G. Elliot Morris posted this embedding of the current Democratic presidential candidates in R^2 on Twitter:

where the edge weights (and thus the embeddings) derive from YouGov data, which for each pair of candidates (i,j) tell you which proportion of voters who report they’re considering candidate i also tell you they’re considering candidate j.

Of course, this matrix is non-symmetric, which makes me wonder exactly how he derived distances from it. I also think his picture looks a little weird; Sanders and Bloomberg are quite ideologically distinct, and their coconsiderers few in number, but they end up neighbors in his embedding.

Here was my thought about how one might try to produce an embedding using the matrix above. Model voter ideology as a standard Gaussian f in R^2 (I know, I know…) and suppose each candidate is a point y in R^2. You can model propensity to consider y as a standard Gaussian centered at y, so that the number of voters who are considering candidate y is proportional to the integral

and the voters who are considering candidate z to

So the proportions in Morris’s table can be estimated by the ratio of the second integral to the first, which, if I computed it right (be very unsure about the constants) is

.

(The reason this is doable in closed form is that the product of Gaussian probability density functions is just exp(-Q) for some other quadratic form, and we know how to integrate those.) In other words, the candidate y most likely to be considered by voters considering z is one who’s just like z but half as extreme. I think this is probably an artifact of the Gaussian I’m using, which doesn’t, for instance, really capture a scenario where there are multiple distinct clusters of voters; it posits a kind of center where ideological density is highest. Anyway, you can still try to find 8 points in R^2 making the function above approximate Morris’s numbers as closely as possible. I didn’t do this in a smart optimization way, I just initialized with random numbers and let it walk around randomly to improve the error until it stopped improving. I ended up here:

which agrees with Morris that Gabbard is way out there, that among the non-Gabbard candidates, Steyer and Klobuchar are hanging out there as vertices of the convex hull, and that Warren is reasonably central. But I think this picture more appropriately separates Bloomberg from Sanders.

How would you turn the coconsideration numbers into an R^2 embedding?