And thus, the Huffington Post predicted a 98% probability for Hillary Clinton to be the next President of the United States. Amen… Let’s tease them a little bit, shall we?
My Bayesian friends, I understand playing with your priors is a very joyful activity but you see, it leads to blindness. It allows you to believe, let me cap & bold this one, BELIEVE that Hillary’s chances to be the next President of United States were 98%! No wonder that betting sites favored heavily Hillary’s side days before the election! I mean 98%! Who wouldn’t put some money there. Right?
But you know, a 98% probability coming from a Bayesian means very little unless, of course, they do some math pirouette to guarantee that the probability has frequentist properties, but then, if they do that, why bother going Bayesian in the first place?
If a frequentist tells you there is 98% probability for an event to happen he/she means that 98 out of 100 times where you find yourself in a situation like where the event is taking place the event will occur. Now, if a Bayesian tells you there is 98% probability he/she means that this is his/her degree of believe (wot?) on the event to happen… Amen again.
In other words, Bayesian results are as credible as the beliefs of the Bayesian statistician making the calculations, now we can understand why they calculate credible intervals instead confidence ones.
If we check on the Huffpo methodology we can read:
Many Bayesian models ― including the Pollster averaging model as it’s implemented for our charts ― use “uninformed” priors that don’t affect the model or provide any background information.
However, we do use information from previous elections in these priors to make predictions in our presidential model.
Ba dum tsssss
Much has been written on the pros and cons of going Bayesian and how evil Frequentists are, but this amazing Bayesian result from Huffpo was just too good to let go as a beautiful example of how blind you can go when playing with your priors.