Quote:
Originally Posted by erimir
And do you think that there's reason to think that primary performance should be interpreted differently depending on which one you're interested in?
|
This isn't merely about interpretation. The type of model you might build, and your criteria for choosing between models, will be different.
Take the
538 primary model. It's in the business of prediction. The demographic component doesn't even try to do any kind of inference. It just runs up to 360 regressions to cover all combinations of predictor variables, then goes with a weighted average of the lot based purely on how well each regression fits. It doesn't attempt, for example, to determine which of the four different race specifications is actually a good representation of voting behaviour. This is fine for a predictive model, but it's just nonsense for inference, i.e. understanding what's actually going on in a complex world.
There's a whole world of discussion and theory around this stuff. When you say things like "it's not used in predictive models, therefore there must be no demonstrable association" you're making a massive logic error.
Quote:
|
why are you so focused on that one interpretation as if it's the only thing I put forward?
|
Gee, maybe because it's the
one you mentioned when you brought this up two and a half months after chunks' post. You know, the same interpretation you ascribed to him in
your follow-up post as if it's established fact.
If I wanted to nitpick I would have plenty more material.