I don't remember the exact number of states, but I distinctly remember a time between New Hampshire and Nevada where they were predicting a Bernie win in a very large majority of states. My point being that they were saying he was going to absolutely crush it. Sorry, should have looked up the exact number.
The confusion is that he was predicted relatively narrow victories in a lot of those states. Saying he's a narrow favourite in 45 (or however many) states is very different from predicting that he's likely to sweep 45 states. If I play a game where I roll a die and win with a roll of 3 or higher, I'm a narrow favourite to win each roll. But if I roll it 45 times I'm extremely unlikely to win every one.
Oh ok, I guess I used the word sweep incorrectly then. I didn't mean he would massively win each one. But overall, the point I'm trying to make is Bernie was predicted to pick up a massive delegate lead over Biden, and the only worry was that he might not have been able to reach 50 percent needed to avoid a contested convention. My point is that this is massively different from the actual outcome.
That's fair enough. On the other hand I remember discussing with a lot of people who thought Pete, Amy or Kamala had some kind of chance, and 538's analysis was extremely helpful in understanding that it was really between Biden and Bernie. It was also thanks to 538 that I already knew that SC was going to be critical and that people high on the Nevada victory treating it as unimportant were mistaken.
Of course 538 didn't predict the orchestrated drop outs, but that's not really the purpose of a model like that, and when they hapened my reaction wasn't "Why didn't Nate Silver tell me about this?!"
Oh for sure, I'm not trying to say the model should have predicted the drop outs. What I'm trying to say is that a model that only uses polls and demographics data to predict the outcome to an election several weeks or more in advance is fundamentally useless precisely because it can't account for things like orchestrated drop outs or buttery males. And also because any prediction made too far away from the date of the election is just unfalsifiable.
For example, what if the model had shown Bernie falling to a 5 percent chance of victory like a year out from the start of voting. What if this then rose back up to something more in line with contemporary polling data by the time voting actually started? Since there is no actual vote a year out from the election, there is no way to actually know if any of that spike corresponded to reality. This line of reasoning has also made me skeptical of polls conducted far out from the date of an election as well. Who's to say primary polls in January of 2019 were actually accurate representations of the electorate? If they were cooked, there would be no way to know because the actual election would still be a year away.
Yes, that I agree with. If you're looking at the 538 model 6 months out you should look at its previous predictions 6 months out to see how confident to be in its predictions (or maybe even look at its worst prediction at 6 months out or less for each election).
But basically yes, far enough from an election its silly to pay much attention to polls or models. But we're actually getting pretty close to the presidential election now!
I don't remember the exact number of states, but I distinctly remember a time between New Hampshire and Nevada where they were predicting a Bernie win in a very large majority of states. My point being that they were saying he was going to absolutely crush it. Sorry, should have looked up the exact number.
The confusion is that he was predicted relatively narrow victories in a lot of those states. Saying he's a narrow favourite in 45 (or however many) states is very different from predicting that he's likely to sweep 45 states. If I play a game where I roll a die and win with a roll of 3 or higher, I'm a narrow favourite to win each roll. But if I roll it 45 times I'm extremely unlikely to win every one.
Oh ok, I guess I used the word sweep incorrectly then. I didn't mean he would massively win each one. But overall, the point I'm trying to make is Bernie was predicted to pick up a massive delegate lead over Biden, and the only worry was that he might not have been able to reach 50 percent needed to avoid a contested convention. My point is that this is massively different from the actual outcome.
That's fair enough. On the other hand I remember discussing with a lot of people who thought Pete, Amy or Kamala had some kind of chance, and 538's analysis was extremely helpful in understanding that it was really between Biden and Bernie. It was also thanks to 538 that I already knew that SC was going to be critical and that people high on the Nevada victory treating it as unimportant were mistaken.
Of course 538 didn't predict the orchestrated drop outs, but that's not really the purpose of a model like that, and when they hapened my reaction wasn't "Why didn't Nate Silver tell me about this?!"
Oh for sure, I'm not trying to say the model should have predicted the drop outs. What I'm trying to say is that a model that only uses polls and demographics data to predict the outcome to an election several weeks or more in advance is fundamentally useless precisely because it can't account for things like orchestrated drop outs or buttery males. And also because any prediction made too far away from the date of the election is just unfalsifiable.
For example, what if the model had shown Bernie falling to a 5 percent chance of victory like a year out from the start of voting. What if this then rose back up to something more in line with contemporary polling data by the time voting actually started? Since there is no actual vote a year out from the election, there is no way to actually know if any of that spike corresponded to reality. This line of reasoning has also made me skeptical of polls conducted far out from the date of an election as well. Who's to say primary polls in January of 2019 were actually accurate representations of the electorate? If they were cooked, there would be no way to know because the actual election would still be a year away.
Yes, that I agree with. If you're looking at the 538 model 6 months out you should look at its previous predictions 6 months out to see how confident to be in its predictions (or maybe even look at its worst prediction at 6 months out or less for each election).
But basically yes, far enough from an election its silly to pay much attention to polls or models. But we're actually getting pretty close to the presidential election now!