https://nothing-so-practical.com/categories/explainer/Recent content in Explainer on Nothing So PracticalHugoen-usSun, 19 Apr 2026 00:00:00 +0000https://nothing-so-practical.com/post/do-the-do-calc/Sun, 19 Apr 2026 00:00:00 +0000https://nothing-so-practical.com/post/do-the-do-calc/<h2 id="do-calculus-for-humans">do-Calculus for Humans</h2> <p>Pearl&rsquo;s structural approach to causal inference is built around DAGs (directed acyclic graphs). In many cases, applying this framework boils down to measuring and adjusting for confounding variables. A key insight is that when working with purely observational data, controlling for all confounders can be equivalent to assigning the treatment level, as we would in an experiment. This claim rests on a curious concept: the do-operator.</p> <p>At first glance, <span class="katex"><span class="katex-mathml"><math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mrow><mtext>do()</mtext></mrow><annotation encoding="application/x-tex">\text{do()}</annotation></semantics></math></span><span class="katex-html" aria-hidden="true"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord text"><span class="mord">do()</span></span></span></span></span> looks like an esoteric way of saying control for confounds. However, that interpretation only holds when key confounders are measured. When we can&rsquo;t measure all confounds, the role of the do-operator becomes clear. In such cases, the rules of do-calculus provide a system for determining whether a causal effect is identifiable from data given a set of assumptions about the data generating process. If identification is possible, do-calculus produces a formula for estimating the treatment effect. It can also tell you when the data are not capable of recovering the target estimand.</p>