Dynamic Admissions Markets

School-choice markets are like any economic market in that they are characterized by prices (score cutoffs), supply constraints (school capacity), and consumer demand (student interest). In general, a t√Ętonnement or price-adjustment process converges quite well for school-choice markets when students’ preferences follow a multinomial logit choice model and their scores at each school are independent or positively correlated.

A unique challenge in school-choice is how to model the preferences of schools over students and vice-versa. In this example, I use a standard model for student preferences known as the multinomial logit choice model. To express the fact that student preferability is imperfectly correlated between schools, I assume schools score students using a convex combination of orthogonal dimensions of student quality. For example, each student may have a language score and a math score, and the school’s decision is how much weight to place on each of these and what minimum score it should require for admission.

(In reality, if the language and math scores correlate, we can use principle component analysis to extract the orthogonal components of the score vector, so the only loss of generality here is that we have limited the score dimensions to two. The number of dimensions used can be increased, but at considerable computation cost.)

The first pane of the animation visualizes the possible sets of admitted schools and associated probabilities. A student who scores high on both tests will be admitted to both schools, while one who scores high on the math test but low on the language test may only get into the Antarctic Institute of Technology, and so on. Each school starts out with a random score cutoff and adjusts it over time to make the number of students who enroll agree with its capacity—when this occurs, we say the market is in equilibrium.

At equilibrium, the total percentage of students who receive no admissions offers (the white area of the graph) must equal 0.4, which is one minus the sum of the schools’ capacities (expressed as a fraction of the number of students).

The second pane shows the motion of the cutoff and demand vector over time. The gray pointers indicate the excess demand at those cutoffs; this graph thus shows that the equilibrium is stable, because all the integral curves lead to it. I configured the step parameter in this example so that it converges to the equilibrium gradually, but with better parameters we can use the t√Ętonnement process to find the equilibrium quite quickly, even in large markets with more schools and tests.

You can find this example in self-contained form here on GitHub. A collection of equilibrium-finding algorithms is part of my DeferredAcceptance Julia package.