It’s not a bad video either.

The video is guilty of 5 fallacies, which I list in order of appearance:

- The issue-space voter model
- IRV as more incentive-compatible than Plurality
- An “honest” approval threshold that is independent of what other voters are expected to do
- Dichotomous preferences
- The idea that the strategic approval vote is a bullet vote for one’s favorite member of the top cycle

(1), (2) and (3) are forgivable, given that they are popular myths and the video is meant for a popular audience. But (4) and (5) are the narrator’s own contribution, so I’m not sure the video is as informative as it is deceptive.

This is the kind of thing we need more of. I’ve been meaning to make something very much like this, with a 2D plane, except showing more pathologies and more systems, but I am not good at following through on things.

For example, T2R dealing with vote-splitting:

vs T2R failing from vote-splitting:

In what way is #1 a fallacy? The spatial model is the most correct model of real-life voter behavior.

I think the modeling is fairly reasonable as a good approximation to observable behavior. Any simulation can be improved but you need to make some assumptions, and modeling assumptions are not fallacies. I think it’s more questionable to think that this kind of model doesn’t give us a lot of useful information.

Just as an example, it’s totally unreasonable to think that the wavefunction of an electron in an atomic orbital is independent of time, but it makes for a very good model where chemistry is concerned.

First, I specifically criticized the *issue*-space model presented in the video, whereas your source tests a more general *attribute*-space model. Second, your source compares its attribute-space model to four attributeless (not merely non-spatial) models; it does not even attempt to demonstrate that its model’s superiority is due to its spatiality.

The issue-space model is fallacious for several reasons, among them:

- Its “ideal point” makes it dependent on irrelevant alternatives.
- The most important attributes are not usefully plotted on interval scales with voters normally distributed about a center. For example, fitness (also trustworthiness, etc.) is not an attribute the average voter prefers an average quantity of, with extremists preferring very fit or very unfit candidates. Fitness is a pure candidate attribute, lower bounded at zero, for which distance between candidate and voter is not a useful concept, and which is a simple factor of utility (not a term to be squared and added as in Euclidean distance).
- Even those attributes that are best understood as relations between voter and candidate lose their predictive power when anonymized and abstracted from each other as in a spatial model. For example, if I own an oil well, my ideal candidate is for an infinite (or maximally favorable) oil subsidy and a 0% wealth tax. According to a spatial model, for any two candidates that differ only in their proposed rate of wealth taxation, I necessarily prefer the one who proposes the lower wealth tax. But that’s not the case. My ideal candidate proposing no wealth tax is a consequence of his making me rich, without which I look at wealth taxation with very different eyes.

You can still map fitness to one of the axes in preference space; it’s just that every voter’s preference is at one extreme of the scale. If you’re electing a chess club president, for instance, and every voter ideally wants a candidate with maximal chess-playing ability and maximal public speaking ability, voters will still have to trade off between those two axes, if no candidate perfectly maximizes both. If you plot them as X and Y, for instance, and then project it at a 45 degree angle, it becomes a chess ↔ public speaking trade-off axis, with voters having an ideal point somewhere in the middle.

I don’t understand why. If Candidate A is (100, 1) in the (oil subsidy, wealth tax) space, and Candidate B is (100, 2), and your ideal point is (100, 0), then you prefer Candidate A. You’re saying that whether you care about the wealth tax axis depends on whether you get what you want on the oil subsidy axis, so they aren’t independent?

But that’s not the way fitness actually relates to other attributes. Fitness has a *multiplicative* relationship to other attributes; it *multiplies* the utility of the candidate’s goals. In the extreme case, if he’s absolutely unfit, his goals are irrelevant; he won’t get it done. That’s very different from a spatial relationship, in which utility loss is the square root of the sum of the squares of the utility losses for the issues abstracted from each other. What is the basis for the assumption that that, Euclidean distance, is a more useful way to aggregate attributes than simple multiplication or anything else?

Of course they’re not independent. But it’s not that I don’t care about the wealth tax if I don’t get what I want on the oil subsidy. It’s that the utility of the wealth tax is inversely related to the size of the oil subsidy. If the oil subsidy is low enough (certainly if it goes deep enough negative, i.e. becomes a big enough tax, but possibly even when it’s 0 or even just rather low), I’ll be poor enough to *want* the wealth tax. To put it in your coordinate terms, if Candidate A is (0,1) and Candidate B is (0,2), I prefer Candidate B, despite my ideal being (100,0), because ideology is not decomposable into the independent dimensions spatial models invariably assume.