THE WILDCATTERS: How Much Confidence Should We Have in the Polls?

The Wildcatters are trying to develop our final election predictions, tweaking the Little Smart Pill Machine and trying to wrap our heads around mountains of data and molehills of conjecture. We have concluded we must consider one important question: How much confidence should we place in the polls?

Because we are social scientists, we can start with a simple declaration (we’ll call it the Wildcatter’s First Law of Electoral Analysis): Evidence is always better than guessing. And, polls are evidence of a sort. They are snapshots in time of the responses and preferences of a sample of the population. So, we should also make this clarification. The polls are never wrong. They are what they are – a snap shot of a reality from a sample of a population, which might or might not be perfectly representative. Our analyses of what they reveal may miss the mark, especially in terms of deciding what to do with undecided voters and predicting who turns out to vote (likely voters).

We should make another clarification on what should be an obvious point: No pollster wants to get the election wrong. They get no points for showing a candidate winning when, in fact, the candidate is losing. They may miss the mark but it isn’t for a lack of trying. Moreover, during a presidential election, there would be little incentive for misrepresenting the data. This is bad for business. Polls results may influence decisions early in a race by establishing the credibility of candidates in a crowded field especially where voters and donors are weighing electability OR in low visibility races where candidates are not particularly well-known and the poll can serve as heuristic for independent voters. But this is not the case in the general election campaign. Potentially, bad polling numbers may also dampen enthusiasm for a candidate, driving down voter turnout, but in most presidential elections—especially this presidential election—this seems unlikely.

In presidential elections, the polls tend to perform pretty well, especially when we consider the results from a wide range of polling organizations and carefully weigh all the available evidence. Here is a quick review:

  • In 2012, the final polls showed a remarkably close campaign. Obama won 51.7 to 47.2. Despite concerns about skewed polls, the polls slightly underestimated Obama’s support. The reason? Obama voters were less enthusiastic than in 2008 but the Obama mobilization efforts were still highly effective in turning out the vote.
  • In 2008, the polls mostly got the presidential race right. Obama finished with 52.9 to 45.7 for McCain.
  • Same story in 2004 when George W. Bush won with 50.7 to 48.3 for John Kerry.
  • In 2000, the polls under-estimated Gore’s final turnout effort and his popular vote total. They subsequently missed the popular vote winner/electoral vote loser scenario. In fact, just before the election, it looked more likely that Bush might win the popular vote but lose the electoral-college.

So what does this tell us? In presidential elections, where polls have missed, they haven’t missed by much and it is mostly because of turnout. This could be a factor in the current election if Democrats turn out at lower than expected levels or if Republicans turn out at higher than expected levels.

But what about Brexit, the 2016 Michigan Democratic Primary, and 2015 Kentucky Gubernatorial Election?

Polling can miss. And these three examples are instances where recent polling is reputed to have missed the final winner/loser call in an election. We can dispense first with Kentucky as an example. Republican Matt Bevin did close the gap in the final days of the campaign to win election in remarkable and stunning fashion. Two factors contributed to his surprise victory and the failure of polling to capture his momentum: (1) not enough polling and (2) abysmally low voter turnout. The only polls close to the election – internal polls from Bevin – showed a race that had tightened to a coin-flip. Neither inadequate polling, nor low turnout are factors in the current presidential election.

The Sanders 2016 Michigan primary victory was even more surprising, as most polls had Sanders trailing far behind Hillary Clinton. The polls missed in Michigan because they underestimated the turnout of younger voters and because they underestimated how strongly these voters supported Sanders. Polling also didn’t continue through election. This is the best and most obvious case of polling “miss,” reflecting the failure of polls to correctly capture the preferences or voter turnout of a hard-to-reach group (younger voters), though notably it also occurred during a more difficult to predict primary election. Polling for caucuses is often even haphazard and prone to error for the simple reason that caucuses are harder to predict.

Finally, in Brexit, the polls did miss the final election call, predicting that remain would win, but we quickly forget how close the final polls were. The final Huffington Post polling average had remain at 45.8 and leave at 45.3. Moreover, only several days prior to the vote (less than a week), leave was slightly ahead in the polls. The predictions (but not the polls) missed because they underestimated “leave” voter turnout and assumed undecided voters would break for the status quo. The final Brexit vote was 51.9 to 48.1

What are the lessons for 2016? Turnout matters.

If Donald Trump wins the election, two things have to happen to the voter turnout models. First, new white voters and low propensity white voters turn out higher than expected rates. Second, Democratic voters stay home. Depression of turnout has been alleged as a strategic goal of Republicans in this contest, either through alleged voter suppression by limiting voting opportunities, or through efforts to demoralize the opposition Democrats. The initial reaction to the FBI’s reactivated Clinton investigation did dampen Democratic enthusiasm for Clinton, and the polls tightened. The primary effect, however, appears to have been to solidify Republican support for Trump (which may have happened absent the revelation). Whatever short-term effect the Comey revelation had on Democratic enthusiasm, it appears to have largely dissipated.

Bottom line: As we sort through the state-by-state data to make our final election predictions, we will rely heavily on polling data but we will also incorporate our understanding of the electoral contexts within individual states and the dynamics of this specific presidential campaign. To ignore the polls or to try to unskew the polls to obtain a specific outcome is an exercise in wishful thinking.