It’s been roughly a month since Donald Trump shocked the world on election night, riding an incredible performance in the Rust Belt to a comfortable electoral college victory despite losing the popular vote by roughly three million votes.  Before the election, Nothing But Numbers gave Clinton a 79% chance of victory.  What went wrong?

  1. Trump Exceeded Expectations in the Rust Belt: Nothing But Numbers expected Trump to do very well in the Rust Belt before the election.  I even wrote an article titled, “Donald Trump Is Turning The Rust Belt Red“.  But Trump managed to not only win Ohio and Pennsylvania, which I expected, but also win Iowa, Michigan, and Wisconsin, states that have gone for the Democrats 17 out of 18 times since 1992.  Even Minnesota, which hasn’t gone for a Republican since Nixon, ended up being highly competitive.  In some ways, this was unsurprising; all four states have larger than average non-hispanic white populations, and our model showed close races in Michigan and Wisconsin.  In Iowa, however, the Facebook Likes data from April proved to be out of date, as Trump managed to win Iowa comfortably despite us labeling the state as “Safe Democrat”.   Iowa, in addition to Michigan and Wisconsin, proved to be our only 3 misses.  I suspect that newer Facebook data would have allowed Nothing But Numbers to call Iowa correctly, as Iowa’s shift towards Trump was very predictable based on demographics.
  2. Trump Outperformed His Projected Popular Vote Margin by 1.5 Points:  Going into the election, Nothing But Numbers expected Clinton to win the popular vote by 3.6 points.  As of this writing, her final popular vote margin is 2.1 points.  While this is a comparatively small error compared to the horrid state polls (which missed in 5 of 13 swing states), this error may have ended up making the difference in an election where Trump won Pennsylvania, Michigan, and Wisconsin by a combined 1.6 points.
  3. Clinton’s Hispanic Surge Didn’t Help Her in the Electoral College: While Clinton exceeded Obama’s winning margin by nearly 10 points in the two largest heavily hispanic states, California and Texas, pointing to a stronger showing with hispanic voters than Obama, it didn’t help her in the electoral college.  Simply put, hispanic voters are mostly clustered in uncompetitive states like California, Texas, New York, and New Mexico.  And while swing states Arizona and Florida have growing hispanic populations, they were drowned out in 2016 by non-hispanic whites, who voted for Trump in droves in Arizona and Florida.  Meanwhile, our classification of Texas as a toss-up state proved to be overambitious, as Trump ended up winning the state by 9 points. Despite doing significantly worse than Romney in the heavily hispanic urban areas of Texas, Trump exceeded Romney’s margin in whiter, more rural sections of northern Texas, and held his own in Texas’ highly educated suburban counties.
  4. Trump Dominated With White Working Class Voters: Going into the election, I expected Trump to improve on Romney’s numbers with white voters without a college degree.  In my Demographics model, I had them going for Trump by a 62-30 margin, compared to 60-38 for Romney in 2012.  While a detailed demographic analysis has yet to be done on the 2016 election, exit polls had Trump winning white voters without a college degree by a staggering 67-28 margin.  While exit polls should always be taken with a grain of salt, given Trump’s extraordinary performance in the largely white working class Rust Belt, these exit poll numbers are probably on to something.  Indeed, many of the typically Democratic white working class counties I pointed to as potential Trump pickups before the election ended up going for Trump, like St. Lucie County, Florida and Erie County, Pennsylvania.
  5. Traditional Conservatives Turned Out For Trump While Rust Belt Bernie Supporters Stayed Home:  One of the things I was watching on election night was how areas where Bernie Sanders and Ted Cruz did well in the primaries related to support for Clinton and Trump.  In the end, the correlation seems clear.  While Clinton was able to win over Bernie voters in deep blue states like Hawaii and Rhode Island, many Bernie supporters in Rust Belt states like Wisconsin, Michigan, and Iowa either stayed home or voted for Trump, as Trump’s victories in those states would not have been possible if most of Bernie’s supporters had joined Team Clinton.  Meanwhile, traditional conservatives who largely opposed Trump during the primaries seemed to have little trouble turning out for Trump in the general election.  Indeed, Trump’s victories in states like Iowa and Texas were largely fueled by voters who opposed him during the primaries, as Trump got less than 30% of the primary vote in both states yet won each state in the general election by 9 points or more.  The Facebook data, which was last updated way back in April, largely missed out on these changing trends in voter preferences.

Despite these errors, in getting 47 of 50 states correct and giving Trump a roughly 20% chance of victory, Nothing But Numbers did better than pretty much every other election prognosticator. FiveThirtyEight, for example, had Pennsylvania, Florida, and North Carolina all going for Clinton, while Nothing But Numbers correctly forecasted Trump winning all three states.  Meanwhile, Sam Wang of Princeton gave Clinton a greater than 99% chance of winning the election, a now infamous prediction that will probably serve as a good lesson in statistical margin of error for future election statisticians.