Distribution of likely number of Democratic Senate seats
Daily Kos Elections is pleased to announce the introduction of its Poll Explorer! The Poll Explorer is a statistical model that allows us to track public opinion in each Senate and gubernatorial race this year, and from there, predict the likely outcome of each race and the number of seats that each party will control after the election is held. This is an ongoing feature that will be available from now until the November election. We'll be doing posts like this one, describing the current state of the election, several times a week from now until Nov. 4; however, there will also be permanent pages for all the
Senate and
gubernatorial data which will be updated every day, and which you can bookmark and access at any time.
As the headline states, today's version of the model shows that if 100 elections were held today, the Democrats would control the Senate in 47 of those elections. That means that it's slightly likelier than not that the Republicans will win the necessary 50 seats for control. The median number of seats that the Democrats end up controlling, in the hundreds of thousands of simulations that we run, is 49. (In other words, if you lined up every single result in order, 49 seats is the point where 50 percent of the results would be higher than that, and 50 percent would be lower.)
However, there's an interesting quirk here: the likeliest distribution of seats would be 50 Democratic-controlled seats and 50 Republican-controlled seats, which, thanks to Joe Biden's tie-breaking vote, would allow the Democrats to retain control. That result is only very slightly likelier than the second-likeliest outcome, though, which is 49 Democratic-controlled seats and 51 Republican-controlled seats, and the more outlying results are more likely to be Republican-friendly (for instance, it's much likelier that the Democrats will end up controlling 46 seats than they will control 53 seats), which explains why the Republicans have slightly better than 50-50 odds overall. (In other words, 49 Democratic seats is the median, while 50 is the mode.) You can see the full distribution of the range of likely numbers of Democratic seats in the bar chart, known as a histogram, shown above.
And there's one more quirk here, too. If you look closely at the specific races, you'll notice that the Democrats have greater than 50 percent odds of winning 50 individual races. The Democrat-held seats that they are, individually, on track to lose are South Dakota (1 percent odds), Montana (1 percent), West Virginia (1 percent), Arkansas (37 percent), and Louisiana (46 percent). The Democrats are currently on track to win Iowa (51 percent), North Carolina (58 percent), Colorado (63 percent), Michigan (72 percent), Alaska (82 percent), and New Hampshire (82 percent). However, you have to take the races into consideration as a
whole.
Think of the fight for the Senate as a plate-spinning contest: the Democrats are currently spinning a lot more plates than the Republicans, and it only takes one more plate falling for them to lose control. The Democrats have a lot more wobbly plates (North Carolina, Iowa, Colorado) than do the Republicans (Louisiana). In a majority of simulations, one of them still falls, even though individually each plate has a better-than-even chance of not falling; such is the nature of probability.
Before you start running around in circles screaming that "even the liberal Daily Kos" says the Democrats are going to lose the Senate, please bear in mind this most important caveat: this is not a prediction that the Democrats will lose the Senate, any more than if we gave the Democrats a 53 percent chance of controlling the Senate, it would be a prediction that the Democrats will win the Senate. It's merely a reflection of the fact that control of the Senate is truly a coin flip, and if there's any contention between us and 538 or the New York Times over what specific percentage should be applied, it's purely an argument about whether that coin, as of today, is ever-so-slightly weighted toward the Democrats or Republicans.
Moreover, the model is merely a description of the polls. If the polls are simply wrong, then the model will be wrong, too. We aren't terribly concerned about that, as poll averaging produced highly reliable results in the 2008 and 2012 presidential races. Nevertheless, statistical modeling is only one potential approach to forecasting an election, and Daily Kos Elections will continue to offer more subjective and gestalt-based Senate and gubernatorial race ratings as well (in much the same way that baseball teams that have incorporated a more sabremetric approach continue to employ scouts, as well).
We'll delve more thoroughly into how the model works and why we're doing this, over the fold:
The Daily Kos Elections Poll Explorer is an adaptation of Votamatic, which was a site developed for 2012 by Drew Linzer that correctly predicted the 332-206 outcome in that year’s presidential election. We've repurposed the Votamatic model to work in the Senate and gubernatorial context as well. It's a Bayesian model that, first, creates smoothed trendlines for each individual race and then runs thousands of Monte Carlo simulations to see how likely Democratic candidates are to win each race, and how likely Democrats are to win the Senate and gubernatorial playing fields as a whole. If you would like to read a much more detailed explanation of each step in the process, and how to interpret the various charts, that can be found at our How It Works page.
The model draws its data from the Daily Kos Elections Polling Database, which is scrupulously maintained by our own Steve Singiser, and which is a great resource if you want to zoom in on the complete polling history of a particular race. We cast as wide a net as possible including all public pollsters (and yes, that includes Rasmussen), under the assumption that averaging all polls will smooth out each individual pollster's house effects or other quirks. We also include polls from partisan pollsters, such as those leaked by campaigns, although we uniformly adjust partisan polls to make them less favorable to the candidates who commissioned them.
We’ve stress-tested the model by retroactively applying it to the 2012 election, which finds it correctly predicting all but two Senate races. Those races are Montana and North Dakota; those two races, it's worth noting, also stumped other prominent aggregators. (And that's simply because there wasn't a robust selection of polls of the Montana and North Dakota races, and the majority of those polls were, at the end of the day, simply wrong; more gestalt-based predictors who weren't bound to models were able to get one or both those races right.)
Now that you know how it works, you might be wondering: why do we need another model? After all, there are perfectly serviceable predictive models courtesy of sites like 538, the New York Times' Upshot, and the Washington Post's Election Lab. Part of the problem, though, is that every model is built on its own set of assumptions; in fact, the next step for some enterprising person might not be to build a model that aggregates the polls, but one that aggregates the aggregators. Models may not just have, for instance, different ways of accounting for polls from partisan pollsters or polls that are too old, but also give different emphases to other factors, like polls of the generic congressional ballot, economic data, or historical trends in midterm elections.
Part of what sets our model apart is that it doesn't account for those economic or historical factors; it's just the polls. Historical factors are relevant earlier in the cycle, when there simply isn't adequate polling data and you need something to fill the gap, but they aren't (as) necessary now that we have a suitable number of polls in all the key races. In other words, this is intentionally set up to be an all-meat, no-special sauce model. Omitting those factors, however, means that this isn't a "predictive" model; we don't add in expectations that voters will swing toward the Democrats or Republicans between now and Election Day. The range of possible outcomes that we calculate for the Senate and gubernatorial races simply starts from where we estimate public opinion to be today, then adds in uncertainty about how voters' preferences might change over the next two-and-a-half months.
To the extent that the models are poll-driven, though, the end results aren't too different; our expectation that the Republicans have a 53 percent chance of taking the Senate is quite close to the NYT's prediction that they have a 65 percent chance of doing so, or the Washington Post's prediction that the GOP has a 63 percent chance of taking the Senate. They see the same polls as us, and there's only so much you can do differently in interpreting them.
The slightly more pessimistic result that these other aggregators arrive at has largely to do with the way that their models rely less on polls and more heavily on economic and fundraising data, which is something that we consider, at this point, already be showing up in the results pollsters are finding. If anything, the Democratic candidates are currently overperforming in the polls compared with what economic and historical fundamentals would lead you to expect. However, if a late-breaking wave or black swan event suddenly causes that Democratic overperformance to head south, that too would be rapidly reflected in the polls.
And there's one other key difference that separates us from the other aggregators: we are also looking at the gubernatorial races. It's not as sexy a topic, since you don't get a prize for holding the majority of gubernatorial seats, and at any rate the number of likely Dem-held gubernatorial seats isn't hovering right between 24 and 25. But gubernatorial races are hugely important, both in terms of establishing progressive policies in individual states and building a presidential bench, so we want to give a similar focus to gubernatorial races using the same method.
As you can see in our gubernatorial histogram, the median outcome would be that, after the election, the Democrats will hold 21 gubernatorial seats. As you can see on the histogram, the modal outcome would be for the Democrats to hold 20 seats. Twenty-one is the current number that the Democrats hold (which is why that column is gray), so that means the likeliest outcome would be either no net change or a net change of one in the GOP's direction.
That may sound surprising, when most pundits would say the Dems are on track to pick up a seat or two. However, the most current round of polls do not support that conclusion. Take a look at the totem pole of races to see the specifics: as most pundits expect, the Democrats have overwhelming odds in picking up the Pennsylvania (99 percent) and Maine (91 percent) races. In addition, they're on track to pick up Kansas (surprising a few months ago but not surprising now, at 63 percent) and Wisconsin (that
is a surprise, but it's a total coinflip at 52 percent).
However, they're also on track to lose not just Arkansas (19 percent) and Illinois (10 percent), but also Connecticut (11 percent) and Hawaii (15 percent). Many people assume that the dark blue lean of Connecticut and Hawaii will save those races in the end (and that's particularly the case in Hawaii, where David Ige beat unpopular Neil Abercrombie in the primary—bear in mind, though, that the model does reflect polling of an Ige/Duke Aiona race). For the moment, though, the polls—which are all we're relying on here—do not support that idea.
To reiterate, none of this—not the percentages for the individual races, nor the percentages for the Dems' overall odds of holding the Senate—are set in stone. The percentages can, and will, fluctuate quite a bit over the coming months. That's why we'll be updating the model with new data every day, which you can see by visiting the permanent pages. I'll also be writing a full post like this one several times a week, discussing what has changed in the model since the last post, and what individual polls may have changed the numbers. So please follow along as we track the stretch run for Election '14.
