The Effective Data Scientist

Why did early COVID-19 studies suggest that smokers were less likely to be hospitalized? The answer wasn’t in the analysis. It was in how the data was collected. In this episode of The Effective Data Scientist, Aziza and Paolo dig into collider bias. It happens when two independent variables start to look related because you’ve conditioned on a third variable that both of them influence. It’s a subtle trap, and it caught a lot of observational research during the pandemic. In this episode: - What collider bias is, explained with a vaccine and infection example - How the “smokers are protected” COVID-19 finding came from the way data was selected - Why unexpected results should make you look at the data generating process - How careful study design and simple causal diagrams help prevent the bias and explain it to stakeholders - How AI and simulations can make these ideas easier to grasp for people without a statistics background “Collider bias reminds us that many data problems are actually poor design problems. Once the data is collected the wrong way, no algorithm can fully fix it.” – Paolo Key takeaway: how your data was selected matters just as much as how you analyze it. Want to go deeper? Aziza and Paolo recommend exploring resources on causal inference.

What does it take to turn elegant statistical theory into methods engineers actually use? In this episode, Aziza talks with Brad Jones, Principal Research Fellow at JMP and one of the most influential figures in modern design of experiments (DOE). Brad invented Definitive Screening Designs, co-authored the textbook Optimal Design of Experiments: A Case Study Approach, and in 2026 received the ENBIS George Box Medal for outstanding contributions to industrial statistics. In this episode, we talk about: - How Brad found his way into statistics and experimental design, and why he combined consulting, software and research - What was missing in traditional DOE, and when it became clear that industry needed something more practical - The role of software like JMP in putting advanced methods into practitioners’ hands - The problem Definitive Screening Designs were built to solve, and the limits of classic screening approaches - What connects his work on split-plot designs, Bayesian design and algorithmic methods - Real examples of experimental design making a difference in industry, and why some methods get adopted while others stay academic - George Box’s influence, and how to balance theoretical optimality with real-world constraints - Advice for engineers and data scientists: the mindset shift DOE requires, and how to go from using methods to innovating

AI can dramatically expand what one data scientist can do. But where does that multiplier stop? Aziza sits down with Glen Wright Colopy of Polaris Data Labs to explore what AI-powered data science actually looks like in practice. They discuss how to choose and configure AI tools, when more capable models really matter, what can go wrong when AI is adopted without a clear framework, and whether AI can genuinely replace a data scientist. They also explore ROI, stakeholder trust, critical thinking, and the skills that will matter most as the profession evolves.

In this episode of The Effective Data Scientist Podcast, we sit down with Jeff Stevenson, CMC MSAT Specialist at Boehringer Ingelheim and one of the most influential “Data Evangelists” in modern biopharma. Jeff shares how 30 years across Quality, Supply Chain, and MSAT shaped his philosophy on integrity, responsibility, and the true meaning of data readiness. Together, we explore the STIPS mindset, why most companies still fall into the IT Trap, how psychological safety drives technical excellence, and which AI tools are finally helping teams achieve analytic‑ready data. Jeff closes with a powerful reflection on what he would do differently if he were starting his career today.

How can you get better, more reliable results from AI? In this episode, Paolo and Aziza explore why effective prompting is about much more than asking good questions. We discuss how structured prompting frameworks such as ROSES, CRISP, CARE, TRACE, and APE can help data scientists provide clearer instructions, reduce ambiguity, and improve the quality of AI-generated outputs. We also share practical examples of applying these frameworks to real-world data challenges and discuss why designing how AI should think is often more important than the answer itself.