🧭Lecture 18 - Causal questions with observational data

1. Without randomisation, a causal claim rests on a diagram of what causes what, drawn from knowledge of the world and not from the data.
2. Adjust for things that cause both. Leave alone things that are caused by both.
3. A model can fit beautifully and be about the wrong thing. Fit is not truth.
4. Show the work to someone who uses statistics for a living before the person who marks it.

Read with

• 🎓 Lecture 13 - Confounding and what regression cannot tell you
• 🔤 Confounder
• 🧩 Confounding
• 00. Statistical Rethinking Hub
• Port Authority open vessel-delay data

• context: Hana Kobayashi rang at four. my best variable in the project is mostly the season. tonight's lecture is that mistake, as a topic
• observational data = nobody randomised anything. which is nearly all data
• draw the arrows FIRST. what causes what. then decide what to adjust for
• adjust for common causes. do NOT adjust for common effects, it invents a link
• natural experiments: something outside changed one group and not the other
• Hale: the data cannot tell you the arrows. you bring the arrows
• showed him the refit at the break. he nodded and asked what else changes with the season. there is always a smaller question