Sunday 13 December 2015

Pollsters Modelling Stupidity

Very occasionally you come across an article so objectionably dumb that it sticks in your head. For almost two months a piece by the name of Forget the debate: Two simple reasons a Republican will likely win in 2016 has been quietly exercising me because it is stupid. The authors are two top American pollsters who should know what they're talking about but, unfortunately, they don't. The reason they fall into error is because they lack a sociological imagination.

In their piece Clifford Young and Julia Clarke confidently predict that the Republicans will win next year's presidential election. This is what they have to say:
First, a Republican will win because voters typically shy away from the party currently in power when an incumbent isn’t running. In fact, a successor candidate is three times less likely to win. Second, President Barack Obama’s approval ratings are too low to suggest a successor candidate will take the White House.
They go on ...
Why are we so confident, especially when opinion poll data now gives Hillary Clinton the edge over most Republican opponents? The simple answer is that we’re relying on models, not polls.
Uh-oh. Models.

Young and Clarke have made predictions based on the outcomes of 450 elections, and found a tendency for incumbents to win elections. In fact they're three times more likely to win than challengers. On the other hand, if we have a situation the US electorate will be facing next year - where the president is retiring at the end of his term limit - challenger parties are three times more likely to defeat the incumbent party. They also argue that a successor candidate can overcome this bias if the outgoing president has high approval ratings. If they're scoring a 55% rating then the successor nominee has an even chance of winning. Presently, Obama's ratings are knocking around 45% giving the Democrat candidate - they calculate - a 14% chance of winning. They concede the prediction could be confounded by a surge in Obama's popularity rating (which, in fact, would confirm the model if the Democrats retain the White House), by a high turn out, or by a "wild card" candidate.

To put it in a single word: baloney. We may be living in the era of big data, but the behaviour of huge numbers of human beings is not the same as the predictable grinding one would find in a mechanism. An election is not like a clock, and expecting one to conform to the parameters of a model are going to be very disappointed. Young and Clarke are right in one respect, human relationships can be modelled as probabilities, but prediction can only be made with confidence if it is constantly fed with fresh data. For example, the weather forecast is always based on the probable directions of wind, cloud formations, zones of high and low pressure, and so on. They tend to be accurate because they're constantly feeding in data gathered by meteorological stations and satellites. It's why forecasting is usually more right than wrong. To use this as an analogy, what Young and Clarke have done with their model is similar to Carol Kirkwood looking at last week's weather charts and extrapolating them into the future. A totally nonsensical approach.

Given their article was written and published in October, Young and Clarke would surely not have missed the rise and rise Donald Trump. If I was interested in working out the probability of his taking the White House, I wouldn't be messing around with results from past presidential elections or analogous contests overseas. I'd be digging into current opinion polling, examining demographic changes, considering the consequences of a split Republican party, and so on. In other words, the answer lies in analysing the situation in front of us and drawing conclusions on that basis. Young and Clarke however believe as if their model has some sway over living, breathing, shifting social processes. I can tell them now, the probability successor candidates are more likely to lose than the incumbent is an artifact of the model, it is a statistical conclusion based on after-the-fact analysis. It has no bearing at all on future outcomes. Just as, for example, it had been established as a "law" of British elections that no incumbent Prime Minister increases their party's vote share. This again was an inference drawn on past behaviour. It did not structure future behaviour as the overall majority won this May by the Conservatives demonstrates.

Young and Clarke have made the elementary scholastic error of confusing the things of logic with the logic of things, as our friend Marx once put it. And there's really no excuse for this sort of sloppy thinking among so-called polling professionals, especially after establishment economics failed to see the crash coming because they preferred the company of mathematically elegant but useless models of markets to the analytical grind of watching the real thing. On this basis, it's safe to ignore this pair of "respected" pollsters. Trump may win the Republican nomination, and should he do so the odds of his candidacy crashing and burning are significantly greater than 14%.

3 comments:

jim mclean said...

Must admit have been following the Presidential race with interest and I have yet to see any evidence supporting an alternative to Hilary Clinton as the next President. Sanders blew it by concentrating on the affluent working class, a great support, but Clinton has the "minorities" signed sealed and delivered and the rest of the "white" vote. The fact that Hilary is the bottom of my like list does not make me a happy bunny, but she is the next POTUS. The Republicans will have to reform as a party and they may make it next time round.

Robert said...

President Trump - it would be fun, if we had a rewind button.

jim mclean said...

Washington renamed Trumpton