It is fun to look at polls, and using such data, decide which candidate will win which state, and ultimately, which candidate will win the electoral college. A lot of people and organizations do that, and for this reason, I don’t. I do not have access to polls that no one else sees. Were I to use polling data to directly predict outcomes per state, I’d use a method like that used by FiveThirtyEight, and probably come up with similar results. How boring. It would be a waste of my time to try to replicate the excellent work done by Nate Silver and his team.
Back during the Democratic Primaries, I decided that I wanted to get a handle on which candidate was likely to win, fairly early on. The polling based estimates were inadequate because most states simply didn’t have polling data that early in the process. So, I invented an alternative method, which made certain estimates of how voters with different ethnic identities would vote. That method accurately predicted several primary outcomes, outperforming the poll based methods such as those used by FiveThirtyEight.
After a while, enough primaries had been carried out that I could switch methods slightly. Using the same exact model, but primed with the results of prior primaries (that year) rather than my estimates of voter behavior, I used the ethnic distribution data for each state to predict the outcome of upcoming primary contests.
Once again, my method was very accurate, and once again, it out performed the polling based methods.
So, recently, I’ve tried to apply a similar method to estimating the electoral outcome for this year’s presidential race. But, it is impossible to use the same exact method because the entire thing happens all on one day. I can’t use the election results from a handful of states to estimate the likely future outcomes in other states.
I recognize that polling data is very limited on a national level. Things happen during an election season that probably change people’s likely voting behavior, especially among independents. Solid states are rarely polled, and small states, swing or not, are rarely polled. Many polls are of low quality. Right now, for instance, fewer than half of the states have polls that were a) taken fully after the final POTUS debate and b) have an A- or better rating from FiveThirtyEight. If I allow the use of B and occasional C ratings for recent polls, and allow a few polls to include periods of time prior to the last POTUS debate, but only in states that are very strongly in favor of one candidate or the other (and thus likely to not move anyway), I can find 32 states that have sort of usable polling data. Interestingly, states with some of the more controversial changes happening, like Utah and Iowa, are not adequately polled.
In order to apply a model like the one I used in the Primaries to the current election, I used the 32 states for which there was somewhat acceptable recent polling data to inform the model (to calculate the regression coefficients) in order to then, separately, predict the likely voting behavior (Trump vs. Clinton) in all of the states.
Before I show you the map, however, I need to discuss something else.
About a week ago the press, especially the somewhat more left leaning press, and various commenters, seeing much reaction to a series of events beginning with the NYT release of Trump’s tax return and ending with the final POTUS debate, events which sandwiched the sexual assault tapes and accusations, collectively decided that a huge gap between Clinton and Trump was rapidly opening up and the race would end with a double digit spread, an electoral rout, and a big party.
Soon after, I pointed out that this may not be correct. That polling data seemed to show, rather, that there was an expansion of the difference between the two candidates followed by a re-closing of the gap, with Clinton still leading but by about as much as before this temporary shift. To this I added a concern. If too many people assumed that the race was over and in the double digit range, perhaps there could be a GOTV backlash effect, or a funding effect, that would shift things to within shooting distance for Trump.
I was not alone in thinking this, and I was probably right. The GOP sunk, via pacs, 25 million dollars into Senate races in response to the Democrats shifting from the national race to the Senate, which was followed by the Democrats shifting back to the national race in certain states, presumably recognizing that the polls were artificially spread. Indeed, some who criticized (arguing mainly from incredulity and good wishes) my admonition noted, correctly, that some of that narrowing was because a bunch of right-leaning polls had come out all at once. This is true, but it ignores that a bunch of left-leaning polls had made the formation of the Great Gap of GOP Defeat look a lot bigger than it ever really was.
I say all this as part one of my preparation for what I’m going to tell you below, which is not the news you want to hear. Part two is some logic I’d like to bludgeon you with.
Consider these points:
1) True Trump supporters could give a rat’s ass about sexual assault, poor debate performance, or tax forms. Donald Trump was correct when he said, weeks ago, now forgotten, that he could gun someone down on the streets of Manhattan and he would not lose support form his base. These people did not abandon him when he was heard to talk about sexual assault. If anything, they were energized by it. And, I’m talking about something just shy of 40% of the voters. We live in a barely civilized asshole country.
2) Please tell me exactly which Hillary Clinton supporters, who were going to vote for Clinton over Trump all along, are NOW going to pick Clinton (if polled or on voting day) that change from not being Clinton supporters to being Clinton supporters? In other words (this is a somewhat subtle point) which people who hated Trump became True Haters of Trump after the sexual assault thing? Almost none. They were already there.
3) The third category of people, the undecideds (who are only lying about being undecided, in most cases) and the so-called “reasonable Republicans” (of which there are very, very few), who could conceivably shift from Trump to Clinton are going to divide their voting activities between Johnson, a write in (as they are being advised by Republican leaders in some cases) or simply staying home.
In other words, over the last few weeks, no source has emerged that hands Secretary Clinton more electoral votes than she probably had about a month ago, and Trump is not going to have any, or at least not many, electoral votes go away.
Those observations (part one) and that logic (part two) cause me to be utterly unsurprised to find out that an analysis of the electoral map I did on October 16th and one I did today do not show Clinton pulling farther ahead. In fact, the two analyses have Clinton being less far ahead than Trump now than ten days ago. The difference is in Ohio (shifting from Clinton to Trump) which is almost certainly going to happen, and North Carolina (which shifted from Clinton to Trump in this analysis) which seems much less likely to happen, and Arizona shifting from Clinton (that was probably wishful thinking) to Trump.
The point here is this, plain and simple. An analysis using a technique that has worked very well for me in the past shows that the difference between that moment of Maximal Clintonosity and today is plus or minus a couple of state. In other words, not different. Maybe a little worse. Really, about the same.
Here’s the current map:
Obviously, I will be watching for more data over the next few days. I assume there will be a spate of polls as we approach November 8th (the day Democrats vote. Republicans vote on the 28th of November). If so, then there will be convergence between my method of calibration and my method of calculation, and the model will consume itself by the tail and become very accurate at the same time.
But between now and then, perhaps that very small number of polls that are both recent and high quality will grow a bit more and I can do this again and resolve those closer states.
By the way, the “swing states” according to my model, the states where things are close, are Ohio, North Carolina, Arizona, and Georgia of those now in the Trump column. Those are indeed swing states. Numerically, the close states that are in the Clinton column are Virginia, New Hampshire, and Pennsylvania.
from ScienceBlogs http://ift.tt/2eGNkKV
It is fun to look at polls, and using such data, decide which candidate will win which state, and ultimately, which candidate will win the electoral college. A lot of people and organizations do that, and for this reason, I don’t. I do not have access to polls that no one else sees. Were I to use polling data to directly predict outcomes per state, I’d use a method like that used by FiveThirtyEight, and probably come up with similar results. How boring. It would be a waste of my time to try to replicate the excellent work done by Nate Silver and his team.
Back during the Democratic Primaries, I decided that I wanted to get a handle on which candidate was likely to win, fairly early on. The polling based estimates were inadequate because most states simply didn’t have polling data that early in the process. So, I invented an alternative method, which made certain estimates of how voters with different ethnic identities would vote. That method accurately predicted several primary outcomes, outperforming the poll based methods such as those used by FiveThirtyEight.
After a while, enough primaries had been carried out that I could switch methods slightly. Using the same exact model, but primed with the results of prior primaries (that year) rather than my estimates of voter behavior, I used the ethnic distribution data for each state to predict the outcome of upcoming primary contests.
Once again, my method was very accurate, and once again, it out performed the polling based methods.
So, recently, I’ve tried to apply a similar method to estimating the electoral outcome for this year’s presidential race. But, it is impossible to use the same exact method because the entire thing happens all on one day. I can’t use the election results from a handful of states to estimate the likely future outcomes in other states.
I recognize that polling data is very limited on a national level. Things happen during an election season that probably change people’s likely voting behavior, especially among independents. Solid states are rarely polled, and small states, swing or not, are rarely polled. Many polls are of low quality. Right now, for instance, fewer than half of the states have polls that were a) taken fully after the final POTUS debate and b) have an A- or better rating from FiveThirtyEight. If I allow the use of B and occasional C ratings for recent polls, and allow a few polls to include periods of time prior to the last POTUS debate, but only in states that are very strongly in favor of one candidate or the other (and thus likely to not move anyway), I can find 32 states that have sort of usable polling data. Interestingly, states with some of the more controversial changes happening, like Utah and Iowa, are not adequately polled.
In order to apply a model like the one I used in the Primaries to the current election, I used the 32 states for which there was somewhat acceptable recent polling data to inform the model (to calculate the regression coefficients) in order to then, separately, predict the likely voting behavior (Trump vs. Clinton) in all of the states.
Before I show you the map, however, I need to discuss something else.
About a week ago the press, especially the somewhat more left leaning press, and various commenters, seeing much reaction to a series of events beginning with the NYT release of Trump’s tax return and ending with the final POTUS debate, events which sandwiched the sexual assault tapes and accusations, collectively decided that a huge gap between Clinton and Trump was rapidly opening up and the race would end with a double digit spread, an electoral rout, and a big party.
Soon after, I pointed out that this may not be correct. That polling data seemed to show, rather, that there was an expansion of the difference between the two candidates followed by a re-closing of the gap, with Clinton still leading but by about as much as before this temporary shift. To this I added a concern. If too many people assumed that the race was over and in the double digit range, perhaps there could be a GOTV backlash effect, or a funding effect, that would shift things to within shooting distance for Trump.
I was not alone in thinking this, and I was probably right. The GOP sunk, via pacs, 25 million dollars into Senate races in response to the Democrats shifting from the national race to the Senate, which was followed by the Democrats shifting back to the national race in certain states, presumably recognizing that the polls were artificially spread. Indeed, some who criticized (arguing mainly from incredulity and good wishes) my admonition noted, correctly, that some of that narrowing was because a bunch of right-leaning polls had come out all at once. This is true, but it ignores that a bunch of left-leaning polls had made the formation of the Great Gap of GOP Defeat look a lot bigger than it ever really was.
I say all this as part one of my preparation for what I’m going to tell you below, which is not the news you want to hear. Part two is some logic I’d like to bludgeon you with.
Consider these points:
1) True Trump supporters could give a rat’s ass about sexual assault, poor debate performance, or tax forms. Donald Trump was correct when he said, weeks ago, now forgotten, that he could gun someone down on the streets of Manhattan and he would not lose support form his base. These people did not abandon him when he was heard to talk about sexual assault. If anything, they were energized by it. And, I’m talking about something just shy of 40% of the voters. We live in a barely civilized asshole country.
2) Please tell me exactly which Hillary Clinton supporters, who were going to vote for Clinton over Trump all along, are NOW going to pick Clinton (if polled or on voting day) that change from not being Clinton supporters to being Clinton supporters? In other words (this is a somewhat subtle point) which people who hated Trump became True Haters of Trump after the sexual assault thing? Almost none. They were already there.
3) The third category of people, the undecideds (who are only lying about being undecided, in most cases) and the so-called “reasonable Republicans” (of which there are very, very few), who could conceivably shift from Trump to Clinton are going to divide their voting activities between Johnson, a write in (as they are being advised by Republican leaders in some cases) or simply staying home.
In other words, over the last few weeks, no source has emerged that hands Secretary Clinton more electoral votes than she probably had about a month ago, and Trump is not going to have any, or at least not many, electoral votes go away.
Those observations (part one) and that logic (part two) cause me to be utterly unsurprised to find out that an analysis of the electoral map I did on October 16th and one I did today do not show Clinton pulling farther ahead. In fact, the two analyses have Clinton being less far ahead than Trump now than ten days ago. The difference is in Ohio (shifting from Clinton to Trump) which is almost certainly going to happen, and North Carolina (which shifted from Clinton to Trump in this analysis) which seems much less likely to happen, and Arizona shifting from Clinton (that was probably wishful thinking) to Trump.
The point here is this, plain and simple. An analysis using a technique that has worked very well for me in the past shows that the difference between that moment of Maximal Clintonosity and today is plus or minus a couple of state. In other words, not different. Maybe a little worse. Really, about the same.
Here’s the current map:
Obviously, I will be watching for more data over the next few days. I assume there will be a spate of polls as we approach November 8th (the day Democrats vote. Republicans vote on the 28th of November). If so, then there will be convergence between my method of calibration and my method of calculation, and the model will consume itself by the tail and become very accurate at the same time.
But between now and then, perhaps that very small number of polls that are both recent and high quality will grow a bit more and I can do this again and resolve those closer states.
By the way, the “swing states” according to my model, the states where things are close, are Ohio, North Carolina, Arizona, and Georgia of those now in the Trump column. Those are indeed swing states. Numerically, the close states that are in the Clinton column are Virginia, New Hampshire, and Pennsylvania.
from ScienceBlogs http://ift.tt/2eGNkKV
Aucun commentaire:
Enregistrer un commentaire