Read on Twitter
Course: Watched week 3 in Intro to Causality from @CasualBrady. It feels like this week was laying down the foundations of the shared language of causality. Specifically graphs- connected nodes and edges that can diagram out relationships.
I had not seen this kind of graph in statistics before, but they remind me of linked lists or trees in comp sci. We also reviewed the assumptions that we make in order for this visual language to be true/valid and translate directly into the math used for causality.
The visuals feel more intuitive. I like this addition. Removing edges between nodes is what we mean when someone says their model “controlled” for a factor. It’s kind of an everyday term to say controlling for a variable but it seems clearer in this form
Another interesting factoid from the week was that conditioning on a collider - a node that is the child of two independent parents, can actually introduce a relationship between those parents (where one did not exist before).
Finally the suggested reading. I liked that it built on the week 2 reading on understanding causality for a factor we can't control experimentally e.g. how BMI relates to mortality. We cannot directly control BMI, we can control a test subjects’s diet and exercise though.
Well in this week is about how some believe we shouldn't measuring the effect of uncontrollable factors. These are more consequential than BMI - like race and gender. Which to me intuitively feel clearly to have a causal effect on certain outcomes. https://ftp.cs.ucla.edu/pub/stat_ser/r483-reprint.pdf
But real academics believe if you can’t alter it in an experiment, it shouldn’t be measured? The paper goes on to outline why there is value in measuring causal effects of things that we cannot directly manipulate and it was an interesting read. Nice build from last week.
