News for hackers. Stuff that matters.

A comment on Marshmallow Test Fails to Replicate. Parent comment

Re:Marshmallow Test Fails to Replicate by astrange on Friday August 02, 2024 @05:51PM (#41142976)
> But when they controlled for household income, they found most of the correlation disappeared.

Controlling for things is mostly bad statistics, although of course all social science is bad statistics.

Confounding variables are bad controls more often than they're good ones, so controlling for them introduces collider bias. Also, finding a result and then controlling for something is a multiple comparison fallacy.

The correct thing to do is to have a theory of causation and then design a study that's capable of detecting it, not the other way round.

[ Parent | on HN ]
Re:Marshmallow Test Fails to Replicate by rahimnathwani on Friday August 02, 2024 @06:03PM (#41143056)
In this case household income is the parents' household income, so it can't be affected by the child's (future) academic achievement.

Can controlling for household income introduce collider bias?

(Sorry I know the words you're using, and a few years ago I started reading Pearl's book, but I did not finish it and do not have a strong grasp of the concepts.)

[ Parent | on HN ]