Find What Works: Schools

← All Philly stories

PhiladelphiaMethodologyeasy

With 287 peer-grouped schools and k=40, the rankings barely move at k=20

Thesis

Only 8 Philly schools (~3%) show a substantively-different top-10 peer set between k=20 and k=40. The KNN matching is stable, so the choice of k matters less than which features feed the distance — and in v1 we use a 7-feature demographic Euclidean (race composition, ELL, IEP, econ-disadv).

Supporting findings

  • 287 of 330 open Philly schools get a peer group; 43 are dropped (matching universe <5 peers).
  • k=20, k=30, k=40 all derived; default = k=40 to match NYC's convention.
  • Flip-diff: 8 schools have ≥3 different top-10 peers between k=20 and k=40. That's the substantive-flip count.
  • Smaller matching universes are the binding constraint, not k.

Reporting directions

  • Show which schools are in the 43-school 'peer-group-less' set — they're a hidden cohort the Outliers UI can't rank.
  • Test sensitivity to one excluded covariate: drop pct_econ_disadv and re-derive; how many top-10 peer sets change?
  • Compare to NYC's peer-group volatility story — same methodological hazard, smaller sample size.

Methodology & replication recipe — the queries + sources behind the findings above.

More on methodology