Bridging the Gap: Statistical Rigor and the Promise of Data Science

As data science increasingly takes center stage in applied estimation problems, a critical tension emerges: while data science offers powerful tools and better point estimates, it often does so by sacrificing the rigorous uncertainty quantification that traditional statistical methods provide. This talk explores how data science, particularly modern machine learning techniques like XGBoost, can augment—rather than replace—statistical theory to improve estimation practices.

We will begin by acknowledging a key limitation of data science: its theoretical foundations remain underdeveloped compared to classical statistics. As a result, despite its empirical strengths, data science still relies on statistical theory to ensure principled inference, especially when dealing with uncertainty. Through examples from statistical consulting and simulation-based modeling, we will illustrate why understanding convergence and variability is essential for sound decision-making.

Importantly, we will also examine practical trade-offs. Data science often imposes fewer assumptions than traditional statistical models, allowing greater flexibility and performance. However, this comes at the cost of interpretability and reliable confidence intervals—an issue particularly concerning in high-stakes fields like medicine, where uncertainty matters deeply. Ultimately, this talk advocates for a hybrid approach: using statistical theory as a foundation, and data science as a powerful augmentation. While not all traditional statistical tools remain essential, many do—and knowing when and how to combine them with modern techniques is the path forward.