Chronic absenteeism in Philadelphia: what we measured, what we found, and why you can trust it

The School District of Philadelphia publishes no school-level chronic-absenteeism rate in its open-data files. What it does publish, in its School Performance open-data collection, is the share of students with 90% or better attendance, per school per year. Our headline series is the complement:

share below 90% = 100 − (“% with 90%+ Attendance (Yearly)”)

— the share of a school’s students who attended less than 90% of their enrolled days. This is a proxy for the chronic-absenteeism rate, and we label it one wherever it appears. The table shows how it lines up against the definitions the publishers actually use.

• Boundary. Under the federal definition (U.S. Department of Education), a student who misses exactly 10% of enrolled days is chronically absent. That same student attended exactly 90% and so is inside the publishers’ “90% or better” group, which puts them outside our below-90 share. For the same students, enrollment spells, and day-counting rules, the proxy therefore cannot exceed a federal-definition rate, and it understates it by the share of students sitting exactly on the line. Neither publisher reports that share, and neither documents its rounding at the line, so the size of that understatement is unknown. The ordering does not carry over to a separately published federal statistic built on different eligibility or attribution rules; the proxy is a related measure, not a close estimate of any particular federal figure.
• Minimum enrollment. The district’s Office of Research and Evaluation documents the rule for this file family: a student is included at a school if enrolled there for at least 10 days, and a student enrolled 10 or more days at more than one school is attributed to each of them (Student Attendance Patterns in Philadelphia, 2017-18 to 2021-22, June 2023, “Inclusion and Attribution Criteria”). Two consequences: the yearly student counts in the file are student records, not distinct students, and a student who moved between schools mid-year can appear in both schools’ figures. The state’s rule is stricter: its Future Ready glossary defines Regular Attendance over students “enrolled in a school for 90 or more school days.” If shorter-enrollment records have lower attendance, excluding them could raise the state’s 90%+ share; we have not measured that relationship or its contribution to the publisher difference. §4 reports the observed difference.
• Local coding. The state standardizes the threshold and the minimum enrollment but not how a partial day counts. Its Guidelines for Reporting Regular Attendance and Chronic Absenteeism (May 2019) state that “missed hours, periods, or fractions of days may be handled differently by each LEA,” and give the example of a student who is a quarter day late four times counting as present 180 days under one local policy and 179 under another. District-run schools and the charter-school local education agencies may use different local attendance policies (district schools share the district’s). We have not verified that they code attendance identically, or that any of them coded it the same way before and after the pandemic; the sector comparison in Figures 9–10 carries that caveat.
• Excused vs unexcused. The 90% threshold here counts all absences; the district’s report states that all classifications and reasons for absence are included.
• The state’s own chronic-absenteeism column. The state’s workbooks carry a PercentChronicAbsenteeism column in its 2024-25 workbook, blank for every Philadelphia school row; the 2021-22, 2022-23, and 2023-24 workbooks have no such column at all (counts in §5). So our served chronic_absenteeism_rate metric is empty and hidden from the ratings dropdowns. The state reports the concept positively, as Regular Attendance; the column is named PercentPersistentAttendance in the 2021-22 and later workbooks and “Percent Regular Attendance” in the earlier ones. We refer to it by the glossary’s name, Regular Attendance, throughout.

Every source file is fetched by pipeline_philly/discover.py, which records sha256, byte-count, and retrieval-time provenance beside the file. The registry of record (URLs, column mappings, year encodings, and per-source caveats) is docs/cities/philly/sources.yaml.

Different kinds of error need different checks, so the status is given per kind rather than as one badge.

The checks, in detail

• Packet-local crosschecks. The generator itself recomputes the district-wide series from the source file using the exact published numerators and compares it with the figure built from the loaded percentages, on the common schools. That is 15 aggregate comparisons (weighted rate, school count, and student count for each of five years), not a cell-by-cell test; the cell-by-cell test is the full QA above. Tolerance for the rates is 0.01 percentage points: the loaded values are published to two decimals, so each school’s rate can differ from its exact-count rate by up to 0.005, and a weighted average of such errors is bounded by 0.005 when the school sets match; 0.01 is that bound plus a small explicit margin. Largest difference observed: 0.0003 percentage points. Counts must match exactly.
• Cross-publisher check, aligned by attendance year. For 216 district-run schools with 2023-24 attendance in both the district’s file and the state’s 2024-25 workbook, the state’s 90%+ share runs a median 1.7 percentage points above the district’s; 27.8% of schools agree within 1 point and 97.2% within 5; 0 differ by more than 10; the correlation is 0.9948. Individual schools fall on both sides (the difference ranges from −3.9 to +6.5), so the median is a central tendency, not a one-directional gap. Across the overlapping years the median difference is 2020-21: +0.8, 2021-22: +1.7, 2022-23: +1.7, 2023-24: +1.7, the same sign each time with the correlation above 0.99 in every year. The previously reported large break disappears when attendance years are aligned; this comparison does not establish that all local reporting practices stayed unchanged. The state’s longer minimum enrollment is a plausible contributor to the difference, but we have not recomputed the district’s measure for the state’s eligible population (the district file has no enrollment-spell detail), so we cannot say how much of the offset it explains; local partial-day coding, transfer attribution, revisions, and rounding are other candidates. The check covers district-run schools in attendance years 2020-21 through 2023-24 only. The threshold for treating the two as one number (median within 0.5 and 80% of schools within 1) is not met, so the site never mixes the two publishers in one series. That threshold was written into the script before its first run on 2026-09-06; readers cannot verify that from the outputs, so treat it as a claim. Full report: cross-publisher attendance check.
• What the misaligned comparison looked like. Before 2026-09-07 the same check matched the state’s workbook year to the district’s attendance year, and the site reported that the state’s figure ran about 11 percentage points above the district’s in 2020-21 and 2021-22 and that the state had changed its attendance rules at 2022-23. Both statements were artifacts of comparing different school years, and both are withdrawn. The misaligned comparison is preserved in the report as a diagnostic so the error is reproducible.
• Closed schools. The database joins on the current directory and so lacks 2 school-years present in the district’s file. The analysis computes its district-wide series from the file with those included, and reports the direction of the effect from the data rather than assuming closed schools were high-absence:

2026-09-07: external methodology audit. A detailed outside review of these pages found several errors, all now corrected:

• The state’s attendance indicator was stored under its workbook year rather than the attendance year it describes. The site had reported an 11-point disagreement between the state’s and district’s figures in 2020-21 and 2021-22 and a change in the state’s rules at 2022-23; both were artifacts of that misalignment and are withdrawn. The state series now runs by attendance year, 2017-18 to 2023-24, with pandemic years hatched, and the charter comparison’s latest year is 2023-24.
• The site said the state never published a 2019-20 workbook. It did; the workbook is now registered and loaded (attendance year 2018-19).
• The boundary case in §1 was stated backwards (a student at exactly 90% attendance is chronically absent under the federal definition, not the reverse).
• The district’s 10-day enrollment rule and the state’s 90-day rule were described as undocumented; both are documented and now quoted.
• The cross-publisher summary said eight schools differed by more than 10 points; the report’s own list showed twelve. The count is now computed, not typed.
• The headline “students” figure is a count of student records, and is now labeled that way; closed schools were assumed to raise the district-wide rate, and the effect is now reported from the data (it goes both ways); the Black–White gap at common schools was described as a within-school gap, and within-school versions are now computed separately; the gender series gained a fixed-panel check; 2020-21 is described as mostly virtual rather than all-virtual; the claim that the two attendance measures show “the same improvement” at a fixed ratio, and the claim that year-to-year mover rankings are mostly luck, were withdrawn.

2026-09-08: second external audit. A follow-up review found that several first-round fixes had not reached all the live text and graphics, plus new errors. Fixed: the charter figures were described as shares of students when they are averages of school rates (now led by medians, with the weighted average named as such and a fixed-panel check); the Black–White takeaway described the equal-weight within-school gap (8.4 in 2021-22, 8.6 in 2024-25) as having declined, and left 8.4 standing as the apparent baseline for the record-weighted 6.2 (now one estimator per pair, plus a 97-school fixed panel); differences between rates were still called “percent” in several places (rewritten as level sequences or as moves on the 0-to-100 scale); Figure 4 still carried the withdrawn “2.9×” annotation (now two level panels); the monthly takeaway claimed a six-year November record that the mostly virtual 2020-21 year beats, and a “worst winter” not supported by a December–February average (superlatives are now composed from a records block the QA recomputes); grade and subgroup wording claimed monotone patterns with exceptions (kindergarten to grade 1, grades 10 to 11, the Asian series); the charter count of 31 schools serving high-school grades was described as all mixed-grade (9 are stand-alone high schools; 22 are mixed-grade); the state’s 90-day rule was presented as the established cause of the publisher offset (now a hypothesis); the chronic-absence column census counted absent columns as blank cells (2021-22 through 2023-24 workbooks have no such column); the district’s published refresh schedule was overlooked; the dashboard-deletion and “most cities” claims were withdrawn; state cell proofs printed template variables; the Black–White proof now shows the subgroup inputs; the QA badge no longer claims to cover every number. Added: the local partial-day coding caveat, the historical governance limitation, a workbook-to-attendance-year map, the graduation caveat, and the New York comparison’s qualification.

2026-09-08: third external audit. The charter and racial calculations reproduced (182 values checked against the reviewer’s own extractions). Fixed: the QA documentation named derived fields as checked that the program never compared (grade shape, subgroup peak years, parts of the monthly records, fixed-panel level pairs, size-band medians — all now compared directly); universal assurance language on this page, in §5 of the analysis, and in the QA report’s title; Figure 10’s claim that its averaged bars are necessarily understated (withdrawn — the direction cannot be inferred across schools) and its income sentence (now a labeled diagnostic of school averages); the district proof pages’ stale denominator description and old publisher-gap label; the packet and sector proof still asserting the 90-day rule as the cause of the publisher offset and mid-year leavers as the worst attenders; the token census, which had skipped the three long-format workbooks (the 2020-21 workbook spells suppression “Insufficient Sample”); the analysis introduction’s “close stand-in”; the small-school reliability inference; the New York heading; the district report’s population; a 0.02 tolerance still quoted in two places; and the “212” school minimum. Added: content hashes on every evidence link so a downloaded packet or QA report always matches the page that cites it, with a shorter cache on raw downloads.

2026-09-08: fourth external audit. The reviewer’s 777 comparisons all matched. Fixed: the QA comparator skipped the student-count comparison whenever the attendance rate was suppressed (2,932 cells) and did not treat a count missing on one side as a mismatch — counts are now compared independently of suppression, with acceptance tests; the evidence identifiers were cache-busting only — the site now refuses a link whose identifier no longer matches, the source viewer preserves the requested version, and the hashing rule is stated; residual “every figure” assurance wording; two methodology sentences that stated a premise (short enrollments have low attendance) and a conclusion (no change in measurement) beyond the evidence; the registry’s dashboard-deletion note and “only published 2025-26 attendance” claim; the packet’s 212–216 school range (now derived from the data); a population convention at the top of the analysis; the charter opening scoped to schools matchable to the current directory; and the over-broad “excluded from every trend statement.”

2026-09-08: fifth external audit. Final wording pass; the reviewer found no numerical change needed and re-ran 777 comparisons with no mismatches. Fixed: the monthly takeaway’s parenthesis and the limitations section now scope the 2020-21 exclusion to the annual comparisons (Figure 8 compares matching months and windows across all years, labeled as such), and the packet’s pandemic-year note says the same; the population convention distinguishes annual school-file totals from Figure 8’s monthly eligibility; chart accessibility descriptions name their denominators; the analysis opening is a short status line with identifiers moved to a technical-details section; the source viewer’s stale-link notice now says accurately that its download is the current version; the comparator’s acceptance tests are linked from §4.

To report a problem with any number on these pages, use the GitHub link on the About page; corrections are logged here with dates.