🌦️Confounding

Two things move together, and it looks as if one drives the other, but a third thing is driving both. That third thing is the 🔤 Confounder. Confounding is what you call the situation once it has fooled you.

For a variable to confound it has to be connected to the supposed cause and to the outcome. Ice-cream sales and drownings rise together. Summer does both.

What regression can and cannot do about it:

• Put the confounder in the model and its share is separated out.
• A regression only adjusts for what is in it. It has no opinion about what I forgot.
• Only 🧩 Randomisation deals with the confounders nobody thought of.

Example: my project, 📌 Statistics II - Project - What predicts vessel delay at Greywater. One variable I was proud of turned out to be standing in for the time of year. The vessel calls that scored high on it bunch in the stormy months, and so do the delays. Hana Kobayashi found it in an evening.

Often confused with: plain coincidence, or reverse causation. Here the association is real. It is the explanation that is wrong.

Taught in 🎓 Lecture 13 - Confounding and what regression cannot tell you, seven weeks before I needed it.

Caught in the wild

• 30 September: Hale put up a tidy line from port data and asked what else changes with the calendar. Everything does.
• 18 November: Hana caught it in the project: my proud variable was the season. Refitted with month in the model and the effect shrank.
• 25 November: submitted the project with month in the model, and one paragraph listing what else I may have left out. It is the paragraph I trust most.