Time for a another iteration of the Polling Post Mortem.

While polling this election cycle would be challenged to rise to the level of “joke”, Wisconsin is an interesting test bed to evaluate polling in general this election cycle. The general defense of polls has been that there are too few of them in any state and more polls should give you a better picture of what is happening. One would assume that any competent poll would look a lot like another because, really, is there any danger that water boils at a different temperature in your garage than in your kitchen? Wisconsin was polled to death, nine in the past three weeks, and so if there was ever a place where the “more polls increases accuracy” argument would work it would be a relatively small and highly polled state like Wisconsin.

Before we go on, let’s look at last night’s result:

Crap. That is so sweet I have to post it again:

Sweet, huh?

This is a matrix of poll results. RED indicates the number is wrong outside the margin of error (MOE column) of the poll. Yellow indicates the order of finish is wrong. From my point of view, an accurate poll must a) call the finishers in the correct order and b) hit the final vote within the margin of error. For this exercise I’m also accepting vote spread within margin of error.

(right click, select open image in new tab, for larger chart)

Overview

Of 27 possible correct answers (9 polls time 3 candidates), only 10 were within the margin of error. When you look at the error for each candidate using the absolute value of the number rather than a raw average, Cruz was, on average, 10 points off in polling. There is not a single case where a poll gave Cruz a higher number than the actual vote. Trump was off by an average of 3 points, making him the most accurately polled candidate. Kasich poll, on average 6 points higher than his vote total.

The polling averages both and RCP and Huffington Post show the fallacy of poll averages. By throwing in horrible polls, like ARG and Optimus, with okay polls you completely dilute the picture of what is happening. The poll averages show a Donald Trump surge this week. That is simply bullsh**.

This is pretty much in line with what we’ve seen all cycle. Cruz routinely does about 5 points higher at the ballot box than he does in polls. Trump and Kasich are the opposite, polling better than they perform.

The Best

There was one poll that could be considered accurate, CBS New/YouGov. It hit the order of finish and all the candidates within a healthy 5.7 point margin of error.

Sorta Okay

Fox Business and Marquette get honorable mention. They caught the vote spread within the margin of error.

The Worst

ARG was a laughable poll. This was part of the reason I wrote this story on Monday. If their polling director has any sense of shame whatsover he or she committed seppuku last night. They missed the finish order and the percentages were off by orders of magnitude. They were one of two polls that called Trump as the winner. And they don’t even have the defense of being an older poll. Their polling should have captured all the crap of Trump’s last week.

Runner up for the worst is Optimus. It used the largest sample (6183 LV) to produce results outside the margin of error for all candidates, it got the order of finish wrong, and it under-polled Ted Cruz by nearly 22 points.

Summary

Polling this election cycle has been generally horrific. There are probably all kinds of reasons but the underlying one is that the pollsters simply are not very good at what they purport to do. I’m not a late-comer to the poll-scoffer community. I don’t know that I’m a charter member but my membership is long-standing. Two years ago, I post on why pollsters completely blew the 2014 races. I re-read the piece today and wouldn’t change a word.

If polling was going to be accurate anywhere it should have been Wisconsin. It is a relatively small state. It has a fairly stable population. It was polled like there was no tomorrow.

The bottom line is that polls are guesses. They aren’t scientific guesses. They are just guesses that use statistical symbols and mathematical mumbo-jumbo to gussy up that guess.