Each dot is a California public high school. The x-axis describes the school's context — its
surroundings (American Community Survey measures for the census tract where the school sits) or its
student body (unduplicated high-need share, and the grade-11 shares socioeconomically disadvantaged
and English learner) — or its application behavior (how much of the eligible class applies to the
selected campus, and in what numbers).
The dot's color is its students' measured achievement (CAASPP grade 11, A–G completion).
The y-axis is the admissions outcome at the selected UC campus. Drag a band on the rail below the
chart to hold that variable roughly fixed, and see how much achievement still separates schools inside it.
±12
15
30
Click a dot (or a twin pair) for details.
How to read this — and important caveats
The question. The main explorer shows that a school's measured achievement relates to its UC
admissions outcomes. This page asks the follow-up: is that relationship its own thing, or is it mostly
the neighborhood? Hold context fixed conditions directly (brush a band of similar
contexts and look at the spread inside). Local correlation formalizes the band: it traces the
conditional correlation function — a kernel-weighted correlation of achievement with the outcome
computed at every point of the context distribution. Twin schools makes it concrete (near-identical
surroundings, very different achievement — do outcomes differ?). What explains more?
summarizes with a variance decomposition, campus by campus.
Neighborhood = the school's tract, not its students. Each school is assigned the American
Community Survey (ACS) 5-year profile of the census tract around its current campus location
(CDE coordinates → Census Geocoder). Students may commute across boundaries, transfer, or attend
charters far from home, and a tract (≈4,000 residents) is not an attendance zone. Read the x-axis
as "the kind of place the school sits in." Where location and enrollment visibly part ways
— the tract's SES percentile and the student body's need (UPP) percentile sitting 30+
points off alignment — the school's profile card carries a note saying so, with the
direction (about one school in four). It runs both ways: a campus in well-off surroundings
enrolling a higher-need student body (an agricultural-belt or gentrified-area school), or a
selective school drawing a lower-need student body to an ordinary tract (a citywide arts or
exam school). Percentile comparisons are the honest scale here: a tract poverty rate and a
UPP share are not comparable numbers — the statewide median high school's UPP is about
70%, so a "typical looking" UPP can sit far from a tract's rank.
Timing. Neighborhood and test measures are taken from each entering class's grade-11 year
— the spring before application — matching the main explorer's CAASPP convention. 2025
classes use ACS 2024 (the latest 5-year release), and where a class's grade-11 join is
unavailable (there are no 2020 entity files, affecting classes entering in 2021) the nearest
available year substitutes. ACS 5-year estimates are rolling averages and
move slowly; tract-level estimates carry sampling error, and top-coded medians ($250k+ income,
$2M+ home value) are kept at the cap.
The SES index is a plain z-score composite — income, adult BA share, home value, minus
poverty, minus unemployment — standardized each year over all California public high schools, so
0 is the average school's surroundings and ±1 is one standard deviation.
School-level student measures. Three context variables describe the student body rather
than the tract. UPP is the CALPADS unduplicated share of grade 9–12 students who are
FRPM-eligible, English learners, or foster youth — taken, as in the main explorer, from the
entering class's senior year. % SED and % EL are the shares of the school's grade-11
CAASPP-reported enrollment that are socioeconomically disadvantaged (FRPM-eligible or neither parent a
HS graduate) and English learners — taken from the class's grade-11 year like the test scores.
These are enrollment counts, not test outcomes, so they exist even in low-participation years;
they start in 2016 (the 2015 files do not report per-group enrollment), and classes entering 2016
use their senior-year value. A school with no reported group row counts as 0%.
Reading the local-correlation curve. Schools are ranked by the context variable
(percentile scale, so skewed dollars behave); at each grid point the correlation between achievement
and the outcome is computed with Gaussian weights (bandwidth in percentile points — the slider).
The blue ribbon is a 95% interval using the kernel's effective sample size; the gray band is a
no-variation envelope: the middle 95% of curves obtained when schools' (achievement, outcome)
pairs are randomly shuffled against their context ranks 200 times — the wiggle pure noise would
produce if the association were the same everywhere. Both are pointwise: judge the curve's overall
shape, not one bump. "Linear moderation β" is the interaction coefficient from a standardized
regression of the outcome on achievement, context percentile, and their product — the one-number
linear summary of how the association trends across the distribution. Near the edges the kernel is
one-sided; the grid stops at the 5th/95th percentiles.
Statistics are school-level and unweighted (each school counts once), computed on schools
passing the min-applicants filter. Pooled periods use ratio-of-sums for funnel rates, as in the main
explorer, with the same reliability gate for suppression-distorted rates. Partial correlation and the
R² decomposition use ordinary least squares with one neighborhood and one achievement variable at
a time — a lens, not a full causal model. "Shared" variance is explanatory overlap that the data
cannot attribute to either factor alone.
This is an observational, school-level (ecological) analysis. It cannot say what happens to an
individual student, and no partial correlation here identifies a causal effect in either direction.
Coverage. 1,468 of 1,518 schools in the main explorer have tract context (the rest are mostly
closed or relocated campuses without current coordinates, plus a few whose UC record could not be
tied to a CDE school — see the repository's crosswalk-repair note). CAASPP is missing for classes entering in
2021 (spring-2020 test cancelled) and 2022 (spring-2021 test non-representative); those periods show
context-only dots in gray.
Application behavior. Three x-variables describe who applies rather than where the
school sits or whom it enrolls: the school's application rate to the selected campus
(applicants ÷ A–G-eligible graduates), its application rate to the other eight UC
campuses (all applications sent to them ÷ eligible — this can exceed 1, since one
student can apply to several), and its raw applicant volume (applicants over the period; drawn
on a log axis). These support a standing question about admit-rate patterns: if only a school's
strongest students apply, its admit rate could reflect who chose to apply rather than how its
applications were read. Holding application behavior fixed — the band, the curve, the twin
match, the partial correlation — asks what remains. Three cautions. (1) The own-campus
application rate and any outcome built on applicants (the admit rate above all) share the applicant
count, so part of any association between them is mechanical; the other-8 rate is the
shared-term-free alternative and tracks the own-campus rate closely. (2) Per-eligible outcomes are
flat-to-positive across schools unconditionally; inside application-rate bands they can turn
negative — a composition effect of the conditioning itself. (3) Applicant counts under 3 are
suppressed in the source, so the other-8 sum can slightly undercount at very small schools.
Applicant-pool GPA outcomes. Applicant GPA and admit GPA are UC-recalculated
weighted-capped GPAs (grades 10–11, A–G coursework), averaged over the school's
applicants and admitted students respectively. Applicant GPA tests the self-selection premise
directly: if thin applicant pools were elite subsets, schools where few apply would send
stronger-GPA pools. Admit GPA is partly downstream of the decision being studied — read it as
the GPA level accepted at the school, not as school strength. On GPA axes the y-scale starts near
the data, not at zero.
Holding two things fixed. The "Also hold fixed" control adds a second context variable to
the conditioning: the side panel then also reports the partial correlation with both
variables removed (OLS residuals), and twin pairs must sit within ±4 percentile points on
both. The chart's band and the local-correlation curve continue to show the first variable only.
School composition variables. % SED and % EL are grade-11 shares from the CAASPP research
files (2016+). The racial/ethnic composition variables are CDE Census Day enrollment shares for
grades 9–12 (2014+; Asian and Filipino are separate CDE categories, combined here; "URG"
follows UC's underrepresented-group definition: Hispanic/Latino, Black or African American, American
Indian). All are aligned to each entering class's grade-11 year, like SED/EL, and describe the
school's students — not its applicants to any campus.
Data vintage
Admissions
UC Information Center, admission years through fall 2025.
Achievement
CDE CAASPP grade 11 (Smarter Balanced), spring 2015–2025. Spring 2020 was cancelled
and spring 2021 is excluded as non-representative; each entering class is matched to its own
grade-11 test the prior spring, so the 2021 and 2022 entering classes carry no CAASPP value.
UC outcomes
Entry cohorts 1999–2024. A rate appears only where its window has closed: six-year
completion through the 2019 entering cohort, five-year through 2020, four-year through 2021,
first-year retention through 2024.
Neighborhood
U.S. Census Bureau ACS 5-year estimates at census-tract level; ACS year set to the school
year, with 2025 school years carrying the 2024 release.
Last updated 31 July 2026. Build dates for each data layer are given with
the sources below; revision history is in
docs/CHANGELOG.md.