The Political Polls Got It Wrong

• This is the appearance
• Polls are imperfect estimates at a single point in time
  • Polls have known average error and error varies over the years
  • They are pretty good, but sometimes error is high
  • We forget this
• The error in the estimate of Clinton’s percentage was actually very low (1%)
  • Difference between percent vote for all polls is in the margin of error
• Obama outperformed polls
  • Created an assumption that Democrats should outperform polls
  • the polls were actually worse (off by 3%)
• Idea is that lots of polls reduce uncertainty (ie don’t pay attention to individual uncertainty)
• Clinton had a 67% probability according to 538
  • that means that there is a real possibility that Trump could win
• Significant differences between models
  • Why? see error in model assumptions below
  • aggregators are taking polls and creating models without considering assumptions properly
• Thought that 8 50/50 states all had to go to Trump
  • this is a bad assumption!
  • if these states are independent the probability is .5^8 = <1%
  • BUT error is correlated - tendency for Trump to overperform
    • thus if Wisconsin is underestimated (and goes for Trump) so will PA and FL etc
• For blue states Clinton over and underperformed equally
• correlation between poll error and % white non-college vote
  • states we talk about as battleground (eg WI, PA, FL, OH) have lots of white non-college
  • this population is increasingly leaning Republican
  • GOP has lost white college grads
  • BUT if this group was sampled correctly it wouldn’t create biased polls
  • this population also isn’t turning out to vote
    • assuming this means that if you’re wrong you introduce error in the poll
    • using a formula to decide if you will vote and whether you should be counted in poll
    • more likely to drop less educated voters from poll
    • small errors lead to just enough underestimation to create consistant bias
• Take poll and weight on the basis of the census
  • weights do not include education, but if education matters than there’s bias
    • the influence of education has changed
• Error in aggregator models
  • greater uncertainty in 2016 due to non-incumbant
  • lots of late deciders influenced by negative Clinton press
• Were people afraid to admit to voting for Trump
  • BUT social pressure would be mostly in blue states, but those states were consistent w polls
  • BUT Republicans consistently outperformed polls
• We will hopefully adjust for error in models but there may be other issues in future
  • reweight data to account for education
  • reconsider likely voter formula
• Polls have been having issues elsewhere
  • eg Brexit passed
  • eg Colombia voted against FARC peace deal
• Confirmation bias
  • media believed it couldn’t happen so they take the pieces that look like this and believe them
  • early voting suggested Clinton
• If polls were wrong about elections what does that say about approval ratings?
  • BUT approval rating is all people; no worries about likely voter and associated formula
  • BUT popular vote estimate was very good
• Other cases of correlated error
  • probability of massive subprime loan defaults in 2007
    • spreading out risk of any given default sounds great
    • HOWEVER if one fails then likely everything else will fail
    • SO we think risk is reduced but instead risk is high throughout

Factors that were underestimated

• Clinton had lost before
• Trump had been winning more than expected
• People looking for something different
• Setbacks immediately prior to voting
• Comey investigation right before election