Find What Works: Schools
← All source documentsRaw file
Source document

docs/cities/miami/05_validation_spec.md

Served verbatim from the project repository. Internal working document conventions apply: documents may reference file paths, branch names, and findings-ledger anchors from the repo.

05 — Validation Spec (Miami-Dade)

Adaptation of verify/METHODOLOGY.md (NYC) and docs/cities/philly/05_validation_spec.md to Florida sources. Same four checks, same posture, different publishers.

The four checks

  • (i) Completeness — every school in the MSID active list (and every school in the year's School Grades roster) appears in each served (metric, year); source-only and served-only cells reported.
  • (ii) Base case — the served value equals an independent re-read of the exact ingested file (xlsx/xls by pandas; PDF tables by a second parser). The binding gate.
  • (iii-a) Spot check, same publisher — FLDOE School Grades components; Florida Report Cards on a stratified sample; M-DCPS ABC book and school dashboards for the closure track.
  • (iii-b) Independent publisher — NCES CCD, EDFacts (EdDataExpress), CRDC, county GIS.
  • (iv) Computed values — derived metrics recomputed in Python from validated inputs (packet recomputation, as in the Philly full QA).

Per-metric expectations

Metric(ii) source(iii-a)(iii-b)Tolerance (ii / iii)
fast_ela_g{3..10}_proficiency, fast_math_g{3..8}_proficiencySRS xls, per gradeSchool Grades ELA/Math Achievement (all grades pooled, ±2 pp; only a sanity check)— (no independent publisher for FL assessments)±1.0 / ±2.0
eoc_{alg1,geometry,biology,civics,ushistory}_proficiencySRS xls (spring administration)School Grades Science / Social Studies Achievement where applicable—±1.0 / ±2.0
learning_gains_{ela,math}, _l25Gains xlsSchool Grades gains components (exact expected)—±0.5 / ±0.5
school_grade, componentsSchool Grades xlsxReport Cards—exact
graduation_rate_4yr (+ subgroups)FedGradRate xlsSchool Grades Graduation Rate (lagged one year — align cohorts)—±0.5 / ±1.0
chronic_absenteeism_rate (10%)ABS21Days10Comparison xlsxReport Cards ESSA absenteeismEDFacts FS195 (ESSA 10%; perturbed)±0.5 / ±1.5
absent_21plus_rateABS21Days10Comparison / ABS21DAYSTDistSchl—CRDC 15+ days: correlation only (different cut)±0.5 / n.a.
attendance_rate (ADA/ADM)ADAADMSchl xlsxdashboards ADA—±0.2
sesir_*_per_100SESIR xlsx—CRDC harassment/bullying counts (correlation)exact counts / n.a.
survey_* rollupsclimate workbooksPower BI dashboard spot sample—±0.0 (exact from raw)
enrollment_octMembBySchoolByGrade xlsxStatistical Highlights district totals; dashboardsCCD membership (federal copy of the same submission; mismatch is a finding)exact / ±0.5% district
fish_capacity_permanent, fish_capacity_school, cofte, utilization_school_capacity_pctFISH LOS PDF (two parsers)ABC book Perm FISH Cap for targeted schools (exact)county GIS CAPACITY (permanent; vintage differences expected and reported)exact
per_pupil_expenditure_*ESSA PPE xlsxReport Cards PPE—exact
teacher_fte, students_per_teacherCCD staff filedashboards staff counts (headcount vs FTE — report both)—exact / n.a.

Known reconciliation hazards (stated before the first load)

  1. School-number padding (0026 vs 26) — normalize before any join; a padding bug shows up as (i) failures, so the completeness check runs first.
  2. Virtual provider numbers — accountability files carry an extra key column; the virtual programs must not collide with the district's own rows.
  3. Denominators differ by survey: October membership (Survey 2) vs annual unduplicated enrollment (Survey 5, ~6% larger; students at two schools counted at each) vs February FTE (Survey 3) vs COFTE (funding FTE). Every served rate names its denominator.
  4. Three utilization numerators (COFTE in FISH LOS; headcount in the ABC book; a 2024-25 count in GIS) and two capacities (permanent; school capacity incl. relocatables). Stored separately; never mixed.
  5. Graduation-year indexing: 2425 = cohort entering grade 9 in 2021-22; School Grades reports it lagged one year.
  6. Suppression tokens vary by file and year (*, **, blank, Less_10); the token census is part of every load.
  7. Assessment regime breaks (2015 FSA, 2023 FAST; no 2020) and the 2023 "informational baseline" — (ii) is per-year and unaffected; (iii-a) against School Grades must use the same year's file.
  8. School Grades achievement pools grades; SRS files are per grade. The (iii-a) proficiency check is a sanity bound, not an oracle.
  9. CEP and the 1.6 multiplier in economic-disadvantage shares; comparisons across CEP status flagged.
  10. CCD lag (~1.5 years): (iii-b) on enrollment covers years ≤ 2023-24 only.

What lands first

  1. (ii) base case for every C1 and P1 metric, full coverage, before anything is published — NYC's bar.
  2. (iii-b) CCD-vs-FLDOE enrollment on every school × year in the overlap (this is the strongest check available in any city so far, since both copies derive from the same submission).
  3. (iii-a) School Grades components and ABC-book capacity on the affected schools.
  4. (iv) packet recomputation for the closure page (mirrors pipeline_philly/verify/absenteeism_full_qa.py).

Authority

Authoritative for Miami-Dade verification expectations; divergences from the NYC spec are intentional and listed above.