Previous projections were derived entirely from polls. This projection attempts to take a 'polls-only' base case and layer on some adjustments for hard-to-poll and late breaking factors. (In this case most of the adjustments are in Harris's favor apart from potential systematic polling miss.) Finally I've tried to factor in what a systemic polling miss might do to the projection.
First, polls-only.
Harris has seen some notable improvement in the Blue Wall states. She was at one stage behind in all 3, and now the aggregated recent polls indicate 65%, 75%, and 85% chance of winning Pennsylvania, Wisconsin, and Michigan, respectively. Since she is also very likely to win Nebraska-2, so winning all of these 3 would be sufficient to barely win the Electoral College.
If she happens to win either of Georgia (25% chance) or North Carolina (20%) then she can afford to lose either WI or MI. Likewise, Arizona (15%) could 'replace' WI, or AZ + Nevada (50%) could replace MI. Making up for a loss in PA requires winning at least 2 long shot states and is not very likely.
Flip the script to Trump's perspective, and he faces equally difficult paths if he loses PA. He would need to take one of the other Blue Wall states and individually they are at best 25% likely to go his way.
Grinding through the numbers, Harris has the edge in "polls-only" at 58%. This is a big jump from sub-40% only a week ago, however this race is ultra-tight and a 1% swing in the polls translates to a really big swing in the final probabilities.
To polls-only, I'm making 3 adjustments for late developments that I expect to be material but won't show up in any of the polls.
Adjustment 1: Get out the vote Harris appears to have the superior GOTV campaign. Opinions vary widely on the effectiveness of such campaigns, but I'm assuming:
a Good GOTV campaign will boost your side's turnout by 1.5%;
a Middling campaign is worth a 0.75% boost;
a Poor campaign will boost your side's turnout by only 0.25%
we can't know for certain the quality of the ground games, but based on consensus among reasonably objective journalists I'd suggest there is a 70% chance that Harris's GOTV campaign is Good, 25% chance Middling and 5% Poor. Trump's outsourced effort is by contrast smaller and not receiving high praise; I'm pegging it at 10% chance of actually being Good, 30% chance Middling, and 60% Poor.
Crunching the numbers the ground game works out to a 0.36% shift for Harris relative to the poll aggregates.
Adjustment 2: Insulted Puerto Ricans
Most would agree that insulting this bloc won't help Trump, but estimating the positive effect for Harris requires a little guesswork. Here's my reasoning to obtain a number:
3 potential (related) effects are possible: An otherwise-Trump voter switches to Harris; A Trump voter stays home; or an otherwise disinterested voter adds a vote for Harris.
It turns out that the pool for potential switches and 'stay-home' isn't especially large. Trump's share of the Latino vote (a proxy for the P.Rican diaspora) is only 34%.
Pennsylvania has the largest contingent of Puerto Ricans at 460,000 population. Eligible voters will be roughly 50% of this number at 230,000. Voter turnout is perhaps ~55%, so a total of 125,000 likely voters. Of these, the 34% presumed Trump voters would be between 40,000 and 45,000 in number.
We can guess at reasonable bounds for the impact on how people will vote. For sake of argument I've bracketed this at between 5% (one in 20) and 20% (one in 5) would-be Trump voters changing course.
We can also guess at similar reasonable bounds for insults being a motivating factor for those who might otherwise not vote at all. This is a somewhat bigger pool (more people in this bloc don't vote than vote for Trump), but an additional vote has only half the impact of a switched vote.
With switched votes and net new vote estimates in hand, we can estimate how much they might influence the tally of roughly 7 million votes for all of Pennsylvania. The number comes out somewhere close to 0.17%.
Repeating the analysis across each of the swing states, the weighted average estimated effect is somewhere near 0.12% in Harris's favor.
Adjustment 3: Republican Women Stealth-Voting for Harris
Judging by the trump-leaning media losing their minds over ads emphasizing that a women can keep her vote secret, the Democrats are onto something here. The potential effect is not especially easy to quantify, but here's an attempt:
Roughly 44% of women are Trump voters, if you believe 12% to be the polling gap.
Evangelical women are probably not drawn to Harris at all; at perhaps 22% of the population the remaining female Republican voters amount to ~22% of all female voters.
Stealth-voting would mainly apply to married women, and roughly half of adult women are married, leaving us with 11% of female voters.
As a share of the electorate, this target bloc amounts to around 5.6% of the total electorate.
Guessing at some bounds, I suggest that at the high end 1 in 10 women in this bloc might be convinced to stealth-vote, and 1 in 50 at the low end. Taking the crude average of these at 6%, and multiplying by the 5.6% share of all voters, we have a net effect of 0.34% in Harris's favor due to stealth voting.
Sum of 3 adjustments above: 0.81% in Harris's favor
Since the above adjustments are intended to offset what pollsters can't see, I'm fairly comfortable tweaking all the poll aggregates (essentially Harris goes up 0.4% in every poll and Trump down 0.4%) and using this as the new base level. It turns out that 0.8% is very consequential, and would shift the base probability to 90% for Harris.
Final adjustment: Systematic Poll Error
Unfortunately polls are not perfectly representative, and sometimes the aggregate bias is material. In theory, a collection of 20 independent polls each with a 3% MOE should have a very tight distribution around the combined average and a MOE less than 1%. In reality all sorts of factors (including less noble motivations such as tilting polls to manufacture a narrative, or crowding in to avoid being a statistical outlier in the post-mortem) are at play that can result in a significant one-directional miss in the aggregate. As a crude correction, I ran 7 scenarios of potential aggregate polling miss:
5% change that the polls are underestimating Trump by 3%; after adjusting for this Harris would have only a 1% chance of winning;
10% chance that polls underestimate Trump by 2%; Harris wins 7% of the time in this scenario;
20% chance polls underestimate Trump by 1%; Harris wins 47% of the time;
30% chance that the polls are correct; with no adjustment either way Harris wins 90%;
20% chance that the polls underestimate Harris by 1%; Harris wins 99%
10% chance that the polls underestimate Harris by 1%; Harris wins 100%
5% chance that the polls underestimate Harris by 1%; Harris wins 100%
Multiplying through, the final estimate is 72% for Harris, 28% for Trump.
Files
Why might you be wrong?
The race is close and the sensitivity to a small change is really high. I've had to make up a lot of estimates for phenomena with little to no precedent. (Oh wait, this involves Trump...)
Whatever the outcome, may we all find peace with it.
I like your analysis, but the problem is that it doesn't take the electoral college into play, and how these insights play in the states as weighted by EVs. I feel like a lot of forecasters here are mistaking this question for popular vote, and not electoral votes.
Your main "why might you be wrong" point should be that you're not taking state by state into account. That's what's made me flip even though I agree with your overall analysis.
Your forecast is largely realistic but might be overly optimistic about Harris’s edge due to reliance on speculative adjustments and uniform shifts across polls.
@sevedb : It clearly wasn't correct! (Apart from perhaps introducing a distribution for systematic polling error. Hard to know how to legitimately estimate that distribution - perhaps a good hobby project.)
@csfinn : Actually my model is/was entirely built around electoral college weightings, and ignored national polls entirely. A past criticism has been the assumption that the state races are independent. In retrospect, one could have modeled either way and been 'victimized' by a pretty substantial one-way polling error (underestimating DJT). To be fair, I did not go to the extreme of simulating independent errors state-by-state, but intuitively I doubt it would have altered the final conclusion in any material way.
You don't have a baseline and methodology that is independent of the current case, something that would be applicable in all elections in the past and future, you only looked at the current case and made up scientifically looking probabilities without any citations of sources that back the probabilities. The word "presumed" shows up in your rationale, the correct word should have been "unknown". There is a training button at the top of the GJO page, it's a tutorial.
Lastly, if a forecaster compliments you for being realistic but simplistic, the forecaster doesn's know, that realistic and simplistic is the hallmark of good science and desirabln. Your model was unrealistic nor simplistic.
It's fair to say, it's a model, it's a thought, it's not your identity or character. Errors are allowed and without consequences.
The big professionals got it wrong, too, by the way. No big deal.
You got duped by coordinated Democrat talking points , parroted on mainstream media, into believing “women stealth vote” was a possibility (think about this for a second), and that it was possible that Harris was being “under polled”.Re-asses your sources, because your process is pretty good.
Or a Bradley effect in which voters tell pollsters they are undecided or supportive of a non-white candidate, but then end up voting for the white opponent on election day.
Why do you think you're right? (optional)
A very eventful final week!
Previous projections were derived entirely from polls. This projection attempts to take a 'polls-only' base case and layer on some adjustments for hard-to-poll and late breaking factors. (In this case most of the adjustments are in Harris's favor apart from potential systematic polling miss.) Finally I've tried to factor in what a systemic polling miss might do to the projection.
First, polls-only.
Harris has seen some notable improvement in the Blue Wall states. She was at one stage behind in all 3, and now the aggregated recent polls indicate 65%, 75%, and 85% chance of winning Pennsylvania, Wisconsin, and Michigan, respectively. Since she is also very likely to win Nebraska-2, so winning all of these 3 would be sufficient to barely win the Electoral College.
If she happens to win either of Georgia (25% chance) or North Carolina (20%) then she can afford to lose either WI or MI. Likewise, Arizona (15%) could 'replace' WI, or AZ + Nevada (50%) could replace MI. Making up for a loss in PA requires winning at least 2 long shot states and is not very likely.
Flip the script to Trump's perspective, and he faces equally difficult paths if he loses PA. He would need to take one of the other Blue Wall states and individually they are at best 25% likely to go his way.
Grinding through the numbers, Harris has the edge in "polls-only" at 58%. This is a big jump from sub-40% only a week ago, however this race is ultra-tight and a 1% swing in the polls translates to a really big swing in the final probabilities.
To polls-only, I'm making 3 adjustments for late developments that I expect to be material but won't show up in any of the polls.
Adjustment 1: Get out the vote
Harris appears to have the superior GOTV campaign. Opinions vary widely on the effectiveness of such campaigns, but I'm assuming:
Adjustment 2: Insulted Puerto Ricans
Most would agree that insulting this bloc won't help Trump, but estimating the positive effect for Harris requires a little guesswork. Here's my reasoning to obtain a number:
Adjustment 3: Republican Women Stealth-Voting for Harris
Judging by the trump-leaning media losing their minds over ads emphasizing that a women can keep her vote secret, the Democrats are onto something here. The potential effect is not especially easy to quantify, but here's an attempt:
Sum of 3 adjustments above: 0.81% in Harris's favor
Since the above adjustments are intended to offset what pollsters can't see, I'm fairly comfortable tweaking all the poll aggregates (essentially Harris goes up 0.4% in every poll and Trump down 0.4%) and using this as the new base level. It turns out that 0.8% is very consequential, and would shift the base probability to 90% for Harris.
Final adjustment: Systematic Poll Error
Unfortunately polls are not perfectly representative, and sometimes the aggregate bias is material. In theory, a collection of 20 independent polls each with a 3% MOE should have a very tight distribution around the combined average and a MOE less than 1%. In reality all sorts of factors (including less noble motivations such as tilting polls to manufacture a narrative, or crowding in to avoid being a statistical outlier in the post-mortem) are at play that can result in a significant one-directional miss in the aggregate. As a crude correction, I ran 7 scenarios of potential aggregate polling miss:
Multiplying through, the final estimate is 72% for Harris, 28% for Trump.
Why might you be wrong?
The race is close and the sensitivity to a small change is really high. I've had to make up a lot of estimates for phenomena with little to no precedent. (Oh wait, this involves Trump...)
Whatever the outcome, may we all find peace with it.
I like your analysis, but the problem is that it doesn't take the electoral college into play, and how these insights play in the states as weighted by EVs. I feel like a lot of forecasters here are mistaking this question for popular vote, and not electoral votes.
Your main "why might you be wrong" point should be that you're not taking state by state into account. That's what's made me flip even though I agree with your overall analysis.
Your forecast is largely realistic but might be overly optimistic about Harris’s edge due to reliance on speculative adjustments and uniform shifts across polls.
Nevertheless I hope it is correct!
@csfinn : Actually my model is/was entirely built around electoral college weightings, and ignored national polls entirely. A past criticism has been the assumption that the state races are independent. In retrospect, one could have modeled either way and been 'victimized' by a pretty substantial one-way polling error (underestimating DJT). To be fair, I did not go to the extreme of simulating independent errors state-by-state, but intuitively I doubt it would have altered the final conclusion in any material way.
You don't have a baseline and methodology that is independent of the current case, something that would be applicable in all elections in the past and future, you only looked at the current case and made up scientifically looking probabilities without any citations of sources that back the probabilities. The word "presumed" shows up in your rationale, the correct word should have been "unknown". There is a training button at the top of the GJO page, it's a tutorial.
Lastly, if a forecaster compliments you for being realistic but simplistic, the forecaster doesn's know, that realistic and simplistic is the hallmark of good science and desirabln. Your model was unrealistic nor simplistic.
It's fair to say, it's a model, it's a thought, it's not your identity or character. Errors are allowed and without consequences.
The big professionals got it wrong, too, by the way. No big deal.