About Achievement Data

Raw test scores mostly tell you who lives in a district or attends a school. Across Colorado's school districts, four facts about a district's families and students — how many parents have a college degree, and the shares of students on free or reduced-price lunch, learning English, or in special education — account for about 90% of the differences in how many students meet state expectations. So "this district scores higher" usually means "this district serves more advantaged families."

This site asks a fairer question: given the community it serves, how does a district — or a school — do? It shows what each district's demographics predict, what it actually scored, and ranks districts by the difference. For schools, it compares each school with schools that serve similar students, next to the state's measure of how much students grow each year.

A non-partisan, non-profit community project by Yoav Lurie, and a sister site to enrollmentdata.org. It is not affiliated with or endorsed by the Colorado Department of Education or any school district. This is a working draft, built in public: methods and numbers may change as it develops.

How the district ranking works

1. Results. The share of students meeting or exceeding expectations on CMAS English language arts and math, grades 3–8, for 2024, 2025 and 2026 combined. District totals include every district-run school — neighborhood, choice, magnet and gifted programs — but not charter schools or online schools, whose results are subtracted from the state's district totals. (A few small charter or online schools have suppressed results and can't be subtracted.) The free/reduced lunch, English learner and special education shares are likewise for district-run schools.
2. Demographics. The share of parents of public-school children who have a bachelor's degree or more (a special tabulation of the US Census Bureau's American Community Survey, 2019–2023, published by NCES), and three CDE measures of the students: free or reduced-price lunch, English learners, and students with an IEP (special education), averaged over 2024–2026.
3. Expected result. A regression across all ranked districts, weighted by the number of students tested, predicts each district's English and math results from those four measures. Each district's prediction comes from a model fit to the other districts, so a district's own scores never set its own benchmark. With districts weighted by students tested, the four measures explain 89% of the differences between districts (87% when each district is predicted from the others). Parents' education matters most by far: a one-standard-deviation difference across Colorado districts is worth about 8.6 percentage points, compared with about 1–2 for free/reduced lunch, English learners or special education.
4. What else was tested. Parents' income, and the Census community measures (all adults' education, median household income, child poverty, SNAP, single-mother families, employment). Parents' income changes nothing. The community measures predict slightly better but describe all residents, including adults without children in school, so they aren't used.
5. Ranks. Predicted rank orders districts by their expected results; actual rank by their real results; the ranking orders them by the over/under (actual minus expected, in percentage points). With English + Math selected, the two subjects are averaged.
6. Groups. Districts can be compared within a size group (large: 3,300+ students tested a year in grades 3–8; mid-size: 330–3,300; small: under 330) and a type (city, suburb, town or rural: the federal NCES locale of the schools that enroll most of the district's students). Ranks within a group are re-counted among its members. Each district's most similar districts are the eight closest on the four prediction measures plus size.
7. Uncertainty. The ± figure is a 90% range based on the number of students tested and the year-to-year swings in each district's results. The likely rank range comes from re-ranking all districts 4,000 times within those ranges. Districts need at least 150 student tests in each subject over the three years, and published Census figures on parents' education; 105 are ranked. The Census doesn't publish parents' education for 41 smaller districts, which are listed but not ranked.

Report cards

Every district and school has a one-page report card: the comparison with what its demographics predict, results by year, growth, how it compares with its most similar districts or schools, and who it serves. They're plain pages that print cleanly.

How the school comparison works

1. Scores. CMAS English language arts and math mean scale scores for 2024, 2025 and 2026. Each school's average for each grade is compared with the statewide average for that grade and year (in standard deviations of the statewide student distribution), then averaged over grades 3–5 (elementary) or 6–8 (middle), weighted by the number of students tested.
2. Who's compared. Neighborhood schools only. Charter schools, online schools, alternative education campuses and choice/focus programs are left out: they don't serve a neighborhood, and some can shape who enrolls. Schools need at least 45 test scores across the three years.
3. Expected score. A weighted regression across all compared Colorado schools predicts each school's score from five characteristics: the income of families with children in the school's neighborhood (relative to the poverty line); the share of adults aged 25–44, the ages most likely to be parents of school-age children, with a bachelor's degree or more in the school's area; the share of students eligible for free or reduced-price lunch; the share learning English; and the share with an IEP (special education). Adding race/ethnicity does not change the results, so it is not used.
4. A school's area. Where we have attendance-zone maps (Boulder Valley so far), the area is the attendance zone: the Census block groups inside it. Elsewhere it is the block groups nearest the school, out to about as many adults as a typical zone (roughly a one-mile radius in town). The Census doesn't publish parents' education below the district level, so adults 25–44 are the closest available measure.
5. Vs. expected. Actual score minus expected score, shown in CMAS scale-score points. The ± range is a 90% interval reflecting year-to-year swings and the number of students tested.
6. Growth. CDE's median growth percentile (grades 4–8): how much a school's students improved compared with students statewide who had the same prior scores. 50 is typical. Growth doesn't depend much on family background, so it is the closer measure of what a school adds in a given year.
7. Look-alikes. For each school, the 25 neighborhood schools statewide whose students are most similar (neighborhood income and education count double). The look-alike percentile is where the school's raw score falls among them.

What to keep in mind

• "Explained" is not "caused." The 90% describes differences between district averages, not individual students, and the demographics stand in for many things that travel with them: preschool, tutoring, home resources, and families choosing districts for their schools.
• Over/under isn't pure school quality. It includes everything the demographics don't capture, as well as the schools themselves. Read it alongside growth.
• Parents' education is an estimate. It comes from the Census Bureau's annual survey, so it carries sampling error, especially in smaller districts. It covers parents of all public-school children living in the district, including those who attend school elsewhere.
• Small districts swing more. With a few hundred students tested, a handful of students can move results several points, and Census figures for small communities are less precise. Weighting every district equally, the model explains a little over half of the differences. Use the size filter and the rank ranges.
• Families who leave aren't counted. Students who live in a district but attend a charter school, another district, an online school or a private school are not in its results. Parents' education, though, covers all public-school families in the district, charter families included.
• A school's area describes where it draws from, not the enrolled families. It fits best where most students live nearby; open enrollment and choice weaken it.
• Free/reduced lunch counts have become less reliable since Colorado began offering free meals to all students, and CDE flags its 2026 counts as less precise. That's why neighborhood income is also in the model.
• Choice, magnet and gifted programs are left out of school comparisons (they don't serve a neighborhood) but stay in district results. The list covers Boulder Valley and a first set of well-known programs elsewhere; corrections are welcome.
• Test participation. Schools where fewer than 90% of students tested are flagged; families who opt out may differ from those who test.
• Small differences are noise. A few points either way, or a handful of ranks, is within normal year-to-year variation.

Data sources

Open source

The code and derived data are dedicated to the public domain under CC0 1.0. Every figure can be traced to its source file; the processing scripts and a list of every file used are in the project repository.

Found an error, or know of a choice/focus program that should be excluded? Please get in touch.

← Back to the data