June 29, 2016
If you’re reading this, there’s an 80-90% chance you know who Nate Silver is. After he correctly predicted 49 of 50 states in 2008 and all 50 in 2012, he was anointed as a predictive savant, something he found more amusing than accurate. Nate’s actual contribution is bringing baseball-style (he cut his teeth at Baseball Prospectus) data analysis to politics.
For all the talk about polls, most pundits aren’t real versed (like at all) in reviewing data, tending to grasp at whatever result bits fit their preferred narrative instead of digging for the most likely reality. Having Silver and his blog-turned-platform FiveThirtyEight around provides some important balance and has definitely influenced how I look at numbers.
His history and status in the political data world made his miss on Donald Trump all the more significant. When Trump was sitting at the top of the polls in late summer, Sliver estimated he had a 2% chance of winning the GOP nomination. Even when Trump’s polling held up for another few months, Nate was still very skeptical, waiting for him to actually start winning primaries to concede a nomination was a real possibility.
Many were fooled, but none are held to Silver’s standard. Price of fame and all that. In the meanwhile, FiveThirtyEight just released their election model for the 2016 general election. Can we trust it? Yes.
A couple/few reasons:
Determining Trump was a 1 in 50 shot to get nominated was very justifiable at the time Nate gave those odds. We now (assuming no convention revolt) have 2 GOP nominees in the past 150+ years who never held political office, served in the cabinet or were a military general. Trump and Wendell Willkie in 1940. He didn’t say it was impossible, just very unlikely.
However, Silver didn’t adjust quickly enough when the poll numbers held up far better for Trump than Herman Cain, Steve Forbes, or any slightly similar past failure. He attributes this to not building an actual model and pulling the estimate basically out of his ass. By the time the primaries began, FiveThirtyEight had estimates for each contest that had a decent amount of polling data. These worked out very well.
Since they have an updated version of their 2008/2012 model for the general election, and the primary projections actually showed Trump doing slightly better than he actually did (they had Trump winning Iowa and Oklahoma), there’s no reason to assume any error from last year is continuing uncorrected.
Most importantly, instead of attempting to pick a definite winner, his system assigns percentages. Currently Trump is between 19% and 26% likely to win the election, depending on which measurement you prefer. There’s choice there too, either a polls-only number, or a polls-plus, which includes economic data and other background factors.
In a world and election where we know there’s some uncertainty, this seems like the best approach. They run 20,000 simulations daily, so if Trump is 20% likely to win, it means he came out ahead in 4,000 of them. They show a distribution that gives the full range of outcomes. There are a few results where Trump wins by a huge amount, others where he gets clobbered.
It’s a useful reminder that both the people who are convinced he could never win and those who think he’s the stronger candidate both have reasonable scenarios to back them up. So, we’re using the FiveThirtyEight projections as a baseline for our analysis this fall.
Sometimes we’ll revert to talking about a specific poll or set of polls that are getting some run in the media. It’s an opportunity to give you a look at the numbers that are being thrown around. Showing how the conventional wisdom is missing something never gets old.
But when we’re talking about how likely Trump is to win Ohio, or Clinton to take Arizona, we’ll begin with what FiveThirtyEight is showing and go from there. Consider it the baseline expectation. If we want to know how Clinton picking Elizabeth Warren might move the needle, start with the odds in a given state or nationwide and then adjust based on how she may impact outcomes.
If FiveThirtyEight thinks Clinton is 99% likely to win a given state, we’re not going to worry about Trump having a ground game disadvantage there. If they show Trump as 99% likely to win somewhere, we won’t overreact if a single poll makes it appear he’s in trouble there.
That still leaves plenty of room for fun speculation. Stay tuned.