Thread by @oliverbeige

While we're killing time I can translate and summarize a podcast talk I gave in 2017 (in German) on how Donald Trump confounded the poll aggregators and won the election. Because clearly the message still hasn't gotten across, or we wouldn't still be waiting. 1/♾

The story started for me in March 2016 when a coworker predicted during lunch that Trump (still mostly a joke then) would win the election. I proposed that the only way for him to do it would be to stay on message and at the same time discourage Clinton voters from voting.

If you remember, a lot of pundits assumed that Trump was playing an outsider role to clear the field during the primaries and then would turn more "presidential" in order to appeal to the centrist establishment voters. As has been the masterplan for decades.

This masterplan has a name: median voter theorem, aka Harold Hotelling's spatial competition model, aka "ice cream vendors on the beach". In the ice cream version, you're supposed to ask yourself where two vendors would position themselves if each bather buys exactly one cone.

The mildly surprising answer is that under certain conditions both vendors, or candidates, should position themselves right next to each other, so that each ice cream vendor serves exactly one half of the beach dwellers.

Translated into politics, this means both candidates try to appeal to the median, middle-of-the-road voter as much as possible, and provide stability in the process. In healthy democracies, elections are boring.

But Hotelling's model depends on a number of assumptions: a roughly one-dimensional beach, two candidates, each beach dweller gets exactly one ice cream cone from the closest vendor, no matter how far that is. Once stripped off these assumptions, things turn interesting.

For one, turnout. You might still schlep from one end of the beach to the center for an ice cream, but picking between two interchangeable candidates neither of which will do much for you is hardly worth getting your butt off the couch on election day.

In the US, presidential election turnout usually hovers around 60%, so the winning candidate if non-votes are considered is usually "neither". This is a sizable group of voters to pick up for an entrepreneurial candidate.

So let's split the common left-right dichotomy in two to get: left disenfranchised, left establishment, right establishment, right disenfranchised. This is a pretty rough model, but it'll do for our purposes.

In a typical primary contest a challenger emerges that champions the outsider vote, but after an initial tussle the establishment candidate wins the day and moves on into the main contest. This was the general expectation that Trump upended for the Republicans in 2016.

There were many signals that the popular support for a Republican establishment candidate had eroded enough to make an outside campaign feasible, but clearly most pundits and pollsters didn't see the signals in 2016, and didn't adjust their models accordingly.

If we remember, Trump spent about 15 minutes in the general election trying to look more "presidential" and then went back to his support base, immeasurably helped by Clinton's "deplorables" comment. It was the "47% moochers" of 2016.

But since I'm talking election forecasting rather than politics, how did this affect forecasting? If you assume a Hotelling world with a broad support for centrist candidates and near-full turnout, the decision csn be expressed as "candidate A vs candidate B".

In a world where establishment support erodes and turnout barely crosses 60%, voters face two separate decisions: "candidate A or stay at home" vs "candidate B or stay at home". The old Hotelling choice loses predictive power.

Turnout becomes the all-encompassing factor and the strategy I described earlier: don't try to convince the opposing voters to defect, but rather convince them to stay at home, don't expand your support base, but convince them to vote, becomes superior.

But turnout is at best a secondary consideration in most forecasting models (except the LA Times which forecast a Trump win 2016 based on a stable panel survey), and this has not really changed in 2020.

But what has changed is that the Democratic party has inherited both the remnants of the Bush-era establishment right and has to bridge the gap to the Sanders-voting post-industrial disenfranchised left, a gap that is only to become bigger and will ultimately erode turnout.

With increasing tension between the camps, this remains a formidable challenge for the Democrats. For the pollsters and poll aggregators who did not understand turnout, it means they misinterpreted the role of a small but coherent support coalition and got things wrong. Again.

For anyone who wants to hear me discuss this in German and far more mathematically, the podcast for @modellansatz is here: math.kit.edu/ianm4/seite/ma…

@threadreaderapp unroll plz