← Back to Econ

Philadelphia's school match is less stable than advertised

In Philadelphia, if your kids want to go to a public school outside their neighborhood catchment, they have to enter a lottery. Since the 2024-2025 cycle, the Philadelphia School District has placed students using the “Deferred Acceptance” (DA) algorithm. This is a modified version of the matching algorithm that won Lloyd Shapley and Al Roth the Nobel Prize in 2012. The matching itself is run by Avela, a company founded out of MIT in 2019.

The District’s advice to families is simple. “There is no way to ‘game the system,’” so families “should always rank schools according to their true preferences.” This is paraphrasing the main mathematical claim about this matching scheme. The deferred acceptance algorithm produces a stable assignment, i.e., if all students report their preferences truthfully, then no student will be shut out of a school she prefers while someone with lower priority (i.e., a worse lottery number) gets a seat there, and in addition, ranking honestly is every student’s best strategy. These claims are formally proven in Abdulkadiroğlu and Sönmez (2003). But these proofs are in an idealized world where we assume:

  1. Students rank every school they would accept
  2. Students only care about which school they get (i.e., they don’t care about which schools their friends, neighbors or siblings get).

The Philadelphia School District’s lottery only allows students to rank 5 schools, which breaks assumption 1. And real students actually care about where their friends and siblings are going, which breaks assumption 2.

Five choices

In the Philadelphia School District’s lottery, each student may submit at most five ranked applications. A student who would happily attend any of six schools has to drop one, and the stability theorem says nothing about schools left off the list. Say her five choices are all long shots and she gets none of them. It’s possible that the sixth school she would have taken admits a student with a worse lottery number. This is exactly the situation that is claimed to be impossible by the stable matching theorem.

Once you lose stability, honest ranking also stops being optimal. A family that lists five long-shot schools in honest order can come away with nothing. A strategic family now should mix in “safe” schools into their top 5 even if the safe schools are not in their top 5 choices. This type of problem was known in the academic literature. So it’s a little disingenuous for the district to claim that students should rank schools according to their true preferences.

Friends in the same grade

The other problem with the matching algorithm as implemented by the Philadelphia School District is that it doesn’t take into account groups of applicants applying to criteria-based schools. Suppose what an eighth grader wants most is to land at the same school as her best friend. The application has no way to say so. Each child is assigned an independent, random lottery number, and at criteria-based schools even twins are processed as unrelated applicants.

In this situation, the matching algorithm also loses the stability and strategy-proofness that make it so appealing in the simple case.

Consider two friends, Maya and Jordan, who want to end up at the same school. For School A, each of them gets in with probability 1/2, and since they have separate lottery numbers, their chances are independent. School B is a sure thing for both of them. Both friends prefer A to B, but each cares even more about being with the other. Say being together at A is worth 4 to each of them, together at B is worth 3, alone at A is worth 2, and alone at B is worth 1.

Each friend can rank honestly (A first, then B) or put B first, which guarantees a seat at B. Their expected payoffs, with Maya’s listed first, are:

Jordan ranks A, then BJordan puts B first
Maya ranks A, then B2.5, 2.52.5, 2
Maya puts B first2, 2.53, 3

If both rank honestly, a quarter of the time they both get into A, a quarter of the time they both miss A and land together at B, and half the time the lottery splits them up. Both friends do best by putting B first, giving up their shot at A to make sure they end up together. When the lottery does split them, the friend at A would rather be at B with the other, and B had room for her, so the outcome is not stable (breaking the other main claim of the ranking algorithm).

The friends also have a coordination problem. If Jordan ranks honestly, Maya’s best response is to rank honestly too, because putting B first on her own leaves her alone at B whenever Jordan gets into A. They only reach the better outcome if they both agree to put B first. With these payoffs, honest ranking leaves both friends worse off whenever A’s admission rate is below 3/4.

This example is inspired by a 2019 paper by Dur and Wiseman on “school choice with neighbors.” In that paper “neighbors” are any two students who want to attend the same school (i.e., friends). In the paper they show that when students want to be with their friends, no stable assignment may exist at all, and no strategy-proof mechanism can find one every time one does exist. If you’re familiar with the famous medical-residency matching algorithm, you’ll know that that algorithm does handle married couples, but that works because couples are a small fraction of the applicants. Dur and Wiseman call that assumption “unrealistic for school choice in an urban setting where many children live close together.”

What would help

Letting families rank more schools would reduce the first issue. For the second, Dur and Wiseman propose a modified algorithm in which a pair applies jointly with the priority of its weaker member. This ensures that no student loses a seat to a pair unless both friends outrank her. Philadelphia already does something in this spirit for twins at catchment and citywide schools. For those catchment and citywide schools (but not criteria-based schools), siblings who apply together and rank all five schools identically, receive sibling preference. Extending some form of sibling-preference to criteria-based schools doesn’t seem like it would be too hard for twins (who typically apply together), but extending this to friends might introduce lots of coordination problems (especially when friendships are not symmetric).

Until then, the honest advice is narrower than the District’s. Ranking true preferences is safe for a student who would accept at most five schools and cares only about her own placement. A student who would take more, or who cares where a friend ends up, faces exactly the strategic choice that deferred acceptance was adopted to remove.