Questions about the Statistics + U.S. Elections
Prof. Maria Tackett
What are the different ways that pollsters determine likely voters vs. registered voters? Have those methods changed at all since 2016?
What does that mean for a race where a lot of energy on both sides is going in to turning out the vote rather than just persuading "moderates." (I'll leave this to the campaign strategists π)
Adults vs. registered vs. likely voters
Adults: People in the U.S. ages 18 and older
Registered voters: People registered to vote.
- Usually determined by a straightforward survey question about whether respondent is registered to vote in their election distribution
Likely voters: More representative of the likely voter electorate
Ask a series of questions about interest in upcoming election, past election participation, and intention to vote in upcoming election.
Some adjustments made, e.g. adjustments made for young people ineligible to vote in previous elections
Use models to predict the likelihood a respondent will vote
- Some pollsters now asking if respondents have already voted
More info:
- Marist Poll General Informaiton
- See likely voter survey questions from Sienna College Research Institute
It has been said that the so-called "hidden votes" gave president Trump an edge in the last election, contradicting the polls. How has that been addressed after four years, if at all possible?
- There is little evidence of the "shy" Trump voter
- Study by the American Association for Public Opinion Research found that "interviewer administered polls did not under-estimate Trumpβs support more than self-administered IVR (interactive voice response) and online surveys".
There is little evidence of the "shy" Trump voter
- Study by the American Association for Public Opinion Research found that "interviewer administered polls did not under-estimate Trumpβs support more than self-administered IVR (interactive voice response) and online surveys".
In 2016, many state-level polls underestimated the number of white voters without a four-year college degree.
- Most polls have started weighting by education to reduce over-weighting of people with a college degree in the sample
There is little evidence of the "shy" Trump voter
- Study by the American Association for Public Opinion Research found that "interviewer administered polls did not under-estimate Trumpβs support more than self-administered IVR (interactive voice response) and online surveys".
In 2016, many state-level polls underestimated the number of white voters without a four-year college degree.
- Most polls have started weighting by education to reduce over-weighting of people with a college degree in the sample
In general, some groups of people are hard to reach, so pollsters have to apply statistical methods to re-weight their sample, i.e. make it more representative of the actual electorate (not just representative of who responds to a poll)
Read more about sampling weights
An Evaluation of 2016 Election Polls in the US by the American Association for Public Opinion Research
Trump Supporters Arenβt βShy,β But Polls Could Still Be Missing Some Of Them by FiveThirtyEight
A Resource for State Preelection Polling by Pew Research
Same as the candidates' campaigns, are election statistical analyses also focused mainly on "swing" states?
Though the discussion may focus primarily on swing states, statisticians are interested in all of the states!
There is useful information to be learned about variability and error in predictions even from states where the outcome is unsurprising.
Though the discussion may focus primarily on swing states, statisticians are interested in all of the states!
There is useful information to be learned about variability and error in predictions even from states where the outcome is unsurprising.
Scenario 1
| Predicted | Actual | Error | |
|---|---|---|---|
| Candidate A | 51% | 49.5% | -1.5% |
| Candidate B | 49% | 50.5% | +1.5% |
General consensus: π
Statistician: π
Though the discussion may focus primarily on swing states, statisticians are interested in all of the states!
There is useful information to be learned about variability and error in predictions even from states where the outcome is unsurprising.
Scenario 1
| Predicted | Actual | Error | |
|---|---|---|---|
| Candidate A | 51% | 49.5% | -1.5% |
| Candidate B | 49% | 50.5% | +1.5% |
General consensus: π
Statistician: π
Scenario 2
| Predicted | Actual | Error | |
|---|---|---|---|
| Candidate A | 65% | 55% | -10% |
| Candidate B | 35% | 45% | +10% |
General consensus: π
Statistician: π±
How long does it take for statistical models to be affected by current events,like a scandal?
Statistical models are primarily rely on polling data, so they take current events into account only as quickly as the polls do.
Also how accurate are models really, when Hillary was favored by a decent margin four years ago and still lost?
What are some types of prediction model used, and what is more efficient? For example in the last election, there are many newspapers (if not say a lot) would predict Hillary to win, and some were against the odds to predict Trump win (and he actually did).
Are models accurate?
Models produce probabilities, not final answers! (Think about your logistic regression models.)
- In 2016: FiveThirtyEight gave Trump a 30% chance to win the day before the election; the Upshot forecast gave him an 15% chance.
- Rare events do happen! March Madness π
- Is 15% or 30% really that rare?
Are models accurate?
But, the models weren't perfect...
Are models accurate?
But, the models weren't perfect...
- In 2016, state polling errors were largely in the same direction
- Clinton over-performed in a lot of blue states and under-performed in key Midwest states (that's a problem! errors should be random)
Are models accurate?
But, the models weren't perfect...
In 2016, state polling errors were largely in the same direction
- Clinton over-performed in a lot of blue states and under-performed in key Midwest states (that's a problem! errors should be random)
Issues with polling get baked into the model (bad data in, bad data out).
Are models accurate?
But, the models weren't perfect...
In 2016, state polling errors were largely in the same direction
- Clinton over-performed in a lot of blue states and under-performed in key Midwest states (that's a problem! errors should be random)
Issues with polling get baked into the model (bad data in, bad data out).
Some models didn't accurately account for correlation between states, especially in the Midwest.
Are models accurate?
But, the models weren't perfect...
In 2016, state polling errors were largely in the same direction
- Clinton over-performed in a lot of blue states and under-performed in key Midwest states (that's a problem! errors should be random)
Issues with polling get baked into the model (bad data in, bad data out).
Some models didn't accurately account for correlation between states, especially in the Midwest.
Uncertainty was not effectively communicated to the general public
Read more reflections on 2016 models
An Evaluation of 2016 Election Polls in the US by the American Association for Public Opinion Research
The Real Story of 2016 by FiveThirtyEight
Presidental Forecast Post- Mortem by the Upshot
Final thoughts
- Focus on state polls, not national polls!
Final thoughts
Focus on state polls, not national polls!
Don't read too much into early results
- Can be potentially misleading based on who is expected to vote early, use mail-in voting, or vote on election day
Final thoughts
Focus on state polls, not national polls!
Don't read too much into early results
- Can be potentially misleading based on who is expected to vote early, use mail-in voting, or vote on election day
There is a lot of uncertainty this year!
Final thoughts
Focus on state polls, not national polls!
Don't read too much into early results
- Can be potentially misleading based on who is expected to vote early, use mail-in voting, or vote on election day
There is a lot of uncertainty this year!
Don't let predictions affect your behavior! VOTE!
