The Daily Kos Elections presidential and Senate forecasts have been live for two weeks now, and we’ve gotten lots of questions about how it works and what it all means. I try to address questions in the comments (either in the twice-a-week model posts, or in other stories as well) when I see them, but, of course, not everyone reads those same comment threads. To make sure as many people as possible are all on the same page, let’s look at some of the most common questions we’ve seen so far.
HILLARY CLINTON STARTED WITH AN 88 PERCENT CHANCE TO WIN, AND NOW SHE’S AT 70 PERCENT. IT’S TIME TO PANIC, RIGHT?
Jesus Christ, no, it’s not time to panic. Please don’t invite me over to your house to watch football with you. OH MY GOD, WE’RE ONLY LEADING BY SEVEN POINTS AT THE HALF! IT’S ALL OVER, WE’RE DOOMED!
It’s true, there’s still a lot of game left to be played, and definitely the potential for things to go unexpectedly wrong during that time. That’s the main reason that there’s still a lot of uncertainty in our projections. Think of it this way: if you have that same seven point lead with only two minutes left to go in a football game, you can feel extremely confident that you’ll win the game. That’s the likely outcome, say, 95 percent of the time. (And the games that make up the other five percent tend to be the ones that take on legendary status.) If it’s halftime and you’re up seven, you still have 70-something percent odds of winning the game, based on the aggregate history of all games ever played. Intellectually, you can feel pretty confident … but there’s still going to be a lot of nervous tension for the next 30 minutes.
The same logic applies to the polls. If you’re up by, say five points on average in the national polls, you have leads of five points or more, on average, in enough swing states to get past the 270 mark, and you have smaller but persistent leads in enough other swing states to give you a substantial cushion, and it’s the last week of October, you’re almost certainly going to win. If you have those same conditions but it’s the first week of September, you still have strong odds of winning, but there’s still a fair amount of uncertainty for which you have to account.
Part of the problem here is that we may have unintentionally started the model at an artificial high-water mark (in other words, if you start at 90 percent odds, there’s really nowhere to go but down, which is bound to leave some people feeling disappointed). We waited until several weeks after the conventions to start publicly discussing the model, in order to let the convention bounces dissipate. What looks like it may have happened, though, was that Clinton’s convention bounce lasted an inordinately long time, in fact, most of a month (usually the short-term effect of the convention lasts a couple weeks at most). It looks like it only started to fade in the last week or two; maybe the convention bounce got conflated with other factors, like the extended blowback Donald Trump received from his attacks on the Khan family, which kept the bounce inflated longer than we would have otherwise expected.
So what we may have seen was a temporary eight-point lead that should have lasted two weeks, but lasted nearly a month instead, before reverting back to what you might think of as the natural state of the race, which is in more of a five-point range. That’s where the HuffPo Pollster national polling average previously had the race, and where they currently have it. (If you’re wondering why I’m linking to their averages, we don't use national polls in our model, only state polls. We don’t even collect them in our database, let alone graph them.)
An eight-point lead, in a presidential race, is pretty darn big. Keep in mind that nobody has won a presidential election by double digits since 1984, when Ronald Reagan drubbed Walter Mondale (despite Mondale’s massive crowd sizes). Even when Bill Clinton beat the hapless Bob Dole in 1996, that was “only” a nine-point margin. So when you’re leading by eight points in early August, think of it in the same way as, say, leading that football game by 10 or 14 points midway through the second quarter. That’s going to give you considerably better than 70 percent odds of eventually winning … at that point in the game. A five-point margin is still big by presidential-race standards, but much more, well, “normal,” and that’s going to bring your ultimate odds down to earth somewhat.
We’re, of course, not the only forecasting model to see this same process happening. Our colleagues at FiveThirtyEight have seen the overall odds fall in a similar manner. In fact, their “polls-plus” model for today gives Clinton 69.9 percent odds, essentially the exact same result. We don’t know what all exactly goes into their proprietary blend of herbs and spices, but as far as we know it has more complex elements than we do; we account for the “fundamentals” simply by using the Alan Abramowitz “Time for a Change” model. The net result of the methodological difference, however, is very small.
BUT DONALD TRUMP IS THE WORST PRESIDENTIAL CANDIDATE EVER. HE SAYS STUPID THINGS, HE CAMPAIGNS IN THE WRONG STATES, HE DOESN’T EVEN HAVE ANY MONEY. WHY AREN’T WE LEADING BY 20?
Well, there’s this overarching problem called “polarization.” In the 1964 election, or the 1972 election, or even the 1984 election, there were a lot more people who weren’t clearly attached to one party or the other, because the parties weren’t that clearly delineated. They were weird, shifting coalitions that were only partly about ideology (you had conservative Democrats in the south, and liberal Republicans in the northeast). Now, there’s a much more obvious liberal party and conservative party, and 40 to 45 percent of the electorate is going to vote for their party no matter what. The days of a Goldwater or McGovern-style blowout simply aren’t going to repeat themselves under our current system.
Think of it this way … if a random wealthy celebrity somehow managed to hijack the Democratic nomination by making a lot of crazy promises that appealed to the Dem base’s lizard brains, and then proceeded to run an incoherent, off-putting campaign against a generic Republican, the average Daily Kos reader wouldn’t say “oh, well, guess I’ll vote for the Republican this year.” The existential threat posed by that Republican would utterly foreclose the possibility of voting for anyone but said rich buffoon!
WELL, MARKOS RUNS A PIECE ALMOST EVERY WEEK THAT SHOWS WE’RE WINNING ALL THE SWING STATES, WITH A HIGHER ELECTORAL VOTE TOTAL THAN YOU’RE SHOWING. HE HAS A MUCH ROSIER DESCRIPTION THAN YOU DO. YOU CAN’T BOTH BE RIGHT, CAN YOU?
It’s not so much a question of right or wrong, as much as we’re trying to accomplish two different things here. Markos's summaries aren’t projections of the likelihood of a future result, as much as they’re a description of where we are right now. If the election were today, and there wasn’t an across-the-boards catastrophic polling error lurking, yes, Hillary Clinton would almost certainly win easily. The polling averages in the key states do, in fact, give Clinton leads in enough states to add up to 340-and-change electoral votes, assuming that each polling average in each state plays out in exactly the same way when it comes to the actual election.
The reality, though, is that each state-level polling aggregate has some built-in uncertainty. That’s what our model tries to account for. In describing our 2014 election model, one metaphor I used several times was the idea of the plate-spinning contest. Markos’s summaries of where we are right now, essentially, assume that the plates are all going to keep spinning. If you have, say, a two-point lead in Ohio, that isn’t a 100 percent guarantee that your plate is going to keep spinning; it says that you have maybe a 60 or 65 percent chance of keeping your plate spinning. When you’re basing that on only a few polls in a given state per month, there’s a lot of room for error there. (There will be less error as we head into the final two months, as we start seeing a lot more polls each week in the major swing states.)
So, we run thousands of different computer simulations (called “Monte Carlos”) and look at the distribution of all the outcomes. In some of those simulations, the Ohio plate falls down, which means you get less than 340 EVs. Sometimes, the Florida plate falls down. Sometimes, even the Pennsylvania plate falls down. Sometimes, the Ohio, Florida, and Pennsylvania plates all fall down at the same time (those are the 20-something percent times where Trump manages to win). Sometimes, on the other hand, those plates stay up and we even manage to keep an Arizona or Georgia plate up, which takes you well past 340. But the median number of electoral votes that Clinton gets, in all the thousands of simulations we run, is most likely going to be less than the 340 that you get simply by extrapolating out the states’ current polling averages. Currently, the median number of EVs in all those simulations is 291.
It’s also worth noting: Markos is using HuffPo Pollster averages in his summaries. They use almost entirely the same polls as we do, though they do have a policy of throwing out pollsters who make landline-only calls. You can especially see that in their Iowa average; Pollster still sees a small Clinton advantage there, because they aren’t counting the recent poll from the Emerson College Polling Society that gave Trump a 5-point lead in Iowa. Emerson’s presence also puts a little downward pressure on our Ohio and Pennsylvania averages, compared with HuffPo’s averages.
SO WHY DON’T YOU THROW OUT EMERSON/RASMUSSEN/GRAVIS/INSERT NAME OF OTHER CLOWN POLLSTER HERE, TOO?
It’s not safe or wise to start throwing out pollsters you don’t like, or that you have methodological doubts about. They could, after all, be right this one time, and everyone else could be wrong and following the herd. The best thing to do is to throw them on the pile with everyone else; if there’s a lot of data, any one pollster’s house effects will be buried under the cumulative weight of all the other pollsters. The one correction we make is to make a 1.5-point correction in the results from specifically partisan polls, i.e. not just ones that we are suspicious of, but polls that were commissioned by campaigns or interest groups and that were performed by pollsters who only work for one particular party, which, if they got leaked, we can assume reflect a best-case scenario for those who commissioned them.
The one potential problem is in states where there isn’t a lot of data yet; Nevada, for instance, is probably the currently least-polled swing state, and several of the most recent polls of Nevada come from notoriously Republican-friendly Rasmussen. That, by itself, could explain why Nevada looks currently closer than its demographics would suggest. But it’ll take more polls added to the Nevada pile before we get a better sense of whether that’s true, or something is truly different in Nevada this year.
OK, BUT IF YOU SAY THE POLL AVERAGE IN OHIO IS A TWO-POINT CLINTON LEAD, WHY ARE HER ODDS IN OHIO UNDER 50 PERCENT?
If you’re particularly observant, yes, you may have noticed that on our landing page we show odds for each state, and we have less than 50/50 odds for Clinton in Ohio, Florida, Iowa, and Nevada. That’s despite the fact that she has a small lead in Ohio and Florida when you do a smoothed average of the polls in those states. (If you want to see the actual trendlines for each state, I’d recommend going to Drew Linzer’s old site, Votamatic. Some of the supplemental graphs and charts associated with the current model, that haven’t been built into the Daily Kos interface yet, are available there. For instance, you’ll see that the Ohio average, when distilled to a two-party share of the total, works out to a 50.5-49.5 average. Despite that, if you look on the Daily Kos model’s landing page, it gives Clinton a 46 percent chance of winning Ohio.)
The reason for that, again, is that we add some weight from the “fundamentals.” Absent any historic precedents, that small lead in the Ohio polls would give Clinton, say, a 55 or 60 percent chance of winning the state. The “fundamentals,” however, act as a bit of a weight on her chances, because historic precedent says what she’s doing is pretty hard, i.e., trying to retain the presidency after two terms of the same party in power, against the backdrop of current presidential approval in the low 50s and middling growth in the GDP in the second quarter.
Just as an example of how unusual that is, the only time in post-22nd Amendment history where that sort of handoff was successful was George H.W. Bush succeeding Reagan. (Obviously, the Democrats stayed in power for five straight terms in the 1930s and 40s, but the handoff involved the death of FDR, and Harry Truman taking over for him.) In fact, for the last time a two-term Democrat successfully handed power to another Democrat in an open seat election, you have to go all the way back to 1836, when Martin Van Buren followed Andrew Jackson!
The “fundamentals,” via the “Time for a Change” model aren’t a big weight; without any polling input, her odds would be only just short of 50-50 overall (instead of 70-30). However, it’s enough of a weight that a one-point advantage in the polls gets reduced to a 46 percent chance of victory in Ohio. But with each passing week, the “fundamentals” will exert a little less weight; by Election Day, they won’t have any effect at all, and the probabilities will be purely based on the polls themselves.
To return to the plate-spinning contest metaphor, our models have a lot of Ohio, Florida, Iowa, and Nevada plates clattering onto the floor … often individually, sometimes all together. At the same time, though, there are other swing states where the polls show much wider margins, and accordingly each of those plates stay up in 80 or 90 percent of all the simulations. In particular, that includes Colorado, New Hampshire, Pennsylvania, Virginia, and Wisconsin. Those are, in fact, the only plates that you need to keep spinning in order to hit 270, which is the only number that really matters … and that’s why Clinton is still winning the large majority of all of our runs of the model.
CAN I ASK SOME FOLLOW-UP QUESTIONS?
Of course you can! I’ll try to be as responsive as possible in the comments, and hopefully not make any more fun of your football-watching skills.