By Stephen Fisher
The majority of forecasts point to Hilary Clinton winning tomorrow’s US presidential election. Several of the poll, market and expert forecasts with probabilities for who will win are helpfully summarised by the New York Times here. The polls-based predictions are all, apart from one, pretty confident that Clinton will win. At the time of writing, Drew Linzer’s model at Daily Kos puts the probability of a Clinton win at 87%, HuffPost has 98% and Sam Wang at the Princeton Election Consortium estimate is as high as 99%. The New York Times’ own model is slightly less confident, on 84%. The exception is Nate Silver’s FiveThirtyEight model which puts Clinton’s chances at just 67%.
The lower probability for Clinton in the FiveThirtyEight model is partly (by comparison with most but not all models) due to 538 having a smaller estimate of her likely lead in votes. By comparison with HuffPost, for instance, this is because FiveThirtyEight put more weight on more recent polls that haven’t been so great for Clinton in wake of the FBI announcement last week of further investigations into her emails.
But the discrepancy is mostly due to the FiveThirtyEight model allowing for more uncertainty. So 538 have a higher probability of a Clinton landslide than other models as well as a higher probability of a triumphant Trump.
There have been some, shall we say, robust discussions of these issues, see here, here, here here, and especially here and here. Personally I think the FiveThirtyEight team are right to allow for more uncertainty in their model. Nate Silver put his case well in his Saturday update which said,
“And it’s important to remember that the outcomes in each state are correlated with one another, so that if Clinton underperforms her polls in Wisconsin (for instance), she’ll probably also do so in Minnesota. Forecasts that don’t account for these correlations are liable to be overconfident about the outcome. It isn’t hard to find examples of candidates who systematically beat their polls in almost every competitive state, as President Obama did in 2012 and as Republican candidates for governor and senator did in 2014.
And that’s before accounting for some of the factors that the model doesn’t consider: the disagreement in the polls, the unusual nature of Trump’s candidacy and the demographic changes it is producing, Clinton’s superior turnout operation, the possibility of “shy Trump” voters, the fact that the news cycle is still somewhat fluid headed into the final weekend, the declining response rates to polls, and the substantial number of high-profile polling misses around the world over the past few years. We think this is a good year for a forecast that calls for more caution and prudence.”
That last sentence still raises intriguing questions about how to calibrate and justify the level of uncertainty.
For most of the last few weeks betting and prediction markets seem to have been roughly half way between the FiveThirtyEight caution and the more confident poll-based forecasters, arguably tracking changes in 538 more closely. Since yesterday’s FBI null-findings announcement. The markets became more confident for Hilary. At the time of writing PredictIt has the probability of a Clinton win at 81% and Predictwise says 89%. The poll based forecasts will, presumably, include polling from before yesterday’s announcement and so they may not adequately reflect the consequences of that information. The markets are supposed to.
Not only markets but polls of citizen expectations have also shown that most Americans expect Clinton to win. As the following is from the New York Times/CBS News polls here shows, the ratio of those expecting Clinton instead of Trump to win has been about 5/3.
A citizen forecast by Andreas Murr, Mary Stegmueller, and Michael Lewis-Beck here shows that such surveys have a good track record in identifying who will go on to win, and based on previous experience this year’s citizen forecast survey figures suggest a comfortable win for Hilary. For instance, the most recent poll in the table above suggests at least a 6-point popular vote share lead.
Pollyvote helpfully summarizes and combines a huge variety of different forecasts and provides explanations of the component methods. They confidently predict a Clinton win with a 5-point popular vote lead and 323 electoral college votes.
In the Pollyvote mix are some forecasts for Trump, including some of the traditional academic ones. Two classics, Alan Abramowitz’s Time for change model based on incumbency and growth rate, and Helmut Norpoth’s Primary model both suggest Trump will win. These are forecasts from some way out and what they mainly tell us is that it is hard for a party to keep hold of the White House when they’ve been there for the last two-terms, and particularly when economic growth is below the 3 per cent the US public appear to expect (see here).
An electoral-cycle cycle based forecasting model did much better than the polls in Britain in 2015, but so too did citizen forecasts (see here). In the US this year these two approaches point in different directions. The PollyVote combined forecast, averaging across forecast types, appears to have a good track record for US presidential elections, and their analysis suggests that citizen forecasts and prediction markets typically outperform the polls in the US. Maybe it will continue to work very well, but my experience trying to replicate the combined forecast method with Rosie Shorrocks for the Brexit referendum showed that all the forecasts can end up pointing in the wrong direction.
Forecasts for the House of Representatives
Even though polls have consistently had the Democrats ahead on the generic congressional vote (albeit currently only by 1.4 points) few are expecting the Democrats to win a majority. This is because of the district boundaries and patterns of incumbency mean that House elections are biased towards the Republicans and they are relatively unresponsive to changes in party support, e.g. as discussed here and here. If the polls are right then there is likely to be about a 3.6 point swing from the Republicans to the Democrats since 2014. Based on 2014 seat change to vote swing ratios, this would mean around 13 seats reverting from Republican to Democrat, undoing the 2014 GOP gains on 2012. This would leave the Republicans on 234 with a comfortable majority over 201 for the Democrats. This calculation does not properly take into account the effects of first-time incumbency in the seats Republicans won back in 2012. So there’s a good chance even fewer seats will change hands. On the other hand, my back of the envelope calculation produces a similar to most current forecasts (e.g. here, here and here) and to the average of some earlier academic forecasts, summarised here. PredictIt has probability of the Republicans continuing to control the House at 93% which seems quite low to me under the circumstances.
The race to control the Senate is apparently the most competitive of the three this year. It is currently split 54 Republican 44 Democrat with two Democrat caucasing independents. Just as the poll-based forecasters differed on the presidential election, so there are corresponding differences for those predicting the outcome of the Senate elections. Part of that linkage is that forecasts of which party controls the Senate have to incorporate probabilities for the party of the Vice-President, since that post holder breaks ties. So a team with a higher probability of a Democratic president is likely to have a higher probability of Democratic control of the senate since a 50:50 outcome on seats is reasonably likely this year. But also the methods again differ according to how much uncertainty they allow. Senate forecasts are summarized by the New York Times here, with the screenshot of the Democratic control probabilities below.
Presumably (I haven’t checked) polling errors for presidential, house and senate elections are correlated and so if the polls have substantially underestimated Trump they are likely to also have have underestimated the Republicans for the Senate and House. This means that there must be a noticeable chance the Republicans will win control of win all three branches of government. PredictIt currently puts that probability of undivided government under the Republicans at 16%, higher than the probability of undivided Democrat government, which is 7%. Clearly continued divided government seems most likely.