¾«¶«Ó°ÊÓ

It’s tempting to think of data samples as neatly organized, with each sample point sitting all by itself like cupcakes in a bakery display: perfectly spaced and unaffected by the rest. But in the real world, especially in scenarios where people are involved, data samples influence and affect each other. A student’s performance might depend on their study group, a shopper’s decision could be shaped by a friend’s review, and one person getting vaccinated might protect their entire household. Ignoring these interactions when working with such data may lead us to wildly wrong conclusions.

In this talk, we explore what happens when standard causal inference tools that are designed under the assumption that data are independent and identically distributed (IID) are applied to messy, interconnected data. We dive into settings where treatments spill over between individuals and where outcomes become entangled. Using causal graphs, we develop a systematic framework to detect and quantify this interaction-induced bias. We identify conditions under which IID-based methods still work surprisingly well, and discuss techniques to correct for bias when they don’t. Whether you're a theorist, a practitioner, or just curious about how cause and effect work in our hyper-connected world, this talk offers a fresh perspective on doing causal inference when data just won't sit still!