Just for amusement, Bradley was 48-40 over Bush in a national poll by LAT in June 1999 whereas Gore was down 49-44. Bradley would have won! Unfortunately for anyone who tried using that metric to pick a candidate, those numbers moved wildly and were all over the place for the duration of the Bradley/Gore race, without much of a consisten…
Just for amusement, Bradley was 48-40 over Bush in a national poll by LAT in June 1999 whereas Gore was down 49-44. Bradley would have won! Unfortunately for anyone who tried using that metric to pick a candidate, those numbers moved wildly and were all over the place for the duration of the Bradley/Gore race, without much of a consistent pattern. But who knows! If you truly think such numbers are predictive, Bradley did better, so maybe it's all the Dems fault for voting for Gore in the primary, lol. (Of course we actually have no idea about the counterfactual and I don't think anyone has seriously put forward that argument.)
The question is whether a busted and broken ruler is better than none whatsoever. Look, I'm not arguing that people shouldn't vote for whoever they want to vote for regardless of actual impact. What if Biden's already lost? Should his supporters just stay home or vote for Stein and West? Of course not. If they want to throw their votes away they should just do what they want.
Unlike what you seem to believe aggregated polling is in fact more accurate than an
individual poll. I have no idea what Bradley v Gore looked like but if the bulk of polling suggested Sanders was a better candidate than Clinton is that somehow worthless? I don't think so. Again, maybe it just means that Sanders loses by a smaller margin to Trump than Clinton but when the bulk of polling shows consistent movement in one direction I am inclined to view it as more trustworthy than a single poll.
Of course aggregated polling is more accurate, I was just finding stuff quickly but I can tell you from memory that the aggregated polling would tell you the same thing. Feel free to look up the aggregated polling for any election this far out in hypothetical H2H polls and you will find the same point I keep making.
Polling can be very solid at predicting a week or two out, and the best way to analyze is to aggregate. No one is disputing that. But literally no one in the polling industry or political science research would tell you that they're predictive this far out. Because this has been studied, and there is absolute consensus: they are not predicative this far out. Sorry!
1. What point is that? That high variance makes aggregates "murkier"? That's your choice of words, not mine.
Also, what do you mean by variance in this case? As related to sample size? That's only meaningful if you're comparing polls with vastly different sample sizes. Even then why not use something like coefficient of variation?
2. Did you even read what I wrote? I am not arguing that the polls now accurately predict the final margins in November. I am arguing that if the polling right now consistently shows black and Hispanic voters defecting to Trump then I would be willing to wager that there is _some_ movement of black and Hispanic voters to Trump even if the exact numbers aren't what the polling suggests.
I did read what you wrote, I've tried to be responsive in detail. And yes if the aggregate shows extremely high variance over time (aggregate on March 15 is radically different than aggregate on April 15) that is murkier, obviously, than if the aggregate is highly consistent over time. That is what I mean by variance. You are vastly overstating the clarity of the Sanders v. Clinton figures even it was predictive; all of this was hotly contested at the time and hard to draw firm conclusions IMO, and I did spend quite a bit of time with the numbers back then.
But listen, I gotta bow out, man. I clearly haven't been clear despite my best efforts and we're going in circles.
I think "electability" is way harder to predict than you do between the top two candidates in a Democratic party. I'm not saying it's impossible to make such predictions or make arguments about it (which is maybe what you think I'm saying?), even if I think a lot of motivated reasoning goes into such attempts. Just think these are contentious arguments, not straightforward reads of the data, and credence on such a prediction shouldn't get too high.
To the point: I think such a prediction is in zero way analogous to a strategic decision to vote for the Dem in a 2-party general election because in that situation your credence that the eventual winner will be the Dem or GOP nominee is near 100%.
That's my position, which I don't think is all that controversial. I think you believe that voters should make these "electability" predictions with the tools available when making their choice in primaries (I'd say people should generally instead simply vote for who they think is best candidate), and I think you have higher confidence than me in how accurate these predictions will be. So that's the rub. That's the disagreement. I'm satisfied we've come as far as we can come.
What exactly do you think the definition of variance is? It has a specific meaning in statistics that I'm not sure you are clear on.
I would also advise you go read up on Nate Silver to get a layman's perspective on what 538 was doing because "the aggregate showing high variance in March versus April" doesn't make sense to me.
Look, at a very simple level if most of the polls show Sanders doing better than Clinton that is actually a pretty good reason to believe that Sanders would do better than Clinton--even if he didn't end up winning. At a similar level if the polls uniformly show a bunch of Hispanic and black voters switching to Trump it is probably happening. What's "murky" is the magnitude of that shift but I would bet dollars to donuts that it's happening at a measurable level.
Finally, what's more important: that a D or R will win (100%) or that the breakdown is R 70% versus D 30%?
I simply mean variance in aggregated results over time, not sample variance. I thought this was obvious from the context but glad to clear it up. I am perfectly aware of how Silver's model works. If Silver's model was static over time -- say it hovered close to the same 50-45 result week after week, month after month-- I would give it more credence in terms of its long-range predictive value than if the aggregate results were swinging wildly week to week or month to month. (By aggregate I mean all the polls conducted within the time frame used, though since you want to get in to the weeds, I realize that past results are baked into Silver's modeling as well.) Silver himself has frequently alluded to this, FWIW, but the point is banal.
We disagree about the predictive power of early H2H polling, which was your original argument. FWIW my position is the consensus in the polling field. Could you pick out crosstabs to make predictions that would have a higher credence? Sure. I still don't think it would be dispositive. There were extremely sophisticated, data-based arguments about why Hillary was the stronger candidate at the time. There are ALWAYS sophisticated, data-based arguments making predictions of this kind. Sometimes they're on to something, sometimes they're the Emerging Democratic Majority.
I am responding to clarify the point of confusion, but otherwise I hope we can say good night.
You think Sanders being the stronger candidate was clear based on empirical data during the 2016 primaries, so clear that Democrats who voted for Hillary despite this clear evidence would be hypocritical to advise voting for Hillary, like her or not, in the general election because that was the best instrumental way to stop Trump.
I disagree. Last word is yours, but FWIW I've quickly perused your previous comments and I believe I've responded to everything in good faith and that my argument is as clear as I can hope to make it. I hope you feel the same.
"I simply mean variance in aggregated results over time, not sample variance. I thought this was obvious from the context but glad to clear it up."
Ok, so why exactly do you think that's relevant to the final predictions his model turns out?
"We disagree about the predictive power of early H2H polling"
Uh, no.
Here we go again.
A bunch of polls show black and Hispanic voters shifting to Trump right now.
What's my stance?
I think there probably is some shift in black and Hispanic voters to Trump. The magnitude may not match what the polling says but I would bet that it's happening.
Note the clause: "happening". I have no idea whether or not it's going to continue. All it's doing is describing current events, not predicting future behavior.
I don't think it was clear that Sanders was the stronger candidate. But in an imperfect world you rely on the tools that you have and the measurements that you can make. The guy at Astral Codex describes a logical fallacy called "isolated calls for rigor". Polling isn't great but it's all we have and pooh poohing the polls when you don't like the answer while turning around and cheering for them when they tell you something comfortable is just intellectually dishonest.
Finally, let's say that the Democrats nominated the Grand Cyclops of the Klu Klux Klan. Would you say that they had some responsibility if they lost the general election?
I think you could have an extreme outlier situation like that, yes--tho the level of extremity required to shift the credence that much is so high that what you're talking about it's gonna be a hypothetical that's not gonna happen. I would again point out that the closest we've come in a presidential primary was Trump, widely regarded as heinously unelectable (backed by the hypothetical H2H polls at the time).
I disagree that the choice between Sanders and Hillary represented anything remotely like the choice between a KKK candidate and a regular old Dem. Not even close.
You were originally explicitly talking about the H2H results between Sanders and Hillary so that's what I was responding to.
re: current trends of Black and Hispanic voters shifting to Trump, that's interesting, I'm not saying it's not worth discussing. I don't think that can give you a definitive answer to whether Biden or Dean Phillips would be likelier to win, or whether Candidate X who polls better with Black voters might be more vulnerable with other voters, etc.
this has nothing to do with the answer I like. I would have preferred Sanders.
I am very happy to use any and all poll data. I think it's important context that H2H hypothetical matchups a year out have no predictive value--itself an analytical point backed by research.
I am continuing to respond in the sincere hope this clarifies.
I think you believe I'm saying that because predictions are hard, we should ignore all data. I'm not. I am saying that given the data we have available, a prediction about which of two highly successful primary candidates will do better in a general months out is simply nothing remotely like an election day prediction that either the R or D will win, so one might choose to act accordingly.
1) Look, how much math (specifically stats) have you taken? Because you're just spouting gibberish at this point.
If you look at Silver's model does he predict final margins? No, he just lays odds: if you roll the dice 1000 times candidate A would win X number of times and candidate B would win Y. The margins of individual polls as they fluctuate up and down are meaningless to that kind of model except in the very important sense that accurate polls are weighted more heavily as inputs. Again, go do some background research. There's plenty aimed at laymen because Silver is so popular.
2) So we're clear that I'm not suggesting that polling is absolutely predictive. The only reason I'm bringing up the black and Hispanic voter thing is because it is something that people are talking about now as it has attracted some attention.
3) Go read my other response. Early polling is held to be accurate--it's just that too far out people aren't thinking about elections, they're not really paying attention, conditions on the ground can change, etc.
But those are obviously just general principles and need to be adapted to specific elections. Again, analysts right now are poring over polling for specific questions such as who voters trust more on issues like the economy and immigration. Why? Because they have a strong suspicion that the economy and illegal immigration are going to be critical issues this November.
Just for amusement, Bradley was 48-40 over Bush in a national poll by LAT in June 1999 whereas Gore was down 49-44. Bradley would have won! Unfortunately for anyone who tried using that metric to pick a candidate, those numbers moved wildly and were all over the place for the duration of the Bradley/Gore race, without much of a consistent pattern. But who knows! If you truly think such numbers are predictive, Bradley did better, so maybe it's all the Dems fault for voting for Gore in the primary, lol. (Of course we actually have no idea about the counterfactual and I don't think anyone has seriously put forward that argument.)
The question is whether a busted and broken ruler is better than none whatsoever. Look, I'm not arguing that people shouldn't vote for whoever they want to vote for regardless of actual impact. What if Biden's already lost? Should his supporters just stay home or vote for Stein and West? Of course not. If they want to throw their votes away they should just do what they want.
Unlike what you seem to believe aggregated polling is in fact more accurate than an
individual poll. I have no idea what Bradley v Gore looked like but if the bulk of polling suggested Sanders was a better candidate than Clinton is that somehow worthless? I don't think so. Again, maybe it just means that Sanders loses by a smaller margin to Trump than Clinton but when the bulk of polling shows consistent movement in one direction I am inclined to view it as more trustworthy than a single poll.
Of course aggregated polling is more accurate, I was just finding stuff quickly but I can tell you from memory that the aggregated polling would tell you the same thing. Feel free to look up the aggregated polling for any election this far out in hypothetical H2H polls and you will find the same point I keep making.
Polling can be very solid at predicting a week or two out, and the best way to analyze is to aggregate. No one is disputing that. But literally no one in the polling industry or political science research would tell you that they're predictive this far out. Because this has been studied, and there is absolute consensus: they are not predicative this far out. Sorry!
1. What point is that? That high variance makes aggregates "murkier"? That's your choice of words, not mine.
Also, what do you mean by variance in this case? As related to sample size? That's only meaningful if you're comparing polls with vastly different sample sizes. Even then why not use something like coefficient of variation?
2. Did you even read what I wrote? I am not arguing that the polls now accurately predict the final margins in November. I am arguing that if the polling right now consistently shows black and Hispanic voters defecting to Trump then I would be willing to wager that there is _some_ movement of black and Hispanic voters to Trump even if the exact numbers aren't what the polling suggests.
I did read what you wrote, I've tried to be responsive in detail. And yes if the aggregate shows extremely high variance over time (aggregate on March 15 is radically different than aggregate on April 15) that is murkier, obviously, than if the aggregate is highly consistent over time. That is what I mean by variance. You are vastly overstating the clarity of the Sanders v. Clinton figures even it was predictive; all of this was hotly contested at the time and hard to draw firm conclusions IMO, and I did spend quite a bit of time with the numbers back then.
But listen, I gotta bow out, man. I clearly haven't been clear despite my best efforts and we're going in circles.
I think "electability" is way harder to predict than you do between the top two candidates in a Democratic party. I'm not saying it's impossible to make such predictions or make arguments about it (which is maybe what you think I'm saying?), even if I think a lot of motivated reasoning goes into such attempts. Just think these are contentious arguments, not straightforward reads of the data, and credence on such a prediction shouldn't get too high.
To the point: I think such a prediction is in zero way analogous to a strategic decision to vote for the Dem in a 2-party general election because in that situation your credence that the eventual winner will be the Dem or GOP nominee is near 100%.
That's my position, which I don't think is all that controversial. I think you believe that voters should make these "electability" predictions with the tools available when making their choice in primaries (I'd say people should generally instead simply vote for who they think is best candidate), and I think you have higher confidence than me in how accurate these predictions will be. So that's the rub. That's the disagreement. I'm satisfied we've come as far as we can come.
What exactly do you think the definition of variance is? It has a specific meaning in statistics that I'm not sure you are clear on.
I would also advise you go read up on Nate Silver to get a layman's perspective on what 538 was doing because "the aggregate showing high variance in March versus April" doesn't make sense to me.
Look, at a very simple level if most of the polls show Sanders doing better than Clinton that is actually a pretty good reason to believe that Sanders would do better than Clinton--even if he didn't end up winning. At a similar level if the polls uniformly show a bunch of Hispanic and black voters switching to Trump it is probably happening. What's "murky" is the magnitude of that shift but I would bet dollars to donuts that it's happening at a measurable level.
Finally, what's more important: that a D or R will win (100%) or that the breakdown is R 70% versus D 30%?
I simply mean variance in aggregated results over time, not sample variance. I thought this was obvious from the context but glad to clear it up. I am perfectly aware of how Silver's model works. If Silver's model was static over time -- say it hovered close to the same 50-45 result week after week, month after month-- I would give it more credence in terms of its long-range predictive value than if the aggregate results were swinging wildly week to week or month to month. (By aggregate I mean all the polls conducted within the time frame used, though since you want to get in to the weeds, I realize that past results are baked into Silver's modeling as well.) Silver himself has frequently alluded to this, FWIW, but the point is banal.
We disagree about the predictive power of early H2H polling, which was your original argument. FWIW my position is the consensus in the polling field. Could you pick out crosstabs to make predictions that would have a higher credence? Sure. I still don't think it would be dispositive. There were extremely sophisticated, data-based arguments about why Hillary was the stronger candidate at the time. There are ALWAYS sophisticated, data-based arguments making predictions of this kind. Sometimes they're on to something, sometimes they're the Emerging Democratic Majority.
I am responding to clarify the point of confusion, but otherwise I hope we can say good night.
You think Sanders being the stronger candidate was clear based on empirical data during the 2016 primaries, so clear that Democrats who voted for Hillary despite this clear evidence would be hypocritical to advise voting for Hillary, like her or not, in the general election because that was the best instrumental way to stop Trump.
I disagree. Last word is yours, but FWIW I've quickly perused your previous comments and I believe I've responded to everything in good faith and that my argument is as clear as I can hope to make it. I hope you feel the same.
"I simply mean variance in aggregated results over time, not sample variance. I thought this was obvious from the context but glad to clear it up."
Ok, so why exactly do you think that's relevant to the final predictions his model turns out?
"We disagree about the predictive power of early H2H polling"
Uh, no.
Here we go again.
A bunch of polls show black and Hispanic voters shifting to Trump right now.
What's my stance?
I think there probably is some shift in black and Hispanic voters to Trump. The magnitude may not match what the polling says but I would bet that it's happening.
Note the clause: "happening". I have no idea whether or not it's going to continue. All it's doing is describing current events, not predicting future behavior.
I don't think it was clear that Sanders was the stronger candidate. But in an imperfect world you rely on the tools that you have and the measurements that you can make. The guy at Astral Codex describes a logical fallacy called "isolated calls for rigor". Polling isn't great but it's all we have and pooh poohing the polls when you don't like the answer while turning around and cheering for them when they tell you something comfortable is just intellectually dishonest.
Finally, let's say that the Democrats nominated the Grand Cyclops of the Klu Klux Klan. Would you say that they had some responsibility if they lost the general election?
I think you could have an extreme outlier situation like that, yes--tho the level of extremity required to shift the credence that much is so high that what you're talking about it's gonna be a hypothetical that's not gonna happen. I would again point out that the closest we've come in a presidential primary was Trump, widely regarded as heinously unelectable (backed by the hypothetical H2H polls at the time).
I disagree that the choice between Sanders and Hillary represented anything remotely like the choice between a KKK candidate and a regular old Dem. Not even close.
You were originally explicitly talking about the H2H results between Sanders and Hillary so that's what I was responding to.
re: current trends of Black and Hispanic voters shifting to Trump, that's interesting, I'm not saying it's not worth discussing. I don't think that can give you a definitive answer to whether Biden or Dean Phillips would be likelier to win, or whether Candidate X who polls better with Black voters might be more vulnerable with other voters, etc.
this has nothing to do with the answer I like. I would have preferred Sanders.
I am very happy to use any and all poll data. I think it's important context that H2H hypothetical matchups a year out have no predictive value--itself an analytical point backed by research.
I am continuing to respond in the sincere hope this clarifies.
I think you believe I'm saying that because predictions are hard, we should ignore all data. I'm not. I am saying that given the data we have available, a prediction about which of two highly successful primary candidates will do better in a general months out is simply nothing remotely like an election day prediction that either the R or D will win, so one might choose to act accordingly.
1) Look, how much math (specifically stats) have you taken? Because you're just spouting gibberish at this point.
If you look at Silver's model does he predict final margins? No, he just lays odds: if you roll the dice 1000 times candidate A would win X number of times and candidate B would win Y. The margins of individual polls as they fluctuate up and down are meaningless to that kind of model except in the very important sense that accurate polls are weighted more heavily as inputs. Again, go do some background research. There's plenty aimed at laymen because Silver is so popular.
2) So we're clear that I'm not suggesting that polling is absolutely predictive. The only reason I'm bringing up the black and Hispanic voter thing is because it is something that people are talking about now as it has attracted some attention.
3) Go read my other response. Early polling is held to be accurate--it's just that too far out people aren't thinking about elections, they're not really paying attention, conditions on the ground can change, etc.
But those are obviously just general principles and need to be adapted to specific elections. Again, analysts right now are poring over polling for specific questions such as who voters trust more on issues like the economy and immigration. Why? Because they have a strong suspicion that the economy and illegal immigration are going to be critical issues this November.
lol.
"The margins of individual polls as they fluctuate up and down." This is precisely the opposite of what I said above and utterly irrelevant.
Everything you've said in the first paragraph is completely obvious and completely irrelevant. You've misread, I'm afraid.
With all respect, this is a waste of time.