Friday, August 5, 2016

The limits of certainty as Trump self-immolates

As Donald Trump's campaign continues to break new ground in campaign stupidity, it is worth taking some time out to consider the limits of how certain we can ever be that he will lose.  Currently, PredictWise puts Trump's chances at 23%.  At this stage of the game, I defer to the betting markets, so let's go with that.  However, I'm a statistician at heart no matter what stage we reach, and that means never putting the chances of anything at 0 or 1.  Dewey/Truman, blah, blah, blah.

So, how certain can we ever be about Trump's chances, and why?  Let's talk about two main reasons to keep Trump's chances above 0.

1)  Intervening events
This is really the big one.  If the polls are right (more on this shortly), then all Clinton has to do to win is maintain the status-quo.  But, a major economic downturn, or a terrorist attack that Trump can exploit, or something like that, and Trump can win.  If Clinton is up by six points, then a shift among three percent of the electorate will flip the race.  The basic partisan closeness of the electorate makes things malleable.  Even with a truly, epically lousy candidate like Trump, who has shot himself in the foot so many times that he's just shooting at blank space right now.

The catch is that the clock is ticking.  With each passing day, there is one fewer day for such an event to happen.  So, as each day passes, Clinton's chances of victory should increase by some small increment while Trump's chances of an intervening event slip away.

However, Clinton's chances won't get to 100%.  Never 100%.  The polls could just be wrong.  And that leads us back to...

2)  The reverse-Bradley
I have written about this before. (Incidentally, if you check the date, the Khan reference there was not related to the current Khan thing. I'm just a Trekkie).  The basic idea is that certain types of voters might be prone to lie about their willingness to vote for Trump.  His embrace of racially divisive language might make some voters uncomfortable admitting that they will vote for him, in which case the polls will systematically underestimate his support.  This would be the reverse of what supposedly happened to Tom Bradley in California in 1982.  Tom Bradley, an African-American gubernatorial candidate, underperformed some of the polls.  The "Bradley effect" is the notion that voters were telling pollsters of their intent to support Bradley to seem not-racist when they had no intention of actually voting for him.  Perhaps in a reverse-Bradley, voters are refusing to admit that they intend to vote for Trump so as not to seem racist.

Now, the evidence for the Bradley effect was never as strong as the legend, and Trump's polling numbers were never that far off from the vote tallies in the Republican primaries, but that could be different in the general election.  However, the volatility in the polls before and after the conventions suggests no reverse-Bradley.

At this point, I should also admit that when I wrote that earlier post on the reverse-Bradley effect, I overestimated Trump.  I didn't think he could possibly be this bad a candidate.  Who attacks the grieving parents of a fallen war hero?  I got nothin'...  Trump just....  I got nothin'...

Still, we can't discount the possibility of polling errors, nor do we know what a candidate like Trump does to turnout models.  We don't know how Trump mobilizes or demobilizes certain groups, and different polls are built on different turnout models, which are difficult to capture in polls. I have addressed the "likely voter screen" problem before, and Trump just makes it even more difficult.

So, how certain can we be this year?  Uh...

Time for the good ole' 2016 disclaimer...

No comments:

Post a Comment