexistent. In addition, finding just one optimal network that describes the domain is not straight-forward, and the classification rates obtained were poor. Thus, the results suggest that factors related to students' background and psychological and cognitive characteristics that were included in this study are not good predictors of programming performance in our context.
