Predictive and explanatory models might miss informative features in educational data

Abstract

We encounter variables with little variation often in educational data mining (EDM) and discipline-based education research (DBER) due to the demographics of higher education and the questions we ask. Yet, little work has examined how to analyze such data. Therefore, we conducted a simulation study using logistic regression, penalized regression, and random forest. We systematically varied the fraction of positive outcomes, feature imbalances, and odds ratios. We find the algorithms treat features with the same odds ratios differently based on the features' imbalance and the outcome imbalance. While none of the algorithms fully solved the problem, penalized approaches such as Firth and Log-F reduced the scale of the problem. Our results suggest that EDM and DBER studies might contain false negatives when determining which variables are related to an outcome. We then apply our findings to a graduate admissions data set and we propose recommendations for researchers working with the kind of imbalanced data common to EDM and DBER studies.

Rubric-based holistic review represents a change from traditional graduate admissions approaches in physics

“…First, we used the AUC permutation feature importance [51] as it is claimed to be less biased than the accuracy-based permutation importance when input features differ in scale (as do our factors listed in Table II) and when the predicted variable is not split evenly between the two outcomes. In practice, our previous work suggests which method we pick will have minimal effect on the conclusions [52]. Under this approach, each feature is randomly permuted and then passed through the model to make a prediction.…”