Donald Trump’s victory was a close-run thing, achieved by the grace of winning margins of 1.2% in Pennsylvania, 1.0% in Wisconsin and 0.3% in Michigan. The New York Times is still projecting that Hillary Clinton 0.2% lead on the raw popular vote will resolve at 1.2%, which presumably has something to do with an incomplete vote in Washington state. Nate Silver has offered a penetrating counter-factual to illustrate how different the result and its implications would look if things had played out just a little bit more happily for Clinton.
Assuming the New York Times projection is right, the crude average of national poll results conducted by RealClearPolitics will prove to have exaggerated her margin by 2.0%, which is quite a bit less bad than the news media consensus would have you believe. Not a few individual pollsters have questions to answer, but the shellacking the industry as a whole is copping is out of proportion to its failure. A similar error in the context of a landslide result would barely have attracted notice, but in this instance it had the effect of encouraging a misapprehension about the likely result.
Of the aggregators, Nate Silver emerged with the most dignity intact because it was his model that had the highest uncertainty factored in. So far as the predicted electoral map, Silver’s model was identical to that of Sam Wang of the Princeton Election Consortium, which fatally estimated Clinton’s chances of winning at 98%. All it took to make a mockery of this was a regional failure of polling in the rust belt states, where the fierce enthusiasm for Trump among normally disengaged white working-class voters was under-reported. Pennsylvania, Michigan and Wisconsin produced 62 polls of presidential voting intention between the start of October and polling day, and only two had Trump in the lead (both conducted on the eve of the election by the Republican-aligned Trafalgar Group).
My own poll aggregate did no better or worse than the others, having won brownie points by tipping Florida correctly, but lost them again by giving Nevada to the Republicans (by insignificant margins in each case). Its findings are of use in ascertaining where the state polls did best and worst, as detailed in the list below showing the size of the error in favour of Trump (so negative results mean the polls were biased to Clinton). This is limited to states where three or more polls were published in October and November. Two patterns emerge: the noted under-measurement of the Trump earthquake in the industrial mid-west, and a tendency for both sides’ winning margins to be underestimated in their strongest states.
Crikey subscribers can enjoy more of my thoughts on these matters here.
North Carolina -3.8%
New Hampshire -2.2%
New Jersey +1.3%
New York +2.3%
New Mexico +3.2%