# Idea Say we're interested in the causal effect of body mass index (BMI) on health outcomes: **BMI -> health outcomes**. We have a problem because there are **many potential ways one could achieve a BMI of a particular value**. Each way might be associated with different outcomes. Problem: There's not **one version of treatment**. The "BMI treatment" here in fact can reflect multiple treatments (e.g., exercise, smoking, genes). So multiple treatments are affecting health outcomes via BMI. BMI **is not directly manipulable**—causal effects refer to the effects of some intervention that can be directly manipulated (Holland 1986: "no causation without manipulation"). Thus, it's better to focus on the causal effects of interventions that aim at manipulating weight. ```mermaid graph LR A[BMI] B[Health outcome] X1[Exercise frequency] X2[Smoking] X3[Genes] A --> B X1 --> A X2 --> A X3 --> A X1 --> B X2 --> B X3 --> B ``` # References - [Hypothetical interventions - Welcome and Introduction to Causal Effects | Coursera](https://www.coursera.org/learn/crash-course-in-causality/lecture/Lgb6O/hypothetical-interventions)