Should the votes from voters combine on a per-issue basis, rather than a per-party or a per-candidate basis?

Why did Kenneth Arrow think that Approval Voting would do a better job of bringing to the surface the real concerns of voters?

A person says they need a job. Various political groups step forward with proposals. One side says that the government should make massive investments in infrastructure and education and science research and this will boost the economy and provide everyone with a job. The other side says it will offer massive tax cuts to encourage business and consumers, and this will boost the economy. Who is correct? Even professional economists sometimes disagree. Can we build a system that discovers the truth without anyone having to be aware of what the truth is? Perhaps we can borrow a few ideas from Machine Learning, or perhaps from general information processing? Every citizen has some information, but is there a way to aggregate that information together in a way that arrives at the right answer, even though no one person knows the right answer?

If we wanted to borrow some ideas from Machine Learning, the essential processes we’d want in such a voting system are:

1. frequent voting (perhaps monthly voting)

2. a very large number of initial options (candidates)

3. the ability to boost many signals, so that entire complex patterns of signals can emerge in fine-grained detail

Please note: we are not talking about using a computer to replace humans. Elections can still happen the old-fashioned way: people go to their polling booth and fill out some paper ballot. But we are looking for a system of voting that does a better job of combining the wisdom of all of the different voters, so that the system ends up being wiser than any one voter.

We could probably design many systems that meet our goals, but my suggestion here illuminate some of the basic things that the system will need: frequent elections, huge numbers of candidates, huge numbers of votes.

Who is Kenneth Arrow? What is approval voting?

Back in 1951, Kenneth Arrow published his Impossibility Theorem. He showed that most of the common forms of voting have some serious problems. He actually said that perfectly aggregating the preferences of voters was mathematically impossible, in the sense that every system of voting has at least a few wild edge cases where the result will seem crazy. In particular, any kind of rank voting, where the voter is trying to pick the top rank candidate, will be flawed.

He was more optimistic about approval voting, where voters get an infinite number of votes (or rather, as many votes as there are candidates), and only the number of candidates elected is finite.

Some systems of voting work well when there are only 2 parties, but then they malfunction when there are many parties, and other systems do well with many parties, but can be sabotaged by “strategic” voters. Read up on all the difference systems.

The results are summarized in this chart:

You can see that approval voting and score voting are the best, in that they are furthest to the right, and fairly high up. Score voting means a voter gets 10 points and can use them to vote for the different candidates, giving perhaps 4 or 5 points to one candidate, and just 1 or 2 points to some other candidate, and 3 or 4 points to another.

(STV is another good option. However, it too has some problems: “STV is a system designed for political scientists and mathematicians, not voters....The local government elections were a disaster and an international embarrassment...” -- New Zealand National Party MP Nick Smith, November 3, 2004)

As a practical matter, there isn’t much difference between approval voting and score voting — both are better than our current system of voting, both are vulnerable to minor distortions. Score voting might be slightly more accurate — but keep in mind it very much resembles the 5 star rating system that Netflix used to have for movies, and Netflix had to abandon that system because people were unable to figure out which movies they liked, on anything like a consistent basis. One month people would be happy and then they would give a lot of movies high scores, but a month later they would be unhappy and give a lot of movies low scores — people’s moods change all the time and so people are unlikely to make consistent scoring decisions over time — so the Netflix recommendation engine was unable to pull a reliable signal from the 5 star rating system. Eventually Netflix switched to a simple up-versus-down rating system. In other words, with years of data and tens of millions of users, Netflix found that approval voting was actually more accurate than score voting. And approval voting is slightly simpler, which might be why people use it more effectively. But in the grand scheme of things, both of these systems of voting are good enough that all of the other problems with our democratic system would make the differences between these two systems completely irrelevant. And both of these two systems of voting run the risk of establishing a “tyranny of the majority,” therefore both would have to be used in modified form.

But for now, let’s focus on approval voting, and let’s see how it might help boost crucial information signals from the voters. Let’s assume a scenario where elections are held every month and where there are 100 candidates and so every voter can cast up to 100 votes.

What is wrong with voting for one candidate?

[[Edited to add: Achen and Bartels suggested that most people vote primarily based on obvious markers of identity, such as race and religion, so in opposition to that sad fact, this next example explores the possibility that a system of voting can discover what people’s real policy concerns are, even when the voters are still biased to first vote on the basis of race and religion.]]

To be clear about why the current system fails, assume a voter who is a Catholic Hispanic male working a low-wage job, who perhaps wants a candidate with these attributes:

1. Hispanic

2. male

3. Catholic

4. supports a higher minimum wage

But instead there is a candidate who is:

1. Hispanic

2. male

3. Catholic

4. a businessman who supports a lower minimum wage

So the voter votes for this candidate because this candidate has 3 of the 4 attributes that the voter is looking for. One attribute is forfeited. If the voter could vote for 50 people a month, then the voter could find different candidates who separately have the attributes that the voter wants. That is:

1. one candidate might be a Hispanic

2. another candidate might be Catholic

3. another candidate might be male

4. another candidate might support a higher minimum wage

There is no longer the need to find every attribute in one candidate.

For a particular voter, there is no particular advantage to this system, but for society there is a big advantage. Every voter has some crucial information about the real status of society, and collectively they have more information than any one person in the system. If they can only vote for one candidate, then they are sending one signal, a very muddy signal that tries to bundle up many concerns. But when the voters can vote for many candidates, they can send separate signals for each of their different concerns. The system itself then aggregates those different signals together, on the basis of issues, rather than aggregating those votes on the basis of a political party or a candidate. This simply happens organically, with no one forcing the process.

The crucial thing is how “many votes” from “many voters” combine

Assume a neighborhood with a large number of poor voters, some of whom want to vote for a Catholic, some of whom want to vote for a Protestant, some of whom want to vote for a Jew, some of whom want to vote for a Muslim, but all of them want to vote for a higher minimum wage. A system where they can send many signals better allows the importance of the higher minimum wage to come through the political system. None of this is meant to suggest that one’s religious identity is unimportant, but again, the goal is to aggregate the concerns of voters so that their most prevalent issues rise to the top. In this case, the desire for a higher minimum wage is more prevalent than the desire for a candidate of any particular religion, so that is the signal that comes through.

If we were serious about borrowing some ideas from Machine Learning, to build a system that is truly smarter than any of the individuals in the system, it would be important to have multiple rounds of such voting, since each round would allow the strong signals to get boosted. In the world of software, it is trivially easy to run such a process for hundreds of rounds. Sadly, that isn’t possible in the real world (except maybe over time, which is another argument for monthly voting). But we can perhaps get two rounds, which should improve the signal. The other day we suggested an intermediate mass assembly of citizens (Notables), between the public and the legislature. If the voters used approval voting to elect Notables and the Notables used approval voting to elect the legislature then we will have built a political system that is capable of boosting the most widespread and urgent issues.

This is a subject where I have seen no research, and where I think we could learn a lot. Take all of the various voting systems (STV, score voting, approval voting, FPTP, MMP) and ask, how do they chain together? What is the output after multiple layers of representation? Do the errors multiply or sublimate? Surely someone is looking for a PhD thesis?

Can approval voting reveal the best economic policy, even when professional economists disagree?

Going back to the example in the first paragraph of this essay, can the right economic policy for a recession be discovered, even when professional economists are disagreeing? Well, different policies could be tried: massive investments, perhaps some tax cuts. People would see how those policies effected their income. They wouldn’t have to understand the wider context, they’d only have to know if their personal situation is getting better or worse. If things are getting better for them personally, they can vote for more candidates like the candidates they voted for previously. If things are getting worse for them personally, they can vote for different kinds of candidates, perhaps belonging to different parties. They can also cast votes on economic issues that are separate from the other things they care about. If the voter is strongly anti-abortion, they can remain strongly anti-abortion when the economy is good and when the economy is bad. They can vote for candidates who are anti-abortion when the economy is good and when the economy is bad. But if things are bad for them personally, they can also give some votes to candidates who are advocating for economic policies that are different from the economic policies that were advocated for by candidates that the voter used to vote for. So with approval voting, the system becomes more fluid, as the voter can more easily change their mind about economic issues, without also having to change their mind about something like abortion.

So here is an argument in favor of approval voting that I think is somewhat original: more than most voting systems, it can be tweaked to work a bit like Machine Learning, enabling a system that is much smarter than anyone in the system. I’m guessing this is why Kenneth Arrow had such positive feelings for it.

If we combine approval voting with monthly voting, then we give voters a way to send signals quickly, based on new information.

There is, of course, the counter-argument: that if we elect politicians for long terms (11 years) then monthly voting only changes a tiny percentage (0.7575%) of the legislature. With 11 year terms it takes 5.5 years to replace the majority. And it is valid to wonder whether the old majority would give a damn about the new crop of politicians who are being elected, while that new crop of politicians is still a small minority.

Then again, there is a counter-counter-argument, which is the speed at which change can begin to happen. In the USA, Herbert Hoover began his second term in March of 1929, the Great Depression began in October of 1929, the voters were not able to put Democrats in the majority of the House till March of 1931, and a new President only took over in March of 1933. With monthly voting, presumably the shift in mood would have begun to register as early as October of 1929. If the opposition party had 40% of the legislature (as the Democrats did) then they would have regained the majority by the spring of 1931, the same as what happened in real life.

A final point, in a world of monthly voting, do we need people to go out into the streets, and stage a big public protest, and demand change from the government? In the current system, when there is a recession, we expect politicians to make a decision about policies such as whether they should fight the recession with tax cuts or massive investments. The politicians will be influenced by various groups in our civil society, such as the newspapers, the television news, public protests, and lobbyists from industry. Aside from protests, the public has little direct input. And arguably, the need for those public protests is a failure of the system. By contrast, with monthly voting there is no longer a need for protests, since whatever rage the public might be feeling they can simply express it in that month’s vote. Just a thought.

(But keep in mind, my co-writer Kathryn Bertoni had some doubts about monthly voting, which she posted in the comments under the essay about monthly voting.)

To summarize:

Approval voting can give us something that in some ways resembles Machine Learning, at least in the limited sense that we can build a system whose output is much smarter than any one of its inputs. Given a huge number of candidates, and frequent voting to enable fast updates based on new information, and a certain amount of signal boosting through multiple levels of representation, then it should be possible to have people’s votes reveal the right policy, and this should happen even when none of the voters know what the correct answer is.

Approval voting limits the ability of politicians to manipulate voters

So far I’ve written of this approach as one that better aggregates the top interests of voters across multiple dimensions. But we should also consider the healthy effect this has on politicians. When voters only have one vote it becomes easy for politicians to cynically manipulate voters around a single issue.

For instance, I have known many Catholics who oppose the death penalty and want to see the government play an expanded role in supporting the poor. Also, they oppose abortion. In the USA, the Republicans have managed to use the abortion issue to gain votes so as to gain power so they can make it easier to put prisoners to death and so they can tear down any program meant to support the poor.

Under approval voting, the Catholics could vote for:

1. a candidate opposed to the death penalty

2. a candidate who wants to expand social programs for the poor

3. a candidate who opposes abortion

Their votes against the death penalty would then aggregate with those votes from other people who oppose the death penalty.

Their votes for expanded social programs would then aggregate with those votes from other people who want expanded social programs.

Their votes against abortion would then aggregate with those votes from other people who oppose abortion.

And so it becomes impossible to manipulate a voter around a single issue — rather than capturing 100% of a voter’s vote, a given issue now only captures a small fraction of a voter’s vote. Therefore a political party can no longer use one hot-button issue to gain power and carry out a program that many of their voters are opposed to.

While this is a complex topic, we can say rather simply that in a world where voters get to vote for many candidates, the Republican coalition that has existed for the last 50 years simply never would have existed.

When votes combine on a per-issue basis, the system is smarter

Some people are under the mistaken belief that the political system, when it delivers dysfunctional results, does so because many voters are poorly informed. For instance, after a strange election, some group of pundits might be left wondering “Why did the working class vote against its own interests?” And yet, Achen & Bartels attacked the idea that there are any high-quality voters. The best educated people are still ignorant, they said. We are all bad voters. The system itself makes us bad voters. But if we design a better system, then everyone can be a good voter, in the sense that every voter can send a signal that can be combined into a stronger signal that expresses an important truth about the circumstances that society now faces. Voters do not need to “understand the issues” in the traditional sense. They only need to send some signals, based on their own experiences — they do this simply by voting for various candidates. And those votes should elect those candidates who are speaking to the broadest concerns of the voters, and this should lead to a system that is much smarter than any one person in the system.