Does "Advanced" Predict College? We Compared 1,726 Schools

This is SchoolScope's analysis, not a CDE published finding. CDE supplies the underlying school test results and college-destination fields. We chose the inclusion rules, pairs and statistical checks.

Who and when. We linked schools, not individual students. The class-of-2023 destination file is paired with 2022 grade-11 tests; class-of-2020 destinations are paired with 2019 tests. The analysis uses ACS 2020–24 neighborhood estimates and school covariates stored in the snapshot. The stored meal-eligibility measure comes from NCES Common Core of Data enrichment; recorded race/ethnicity shares come from CDE enrollment data. Those are contextual measures, not historical family records for the students in either completion year. We checked the score and destination fields against retained CDE files. We have not independently checked the raw neighborhood and current school-demographic sources.

CDE's College-Going Rate concerns high school completers, a broader category than diploma graduates. Enrollment means enrolling at a college within 12 months of completion. The analysis is not a comparison of a particular autumn's incoming class.

CDE’s report definitions assign someone with multiple college enrollments to the institution where they had the longest enrollment during the reporting window. These are recorded destinations, not counts of everyone who attended each type at any time. Privacy-blocked National Student Clearinghouse records can lower recorded college-going rates; we have not estimated a school-specific correction for that missing enrollment. Here, community college means the California Community College system, not out-of-state or private two-year institutions.

The test measure averages English and math subject percentages without weighting them by subject count. Its minimum of 50 scored student-subject records sums the English and math Students with Scores counts; the same student can appear in both. It is not a minimum of 50 distinct juniors. Schools also need usable destination results and covariates. We exclude ZIP income reported at the top-coded value, which means "$250,000 or more," rather than an exact income.

What the intervals mean. A nominal 95% bootstrap interval summarizes the uncertainty from resampling these matched pairs under our procedure. These intervals condition on the selected pairs; we do not repeat the school-matching step inside each resample. They do not account for every source of bias or establish a causal effect. We have not established that this procedure gives 95% coverage for these selected schools or modeled dependence from shared district and neighborhood context. An interval crossing zero leaves a zero average gap compatible with this procedure.

Main comparisons, both completion years

All gaps and intervals are percentage points, higher-Advanced school minus lower-Advanced school. UC is also part of the mixed proxy; the destination categories are not four separate groups of students.

We use a paired sign-flip test, which compares the observed average gap with gaps obtained by randomly reversing its direction within pairs. The sign-flip test assumes that reversing directions within pairs is appropriate under its null; matching schools does not establish that assumption. We also adjust for testing four outcomes within each design using the Benjamini–Hochberg procedure. These adjusted q-values do not establish false-discovery control for this study’s outcome dependence. A p-value is not the probability that a finding is due to luck or that no difference exists.

At an adjusted q threshold of 0.05, the main UC q-value is below the threshold in 2023 but not in 2020. The adjustment is within each four-outcome design; it does not cover the full set of exploratory matching choices. Neither the threshold nor the positive UC intervals make these school comparisons causal.

Which schools made it into pairs?

Only 252 of 1,726 eligible schools matched in 2023, and 254 of 1,691 in 2020: about 15% in each year. They are a selected sample, with higher test results and college-going rates than the schools left unmatched.

These figures describe the schools selected by this design, not California high schools generally. Some schools can appear in both completion years; this is not two independent samples of schools.

What changes when we change the pairs?

The table gives gaps followed by nominal 95% bootstrap intervals. Neighborhood education matching limits the difference in the share of adults with a bachelor's degree. Racial-composition matching limits differences in four recorded student race/ethnicity shares. Both are school or area measures; neither tells us an individual child's background.

The 2023 neighborhood-matched UC estimates approach zero. In 2020 the strictest neighborhood match also approaches zero, but excluding charters or choosing larger completion groups produces bigger positive estimates. Different choices select different schools and change the uncertainty. The five same-district pairs in 2023 are especially thin evidence.

A separate exploratory regression within the original pairs tells a less decisive story. Multiplying its neighborhood-education coefficient by the average remaining imbalance gives about 0.50 points in 2023 and 0.34 in 2020. That calculation cannot explain the whole UC gap or separate school effects from family selection. It also cannot resolve why re-matching changes the result.

Why we cannot claim a four-point limit for California community colleges

The California community-college estimate changes from −1.60 points in 2023 to +0.80 in 2020. Both primary nominal bootstrap intervals cross zero. That is uncertainty, not proof of equal outcomes.

For this California system, the script also computes an approximate equivalence envelope: the absolute estimated gap plus 1.645 standard errors. In the primary designs it is ±4.24 points for 2023 and ±3.13 for 2020. The 2023 alternative designs produce envelopes from ±4.28 to ±6.53, apart from the five-pair same-district case at ±11.63. In 2020 the alternatives range from ±3.62 to ±4.79, apart from the same-district case at ±7.56.

These are normal-approximation calculations within selected pairs. They do not guarantee that a larger difference would have been detected, and they do not bound a causal school effect. Our earlier claim that every check ruled out a shift larger than about four students per hundred was too strong.

What changed on this page?

Correction, September 28, 2026: We restored the exclusion of ZIP income reported as "$250,000 or more." The 2023 pool changed from 1,732 schools and 127 pairs to 1,726 schools and 126 pairs; the 2020 pool changed from 1,697 and 128 to 1,691 and 127. We updated the primary results, alternative matches and sample descriptions for both years.

An outside review also found that our original random-number generator repeated resamples. We replaced it and recomputed every confidence interval, p-value and adjusted q-value for both years; the paired schools and average gaps did not change. These engineering checks do not certify the statistical assumptions. We also corrected the multiple-testing calculation, the mixed destination label, the distinction between scored subject records and distinct students, and the California-only community-college category and the strength of that conclusion. The chart now covers the four destination measures from this analysis. Graduation, AP and A–G results came from separate exploratory work and are no longer presented as additional outcomes from the same matched destination study.

The earlier published page reported an August 15 correction to its stored record-count input. The figures here come from the September 28 run. The analysis uses a fixed random seed; its school census, exact queries, pair lists and output logs are retained with the correction record. Contact us if you want to discuss the method.