LogFAQs > #939683181

HeroDelTiempo17 posted...

I work at a medical university that has been doing data modeling just for the cases in the metroplex. As far as I can tell, their model is very accurate now. While they didn't retroactively model what it would look like if they shut down earlier, they could model what the case numbers would look like if we had shut down any later than we did. And the numbers for one and two week delayed are absolutely ridiculously larger. Delaying 1 week is 4x the amount of cases. Delaying 2 weeks is 18x higher.

And this is in an area that did not have an early spike, mind you. It absolutely stands to reason that if the places hit hardest had shut down even a little bit sooner it would have made an enormous difference.

The problem with retrospective modeling is you can always find one that fits the data after the fact. If I make 100 models, I'll probably get one that comes close to the data after the fact. It is not the same as predicting it beforehand, because no one will take me seriously if I present 100 models and then brag that one of them was right. But with retrospective modeling, you can just not publish the models that were wrong and it looks like you "predicted" things.

As an Econ major I'm well aware of this problem which absolutely plagues the field. The hard sciences generally give very little credibility to retrospective modeling but in Econ we have no choice as it's hard to design macroeconomic experiments due to the scale and number of variables.