The electology.org poll, part 2/3: Feeling the Bern? Or feeling the Johnson?

This is part 2 of a series of 3 posts by Jameson Quinn, based on a pre-election poll run by electology.org and GfK research. Part 1 introduced the poll and analysis methodology, and discussed the lessons of the poll regarding the two-way Hillary/Trump race. This part talks about what the poll tells us about support for other candidates or potential candidates. Part 3 will talk about the implications for voting systems.

I want to clarify especially that I’m speaking for myself in these posts. Specifically, Electology (aka the Center for Election Science or CES) is a nonpartisan organization; our board members and supporters include active and committed members of all of the top 4 US parties (Republican, Democratic, Libertarian, and Green). I honestly believe that our proposals for voting reform would bring the kind of positive-sum changes that could be win/win for all of these groups and more. We value our nonpartisan position, and any statement I make here that’s partisan is NOT the position of the organization.

The focus of electology’s part of this poll was on alternative voting methods. To that end, voters were asked about how they would have voted under 3 such methods (approval voting, 0-5 score voting, and IRV), as well as under plurality (aka First Past the Post). They were also asked for their honest opinions of the candidates on a 0-5 scale. Interviewees were randomly given either a short list of 4 candidates — Clinton, Trump, Johnson, and Stein — or a longer list of 9 possible candidates — Clinton, Sanders, Trump, Cruz, Johnson, Stein, McMullin, Bloomberg, and Castle.

In this installment, I’m focusing on the question of how people felt about all of these candidates, not on the effects of voting system per se. To that end, I’ll look at the results from the long list of candidates, under approval and score voting. In the next installment, I’ll dive much further into the differences between the long and short lists, the various voting methods, and what that shows. (Teaser: plurality still sucks and needs reforming, but the other lessons are not entirely what I’d hoped for.)

So, here’s the flashy, yet controversial, picture. (I know that there are a lot of caveats and disclaimers with this picture, and I’ll try to cover them below. And I’ll still probably fail to give all the caveats I should, so you can yell at me in comments if you want.)

What are you looking at? I made demographic models, as detailed in the previous post, for the score results of all 9 candidates mentioned above. Each model used ordered logistic regression; I’ll explain what that is below. Then I applied the models to each state’s demographics, as given by the Current Population Survey, and voting propensity by ethnicity, gender, and state from 2012. I then corrected the Clinton and Trump results by the factor by which my plurality model from last post got the state wrong; and the other candidates’ results by those factors for Clinton and Trump, each raised to the power of the correlation between the major-party candidate’s simulated plurality total and the minor candidate’s simulated score voting total. (The sum-absolute-value of those correlations was between 0.5 and 1 for both Cruz and Sanders; and none of the other candidates came close to winning anywhere).

Or, in plain English, I used the survey results to model how each state’s population would have responded to the score voting question, and then rescaled the results to agree better with the actual voting outcome.

Warning: Poll results do not guarantee contrafactual performance.

If you’re wondering about the strange empty area in the Midwest: that’s because nobody actually lives there. That is, the states above are scaled proportionally to their electoral votes. (If I’d scaled them to their population, Wyoming would be microscopic.) This excellent map format is from Daily Kos Elections.

Of course, the obvious disclaimers apply. I’m aware that this involves unfair comparisons. Clinton and Trump went through bruising general election campaigns and both got large amounts of primarily-negative media coverage; Johnson and Stein were mostly not treated seriously; and during the general election campaign, the other hypothetical “candidates” like Sanders basically only got attention from their supporters. I’m happy to discuss why or how that happened in the comment section, but for now: yes, I am fully aware that if Sanders had been the nominee or one of many nominees, this would not have been the map.

Furthermore, if the real-world election were being run using score voting, voters would almost certainly vote more strategically than they do on a poll. (I’ll talk more about that in the next installment).

Specifically regarding McMullin: my model would probably underestimate the support of regional candidates like him, so it’s possible that a more perceptive model would give states like Utah or Nevada to him.

And finally: even if you ignore all the above disclaimers, this map above is a visualization of what “might” happen if we kept the electoral college but used score voting in each state. That’s a crazy hybrid of election methods that nobody actually advocates. We need to get rid of the electoral college. (I’ll talk more about this point too in the following installment.)

Still. Those disclaimers notwithstanding, I find this map very interesting. The electoral totals would be: Sanders 280, Trump 174, Clinton 73, Cruz 11. Sanders wins this crazy hybrid election outright.

This map definitely suggests to me that Sanders would have had an easier time beating Trump, in the EC as well as in the popular vote, than Clinton did. Again, if he’d actually been on the ballot and faced a negative campaign from the Republicans, I imagine he’d have done worse than this suggests (though remember, in a two-way race against Trump, he gets almost all the Clinton EVs above too, because he’s in second place there). But on the other hand, I can’t imagine that even without the bruising general-election coverage, Clinton could really win states like Nebraska, West Virginia, or even perhaps Ohio.; so I think this map does actually demonstrate he could have had some strengths she didn’t. I understand that there’s room for reasonable disagreement on this point, and I expect some strong push-back in the comments.

So, should he have been the Democratic nominee? Depends what you mean by “should”. I’m not saying that we should have ignored the primaries. Clinton won them fair and square, and I think that somehow ignoring that victory would not have been a good idea. Her supporters — many of them Black Southern Democrats — should not be silenced; and much of Clinton’s weakness has to do with the gross unfairness and sexism of the media and electorate. But if it’s just a matter of who had a better chance of winning, I think that even subtracting a few states at the margin, this map looks very good for Sanders in a two-way race against Trump.

Clearly, this finding (or at least, suggestion) has strong implications for the potential impact of different voting systems. But I’ll save those for the next part.

What else does this data tell us about candidates besides Clinton and Trump? Well, for one thing, neither Stein nor Johnson come close to winning a single state in any voting system in my model.

We could ask about how Stein and Johnson affected the major candidates.

Note that this data set does not allow me to give any definite answers to that question. That would of course be true for any counterfactual; I’d always run into the fact that how people say they will vote on a poll and how people actually vote are two different things. But when it comes to voting strategy — whether people are willing to “betray” a third-party candidate they actually prefer in order to cast a useful vote for a major candidate they like less — the difference between polls and reality is especially important.

So, in order to address the question of what would have happened without Johnson and Stein in the race, I have to make some assumptions:

1. Tendency to support a given third-party candidate — Johnson or Stein — exists on a continuum, and the multiple alternative election method questions in our poll can give a good idea of where each demographic subgroup lies on that continuum.
2. Voting for a third-party candidate in real life can be predicted by where a group lies on that spectrum. That is, insofar as different groups are more or less likely to strategically vote for whomever they see as the “lesser evil” main party candidate, those differences will show up in the poll.
3. Similarly, voting or not voting is governed by some overall tendency, and the polling questions give an accurate picture of that tendency for each demographic subgroup.
4. The association between voting likelihood and likelihood of support for a given third-party candidate is accurately shown by the polling data, and is itself constant across all demographic groups.

The above assumptions can be put in much more technical form. But I slogged through all the math and programming (believe me, it was no easy slog) to get some simple answers, and here they are:

Again, getting these numbers required some pretty strong assumptions, so like the map above, I wouldn’t trust them overly much. But they are one plausible, evidence-based counterfactual outcome. And we do know how many Stein and Johnson voters there actually were in each state, so the statewide totals do come out right. And the answer is, Johnson had several times the impact of Stein, and Stein probably “spoiled” only Michigan.

…

I was trying to write a nice explanation of ordered logistic regression here, but I realized that it would be better if I just posted that in the comment section.

Final note: this poll cost almost $10,000 for electology, and the analysis behind these three posts took significant unpaid work and expertise too. If you think this is worth it, consider donating to electology.org to help us support better voting methods in the US and around the world.