Online learning has been adopted rapidly by educational institutions and organizations. Despite its many advantages, including 24/7 access, high flexibility, rich content, and low cost, online learning suffers from high dropout rates that hamper pedagogical and economic goal outcomes. Enhanced student dropout prediction tools would help providers proactively detect students at risk of leaving and identify factors that they might address to help students continue their learning experience. Therefore, this study seeks to improve student dropout predictions, with three main contributions. First, it benchmarks a recently proposed logit leaf model (LLM) algorithm against eight other algorithms, using a real-life data set of 10,554 students of a global subscription-based online learning provider. The LLM outperforms all other methods in finding a balance between predictive performance and comprehensibility. Second, a new multilevel informative visualization of the LLM adds novel benefits, relative to a standard LLM visualization. Third, this research specifies the impacts of student demographics; classroom characteristics; and academic, cognitive, and behavioral engagement variables on student dropout. In reviewing LLM segments, these results show that different insights emerge for various student segments with different learning patterns. This notable result can be used to personalize student retention campaigns.