Classification Trees vs. Logistic Regression in the Generic Skill Development in Engineering

Abstract

From an experimental approach, the performance of Logistic Regression vs Classification Trees is evaluated in the context of two generic engineering skills (quantitative reasoning and reading comprehension). These methods incorporate two separate predictor scenarios (indicators only, and only constructs derived from Principal Component Analysis: ACP). The sample is 7,395 instances of Saber Pro 2015-3, 2014-3 (Colombia). The study considers: training (70% of the sample), prediction (30% remaining) and experimentation (176 original observations, design of three factors: method, type of predictor and competence). The response variable is a new metric (Underlying Asset Ratio obtained by ACP). Both methods present similar performance in the scenario of only indicators, but not in the scenario of constructs (Regression Logistics, better performance).