NYC chronic absenteeism — first descriptives (Phase 5, loop 1)

Status: internal working memo — computed on the dev replica (loader-faithful to prod, base-case-verified); pre-Chalkbeat-checkpoint. Findings indexed in docs/design/08_absenteeism_findings.md.

2026-06-09 · computed from the local replica (built by the repo's own loaders from the official InfoHub file; cell-level fidelity verified). Universe: NYC district schools (include_in_default_comparisons, no charters/D75/D79). Conventions per the metric memo: 2019-20/2020-21 excluded from trends; school-level stats unweighted unless labeled; "weighted" = student-weighted by denominator.

Headlines

1. Recovery is real but decelerating and incomplete. Student-weighted chronic absenteeism: 25.8% (2018-19) → 39.5% peak (2021-22) → 32.3% (2024-25). Annual improvement has slowed: −4.4, −1.3, −1.5pp. At this pace the pre-pandemic level is ~4+ years away.
2. High schools have nearly recovered; elementary/middle have not. HS is +3.3pp vs its 2018-19 baseline (and recovered most from the peak, −8.9pp); ES +7.1, MS +8.5, K8 +9.5. Grade 12 is the only grade fully back to baseline (−0.1pp); grades 1–8 remain +7–9pp. The persistent damage is concentrated in younger grades — a strong story lead and a reversal of the usual "HS is the problem" framing.
3. The grade U-shape is confirmed (new per-grade data): PK 48.2% and grade 12 40.1% highest; grades 5–6 lowest (~26.6%) in 2024-25 — same shape pre-pandemic.
4. Metric choice validated: CA and ADA rank schools almost identically (r ≈ −0.91 every year) but CA amplifies the same attendance loss ~4× (ADA fell 3.2pp at the peak; CA rose 13.7pp). CA is the right headline; ADA the right companion.
5. Most variation is within districts, not between them: 23.3% of school-level variance is between-district (still: D20 21.8% vs D23 45.2% weighted — the worst district is >2× the best).
6. School CA is signal, not noise: consecutive-year correlation 0.89–0.94. 58 schools sat in the worst decile in all four post-COVID years, 83 in the best — stable outliers exist and are worth naming. But small schools are 1.5–2× as volatile (a 5pp one-year swing at a 150-student school is ~noise), and the worst 2022-23 decile improved 3.4pp more than average by 2024-25 — classic regression to the mean. Any outlier/mover claim needs a multi-year + size-aware gate.
7. Demographics predict roughly half of school-level CA (R² 0.43–0.50/yr; 0.46 in 2024-25) — poverty is the strongest correlate (r ≈ 0.57). Prediction is much better for ES (0.57) than HS (0.42): the most school-attributable (or unmeasured) variation lives at the HS level.
8. Demographic-adjusted residuals are stable (r = 0.88 year-over-year) — a "beating the odds" list will not be noise. Early preview: Eagle Academy campuses (3 of the top 10 negative residuals) and Concourse Village ES beat predictions by 27–39pp; the worse-than-predicted tail is dominated by small Manhattan D2 high schools — likely school-type confounding to resolve before publishing (admission type/transfer-adjacent controls).
9. Equity is the sharpest finding: the poverty gap is at its 7-year maximum. Weighted gaps 2024-25: students in temporary housing 49.0% vs 29.5% (19.5pp); poverty 36.5% vs 17.8% (18.7pp, vs 14.7pp pre-pandemic). Recovery has been faster for advantaged students; the ELL gap flipped from ~0 pre-pandemic to 5.5pp. Roughly a quarter to a third of each citywide gap is between-school sorting rather than within-school difference.
10. Pseudo-cohort tracking works (grade 3 → grade 4 within school: r = 0.87; only 5.2% of cohorts churn >20%), opening cohort-style analyses with the new per-grade data.

⚠ Post-memo finding (NYSED cross-publisher reconciliation, same day): five of six publisher-comparison panels agree (median per-school diff ≤0.4pp, r ≥ 0.99), but 2024-25 grades 1–8 diverges systematically: DOE shows EM chronic absenteeism falling in 2024-25 while NYSED shows it rising on the same schools with near-identical denominators (weighted 29.0% vs 30.8%). Cause not established. Document review (nysed-recon/NARROWING.md) rules out any documented 2024-25 definition change on either side (NYSED's enrollment-window change starts 2025-26; ReadMes/glossaries word-identical; DOE internally consistent incl. the MMR), and the one standing definitional difference — NYSED excluding suspensions/extended-medical absences — predicts the opposite sign. One documented candidate mechanism remains (SIRS defaults missing/unpaired attendance records to "absent, unexcused"; whether NYC's 2024-25 EM submission was affected is unestablished — inquiry drafted to NYSED). Until resolved, every "continued improvement into 2024-25" claim carries a publisher asterisk for elementary/middle grades — HS agreement is excellent (89% within ±1pp), so HS-recovery findings are unaffected. See docs/qa_reports/nysed/RECONCILIATION.md.

Implications for the next phases

• The bake-off should treat stability as a first-class metric (we now have baseline numbers to beat) and must include school-type/admission controls — the D2 positive-residual cluster is exactly the confound the K-NN hard filters and the residual model's covariates need to handle.
• HS needs the most careful adjustment (lowest R²) — also where NYC's own Comparison Group method leans on richer student-level controls we don't have; expect wider uncertainty bands there.
• The pct_econ_dis top-coding at 95% (stored as NULL for 8–18% of schools) is a data wart that biases any SES adjustment — worth fixing in the demographics loader (ingest the "Above 95%" flag) and worth knowing about for the ENI question.
• Early story leads (all need validation passes before sharing beyond the team): poverty gap at 7-year max; younger grades stuck while HS recovered; the 58/83 stable outlier schools; 24 schools where poverty students attend better than their non-poverty schoolmates.

Trends & variation

NYC district-school chronic absenteeism (CA) peaked at a student-weighted 39.5% in 2021-22, has fallen for three straight years to 32.3% in 2024-25, but remains 6.5 points above the 2018-19 baseline of 25.8%. The recovery is decelerating (-4.4, -1.3, -1.5 pts/yr). HS is closest to its pre-pandemic level (+3.3 pts) while K8/MS/ES remain 7-10 pts elevated. CA and ADA are near-mirror images across schools (r ~ -0.91 every year), but CA is far more sensitive: ADA fell only 3.2 pts at the peak while CA jumped 13.7. The by-grade U-shape is confirmed in both 2018-19 and 2024-25 (PK and grade 12 highest; grades 5-6 lowest); grade 12 and PK recovered most from the 2021-22 peak, and grade 12 is the only grade fully back to baseline. In 2024-25 district weighted CA spans 21.8% (D20) to 45.2% (D23); Bronx (37.5%) is the worst borough, Staten Island (28.7%) the best. Only 23.3% of school-level CA variance is between districts; 76.7% is within districts.

Findings

• Citywide student-weighted CA (subgroup ALL, school-wide key, default universe of ~1,450-1,470 district schools): 25.8% in 2018-19, peak 39.5% in 2021-22, then 35.1% (2022-23), 33.8% (2023-24), 32.3% (2024-25). Net change vs pre-pandemic: +6.5 pts; recovery from peak: -7.2 pts, with the annual improvement decelerating from -4.4 to -1.3 to -1.5 pts.
• The whole school distribution shifted up, not just the mean: median school CA 35.7% in 2024-25 vs 27.9% in 2018-19; p10 16.8 vs 10.5; p90 54.1 vs 49.2. Cross-school SD is essentially unchanged (14.1 in 2024-25 vs 14.7 in 2018-19), after spiking to 17.2 in remote-attendance 2020-21.
• Grade-band trends (weighted): HS is highest every year (34.1% in 2024-25) and MS lowest (29.2%), but HS recovered the most from the 2021-22 peak (43.0 -> 34.1, -8.9 pts) and sits closest to its 2018-19 baseline (+3.3 pts). Elevation vs 2018-19 elsewhere: ES +7.1 (24.6 -> 31.7), MS +8.5 (20.7 -> 29.2), K8 +9.5 (23.7 -> 33.2).
• CA vs ADA contrast: school-level Pearson r between the two metrics ranges -0.907 to -0.948 across the seven years (R^2 0.82-0.90; n 1,450-1,470 schools/yr), so they rank schools almost identically. But citywide weighted ADA moved only 92.0% -> 88.8% at the 2021-22 trough (-3.2 pts) while CA moved 25.8% -> 39.5% (+13.7 pts) - CA amplifies the same attendance loss roughly 4x because it is a threshold (share of students missing >=10% of days), making it the more sensitive headline metric.
• U-shape by grade confirmed in 2024-25 (weighted CA): PK 48.2%, K 37.8%, falling to a minimum of 26.6-26.7% at grades 5-6, then rising through 28.7% (gr 7), 31.2% (gr 8), 33.6-33.7% (gr 9-10), 31.5% (gr 11), to 40.1% at grade 12. The identical shape existed in 2018-19 (PK 43.5%, min 18.1% at gr 6, gr 12 40.2%).
• Recovery vs the 2021-22 peak is largest at the ends of the U: grade 12 -13.5 pts (53.6 -> 40.1), PK -12.2 (60.3 -> 48.2), grade 11 -10.3, grade 10 -8.1, K -8.0; smallest in grades 7-8 (-4.6 and -4.5). Grade 12 is the only grade at/below its 2018-19 level (-0.1 pt); elementary and middle grades (1-8) remain 7.1-8.9 pts above pre-pandemic, the most persistent residual damage.
• District variation 2024-25 (weighted CA): best five - D20 21.8%, D26 22.0%, D25 22.5%, D15 24.3%, D24 27.4%; worst five - D23 45.2%, D05 44.6%, D16 44.4%, D08 43.2%, D19 42.5%. Best-to-worst spread is 23.4 pts, i.e., the worst district's rate is more than double the best's.
• Borough weighted CA 2024-25: Staten Island 28.7%, Queens 29.8%, Brooklyn 31.9%, Manhattan 33.5%, Bronx 37.5% (unweighted school means: 30.7, 31.0, 36.4, 38.1, 38.6 respectively; borough codes verified B=Bronx, K=Brooklyn, M=Manhattan, Q=Queens, R=Staten Island via district mapping).
• Variance decomposition (one-way ANOVA on 1,450 school-level CA values across 32 districts, 2024-25, unweighted): SS_between = 66,833 of SS_total = 287,194, so 23.3% of variance is between districts and 76.7% within districts (eta-squared = 0.233; F(31, 1418) = 13.9). District membership matters and is highly significant, but most of the school-to-school variation lives inside districts.

Tables

Table 1. Citywide chronic absenteeism by year (subgroup ALL, school-wide key, default district-school universe)

*Excluded from trend fits: 2019-20 truncated by COVID closure; 2020-21 counted remote attendance.

Table 2. Student-weighted CA % by grade band and year (unweighted school mean in parentheses)

*COVID-affected years, excluded from trend interpretation. N schools/yr: ES ~638-644, MS ~238-244, K8 164, HS ~406-414. (K12 band, 46 schools, omitted.)

Table 3. CA vs ADA: citywide weighted means and school-level correlation by year

*COVID-affected years. Note the asymmetry: at the 2021-22 trough ADA fell 3.2 pts vs 2018-19 while CA rose 13.7 pts.

Table 4. Student-weighted CA % by grade: 2018-19 baseline, 2021-22 peak, 2024-25 (default universe)

U-shape: maxima at PK and grade 12, minimum at grades 5-6 in all three years. Largest recoveries from peak at grade 12, PK, and grade 11; grade 12 is the only grade back to its pre-pandemic level.

Table 5. District and borough variation, 2024-25 (student-weighted CA)

Best five districts:

Worst five districts:

Boroughs:

Table 6. Variance decomposition of school-level CA across districts, 2024-25 (one-way ANOVA, unweighted school values)

n = 1,450 schools, k = 32 districts; eta-squared = 0.233; F(31, 1418) = 13.9.

Caveats

• 2019-20 (COVID-truncated year) and 2020-21 (remote attendance counted) are shown for completeness but excluded from trend interpretation per the project's metric memo; do not fit trends through them.
• Universe is NYC district schools with include_in_default_comparisons = true (2,006 of 2,131 schools), which excludes D75/D79/alternative programs; charters have no attendance rows at all. Table 2 omits the K12 band (46 schools) and the citywide table includes it, so band rows do not sum to the citywide row.
• Percentiles and SDs are unweighted across schools; weighted and unweighted means diverge (e.g., 32.3 vs 35.7 in 2024-25) because smaller schools tend to have higher CA - both are reported.
• School counts vary slightly by year (1,450-1,470) due to suppression, openings, and closings; year-over-year comparisons are repeated cross-sections, not a fixed panel.
• Per-grade weighted rates use denominator = students contributing per school-grade cell after suppression filtering; suppression removes small cells, which could bias small-grade estimates slightly (PK n = 32,289 is the smallest cell).
• District/borough rankings are point estimates with very different N (e.g., D16: 5,526 students vs D31: 58,697); small districts' rates are noisier. Boroughs were mapped from codes (B=Bronx, K=Brooklyn, M=Manhattan, Q=Queens, R=Staten Island), verified against known district-borough assignments.
• The 23.3% between-district variance share treats districts as fixed groups on unweighted school values; weighting by enrollment or adding grade-band controls would change the split (grade-band composition differs across districts and is itself a CA driver).

Year-over-year stability & noise

School-level chronic absenteeism in NYC district schools is highly stable year-over-year (Pearson r ≈ 0.89-0.94 for consecutive post-COVID pairs, ~0.78 across the 2018-19→2024-25 arc), so a school's CA rate is mostly signal, not noise — but individual YoY changes of a few points are routine (median |change| 2.9pp in 2023-24→2024-25), small schools are roughly 1.5-2x as volatile as large ones, and the worst 2022-23 decile improved ~3.4pp more than average by 2024-25 (classic regression to the mean). 58 schools sat in the worst decile and 83 in the best decile in all four years 2021-22..2024-25 — credible stable outliers. The grade-3→grade-4 pseudo-cohort check (r=0.87, median |change| 4.4pp, only 5.2% of schools with >20% cohort-size churn) confirms persistence holds even within followed cohorts.

Findings

• (1) YoY stability is high and rising: Pearson/Spearman correlations of school-wide CA (ALL subgroup, default universe) for consecutive years are 2021-22→2022-23 r=0.892/rho=0.889 (n=1,462), 2022-23→2023-24 r=0.928/0.928 (n=1,459), 2023-24→2024-25 r=0.936/0.934 (n=1,450). Long arc 2018-19→2024-25: r=0.777/rho=0.789 (n=1,448). By Kane-Staiger standards (test-score gain measures often r<0.5 YoY), CA is a comparatively low-noise school measure; a single year's rate is a reasonable proxy for the school's standing.
• (2) 2023-24→2024-25 change distribution (n=1,450; negative = improving): 42.7% of schools improved by >2pp and 19.5% by >5pp; 21.1% worsened by >2pp and 7.4% by >5pp. Median change -1.4pp, mean -1.4pp, median |change| 2.9pp. So citywide improvement is broad but a fifth of schools still moved the wrong way by a non-trivial margin.
• (3) Noise scales with size: pooling the three consecutive YoY pairs 2021-22..2024-25 (bucketed by earlier-year denominator), SD of YoY change falls from 7.6pp for schools with 100-249 students to 5.0pp for 1000+ (ratio ~1.5x); median |change| falls 4.7pp → 2.3pp (~2x). The <100 bucket (SD 8.0pp) has only 15 school-year obs because the default universe excludes most tiny schools. Latest pair only (2023-24→2024-25): SD 6.8pp (100-249) vs 3.3pp (1000+), ratio ~2.1x. Practical rule: a 5pp single-year swing at a 150-student school is within ~1 SD noise; the same swing at a 1000+ school is ~1.5 SD and more meaningful.
• (4) Regression to the mean is real but does not erase the gap: among 1,450 schools observed in both 2022-23 and 2024-25, the worst 2022-23 decile (mean CA 63.1%) improved by 6.3pp on average vs 2.9pp for all schools (3.4pp excess), while the best decile (mean CA 14.3%) improved only 0.8pp (2.1pp less than average). The worst decile still averaged 56.9% CA in 2024-25 — over 4x the best decile's 13.5% — so the extreme groups compress toward the mean but remain far apart.
• (5) Persistence: of 1,450 schools with non-suppressed CA in all four years 2021-22..2024-25, 58 schools (4.0%) were in the worst within-year decile all four years and 83 (5.7%) in the best decile all four years (vs ~0.01% expected if deciles were random). Relaxing to 3-of-4 years: 112 worst, 133 best. These are the stable-outlier candidates worth naming/profiling.
• (6) Pseudo-cohort sanity check: for 770 default-universe schools with grade-3 CA in 2023-24 and grade-4 CA in 2024-25, Pearson r=0.868, Spearman rho=0.872, mean change -2.6pp, median |change| 4.4pp (larger than the school-wide 2.9pp, consistent with smaller per-grade N — median cohort size 69). Only 5.2% of these school-cohorts had cohort N change by >20%, so same-school grade-to-grade comparisons are usually tracking substantially the same students; the per-grade data supports cohort-following analyses.

Tables

Table 1. Year-over-year correlation of school-wide chronic absenteeism (subgroup ALL, default universe, non-suppressed both years)

Table 2. Distribution of YoY change in CA, 2023-24 → 2024-25 (n=1,450; negative = improvement)

Table 3. Volatility of YoY CA change by school size (pooled consecutive pairs 2021-22..2024-25; bucket = earlier-year CA denominator)

(Latest pair only, 2023-24→2024-25: SD = 6.8pp for 100-249 vs 3.3pp for 1000+.)

Table 4. Regression to the mean: 2022-23 CA deciles followed to 2024-25 (n=1,450 schools in both years)

Table 5. Decile persistence, 2021-22..2024-25 (within-year deciles; 1,450 schools observed all 4 years)

Table 6. Pseudo-cohort check: grade 3 CA in 2023-24 vs grade 4 CA in 2024-25, same school (n=770)

Caveats

• All stats are unweighted across schools (per project convention); samples are schools non-suppressed in BOTH years of each pair, so pair Ns differ slightly (1,448-1,462) and survivorship excludes closed/opened/suppressed schools — likely the most volatile cases, so volatility is, if anything, understated.
• The <100-denominator bucket in Table 3 has only 15 school-year observations (4 in the latest pair) because the default universe excludes D75/D79/alt schools; its SD is unreliable. Use the 100-249 vs 1000+ contrast (~1.5-2.1x) as the headline small-vs-large ratio.
• Pooled bucket SDs in Table 3 mix three year-pairs with different citywide level shifts (mean change ~ -2 to -2.6pp per pair); SDs therefore include common-year shocks, not pure school-level noise. The latest-pair-only row addresses this and shows the same gradient.
• Regression-to-the-mean estimates (Table 4) confound true noise reversion with any real targeted improvement at high-CA schools (e.g., attendance initiatives); this analysis cannot separate the two. The decile-persistence counts (Table 5) are the better evidence that extreme performers are largely real.
• Decile membership uses ntile(10) within each year among non-suppressed default-universe schools, so decile cutoffs move year to year; persistence counts measure relative standing, not fixed CA thresholds.
• Pseudo-cohort grade 3→4 is not a true student-linked cohort: even with <20% N churn, individual students move in/out; 5.2% churn-flag rate uses |N2-N1|/N1 > 0.2 on metric denominators. Gravity toward ES/K8 schools only (n=770).
• Pearson on ranks was used for Spearman (ties handled by rank(); average-rank ties would shift rho only in the 3rd decimal). Numbers reported to 1 decimal except correlations (3 decimals).

SES/demographic prediction (adjustment preview)

Demographics predict roughly half of school-level variance in chronic absenteeism: OLS of CA on [pct_econ_dis, pct_ell, pct_swd, pct_black, pct_hispanic, log(enrollment)] gives R2 = 0.43-0.50 per year 2021-22..2024-25 (0.459 in 2024-25; 0.489 student-weighted). Economic disadvantage is the strongest single correlate (r = 0.53-0.58 with CA). Adjustment works as designed: corr(residual, pct_econ_dis) = 0.000 vs 0.567 raw in 2024-25. Residuals are highly stable year-to-year (corr 0.884 between 2023-24 and 2024-25 model residuals), so a residual-based 'beating the odds' list is not noise-driven. Method: local Postgres (psql) + python3/psycopg/numpy lstsq OLS, default universe (include_in_default_comparisons), subgroup ALL, not suppressed; pct_econ_dis nulls ('Above 95%' top-coding) imputed at 0.95.

Findings

• (1) Pairwise correlations with school-wide CA, default universe, n=1,450/yr: pct_econ_dis is the strongest SES correlate every year (r = 0.581, 0.533, 0.545, 0.567 for 2021-22..2024-25), followed by pct_swd (0.43 falling to 0.37) and pct_black (0.46 falling to ~0.43). pct_ell is near zero pairwise (0.01-0.13, rising over time). total_enrollment is moderately NEGATIVE (-0.28 to -0.32): bigger schools have lower CA.
• (2) Multiple-regression R2 of CA on [pct_econ_dis, pct_ell, pct_swd, pct_black, pct_hispanic, log(total_enrollment)], numpy lstsq OLS, unweighted, n=1,450/yr: 2021-22 = 0.496, 2022-23 = 0.439, 2023-24 = 0.432, 2024-25 = 0.459. Student-weighted (w=CA denominator) 2024-25 R2 = 0.489. Demographics explain slightly less variance post-2021-22 than in the immediate-reopening year.
• (2b) R2 by grade_band, 2024-25: ES 0.569 (n=638), K8 0.593 (n=164), MS 0.498 (n=238), HS 0.420 (n=406). Demographics predict elementary CA much better than high-school CA, so any demographic adjustment leaves the most unexplained (school-attributable or unmeasured) variation at the HS level. K12 has only 4 schools in-universe (not fittable).
• (3) 2024-25 residual fit on denominator>=100 (excludes only 7 of 1,450 schools; fit n=1,443, R2 = 0.458). Most NEGATIVE residuals (beating the odds): three of the top 10 are Eagle Academy for Young Men campuses (29Q327 resid -38.8, 23K644 -27.2, 09X231 -27.2), plus Concourse Village ES (07X359, -34.2) and STAR Early College (17K543, -31.6) -- all high-poverty (econ 74-95%). Most POSITIVE residuals (worse than predicted) are dominated by small Manhattan District 2 high schools: Urban Assembly Business for Young Women (02M316, CA 83.2 vs predicted 43.0, resid +40.2), Bronx Collaborative HS (10X351, +36.3), Murry Bergtraum (02M520, +32.7); 7 of the 10 are HS, 6 of 10 are 02M.
• (4) PEER diagnostic, 2024-25 denom>=100 sample: corr(raw CA, pct_econ_dis) = 0.567; corr(OLS residual, pct_econ_dis) = 0.0000 (zero by construction since econ_dis is a regressor). This is exactly what demographic adjustment does: the raw metric ranks schools substantially by poverty; the residual metric is poverty-orthogonal.
• (5) Stability: same model fit on 2023-24 (n=1,446, R2 = 0.430); for the 1,443 schools in both years, corr(residual 2023-24, residual 2024-25) = 0.884. For context, raw CA itself correlates 0.937 across the same two years. Residuals carry persistent school-level signal, not one-year noise.
• Coefficients of the 2024-25 fit (percent-CA per unit-fraction predictor): intercept 18.3, econ_dis +19.7, ell +3.3, swd +22.4, black +21.4, hispanic +12.3, log(enroll) -2.5. Race coefficients remain large conditional on econ_dis (collinearity caveat applies; descriptive only).
• Data quirk found: school_year_demographics.pct_* are FRACTIONS 0-1 (CA value is percent 0-100), and pct_econ_dis is top-coded at 0.95 with 'Above 95%' stored as NULL (138/190/310/293 schools in 2021-22..2024-25, i.e., 8-18% of the universe, and these schools have HIGHER mean CA: 41.9 vs 34.2 in 2024-25). Dropping them would bias correlations down via range restriction; they were imputed at 0.95. Complete-case sensitivity: corr(CA, econ_dis) shifts by at most +0.026 (e.g., 2024-25: 0.567 imputed vs 0.582 complete-case, n=1,177).

Tables

Table 1. Pairwise correlations of school-wide chronic absenteeism (subgroup ALL, not suppressed) with demographics, default universe, n=1,450 schools/year. pct_econ_dis nulls imputed at 0.95 top-code; complete-case (cc) shown for econ_dis.

Table 2. R2 of OLS: CA ~ pct_econ_dis + pct_ell + pct_swd + pct_black + pct_hispanic + log(total_enrollment). Unweighted across schools (numpy lstsq).

Table 3a. 'Beating the odds' candidates: 10 most NEGATIVE residuals, 2024-25 model (fit and listing restricted to CA denominator>=100; excludes 7 schools). CA, predicted, residual in percentage points; econ% = pct_econ_dis x100 (95.0 = top-coded 'Above 95%').

Table 3b. Worse than demographics predict: 10 most POSITIVE residuals, 2024-25 (same model/filters as Table 3a).

Table 4. What adjustment does + how stable it is (denominator>=100 samples).

Caveats

• pct_econ_dis is top-coded at 0.95 and 'Above 95%' values are stored as NULL (8-18% of schools per year; these schools have higher CA). Nulls were imputed at 0.95, which compresses the top of the poverty distribution and likely UNDERSTATES the true econ_dis correlation and R2 slightly; complete-case correlations run up to +0.026 higher but suffer range restriction. The 95.0 econ% entries in the residual tables are top-coded, not exact.
• Demographic table units: all pct_* fields are fractions 0-1 while CA is percent 0-100; correlations and R2 are scale-invariant, but the printed regression coefficients are 'pct points of CA per unit fraction' (divide by 100 for per-percentage-point effects).
• The denominator>=100 exclusion for the residual preview removed only 7 of 1,450 schools in 2024-25 (and 4 in 2023-24), so it barely changes the fit; it mainly protects the top-10 lists from small-N extremes. Several listed schools are still small (125-200 students), so their residuals have wider sampling error.
• This is a descriptive OLS preview, not the final adjustment model: residual = 0 correlation holds only for variables in the model (econ_dis shown); residuals may still correlate with omitted factors (grade configuration beyond band, screened admissions, temp-housing share, district effects). The positive-residual tail's concentration in small District 2 Manhattan transfer-adjacent high schools suggests school-type confounding worth checking before publication.
• Race/ethnicity coefficients (black +21.4, hispanic +21.4/+12.3 conditional on econ_dis) are collinear with poverty and should not be read causally; whether to include race in a public 'expected CA' model is an editorial decision, not a statistical one.
• schools.grade_band for 09X593 is ES but its name says 'South Bronx International Middle School' -- possible grade_band labeling issue in the schools table, unverified.
• Demographics exist only from 2020-21, so the 2018-19 pre-pandemic baseline cannot be included in this joined analysis; per project conventions 2019-20 and 2020-21 were excluded. Universe is NYC district schools with include_in_default_comparisons = true (no charters, no D75/D79/alt). School-level stats are unweighted unless labeled student-weighted.

Within-school subgroup gaps

Within-school chronic-absenteeism (CA) subgroup gaps for NYC district schools (include_in_default_comparisons=true), 2024-25, using the newly recovered complement subgroups (NOT_ECON_DIS, NOT_ELL, NOT_TEMP_HOUSING, OTHER_ETHNICITY). Citywide student-weighted aggregates closely match published figures (STH 49.0 vs published ~48.7; non-STH 29.5 vs ~30.7). The largest divides are housing (19.5pp) and poverty (18.7pp); ELL is small (5.5pp) and gender is negligible. Within schools (both cells N>=50), the median STH gap is 12.0pp and median poverty gap is 12.8pp — both meaningfully smaller than the citywide gaps, so a nontrivial share of each citywide gap reflects between-school sorting. Poverty gaps are moderately larger in high-CA schools (r=0.27); STH gaps are uncorrelated with school CA (r=-0.00). 24 schools have reversed poverty gaps (< -2pp); the trend table shows both gaps are now WIDER than pre-pandemic (poverty 14.7->18.7pp, STH 18.5->19.5pp) even as overall CA recedes — recovery has been faster for advantaged students.

Findings

• (1) Citywide student-weighted CA 2024-25 (district schools, default universe): ALL 32.3% (N=836,268). Largest gaps: TEMP_HOUSING 49.0 vs NOT_TEMP_HOUSING 29.5 (19.5pp); ECON_DIS 36.5 vs NOT_ECON_DIS 17.8 (18.7pp); SWD 40.5 vs NON_SWD 30.1 (10.4pp); ELL 37.1 vs NOT_ELL 31.6 (5.5pp). Race: BLACK 39.3, HISPANIC 38.9, OTHER_ETHNICITY 26.1, WHITE 23.4, ASIAN 17.8. Gender essentially flat: MALE 32.7 vs FEMALE 32.4 (0.3pp).
• (1b) Match to published citywide figures: our STH 49.0 / non-STH 29.5 vs expected ~48.7 / ~30.7 — within ~1pp, consistent with our universe being district schools only (no charters, excludes D75/D79/alt) and school-level weighted aggregation.
• (1c) 2018-19 baseline (same universe): ALL 25.8; STH 43.1 vs non-STH 24.6 (18.5pp); ECON_DIS 29.6 vs NOT_ECON_DIS 14.8 (14.7pp); SWD 35.5 vs NON_SWD 23.4 (12.1pp); ELL 26.8 vs NOT_ELL 26.3 (0.5pp — ELL flipped from no gap pre-pandemic to a 5.5pp gap in 2024-25). Every subgroup is worse in 2024-25 than 2018-19; NOT_ECON_DIS deteriorated least (+3.0pp) while ECON_DIS rose +6.9pp.
• (2) Within-school gap distributions, 2024-25, both cells denominator >= 50 (cutoff chosen so a 1-student change moves a cell <= 2pp): STH gap (TEMP_HOUSING - NOT_TEMP_HOUSING), 976 schools: p10 +1.6, median +12.0, p90 +27.4, mean +13.9, 6.6% of schools reversed (<0). Poverty gap, 795 schools: p10 +2.9, median +12.8, p90 +28.5, mean +14.4, 4.4% reversed. SWD gap, 1,321 schools: p10 -2.2, median +7.5, p90 +17.2, mean +7.4, 16.7% reversed.
• (2b) Within-school mean gaps (poverty 14.4pp, STH 13.9pp) are 4-6pp smaller than the corresponding citywide weighted gaps (18.7pp, 19.5pp): roughly a quarter to a third of each citywide gap is attributable to disadvantaged students being concentrated in higher-absenteeism schools rather than to within-school differences.
• (3) Are gaps bigger in high-CA schools? Poverty gap: yes, moderately — corr(school ALL CA, poverty gap) = +0.268 (n=795). STH gap: no — corr = -0.004 (n=976); the housing penalty is essentially constant across the school CA spectrum.
• (4) Reversed poverty gaps (ECON_DIS attends BETTER, gap < -2pp, both N>=50): 24 schools of 795 (3.0%). Most extreme: 19K159 P.S. 159 Isaac Pitkin (Brooklyn ES, -12.6pp), 21K238 Anne Sullivan (Brooklyn K8, -9.3pp), 25Q244 Active Learning Elementary (Queens ES, -9.1pp), 24Q019 Marino Jeantet (Queens ES, -8.1pp), 20K105 The Blythebourne (Queens borough field K, ES, -6.8pp). 7 of the top 10 are Queens/Brooklyn elementary or middle schools, several with small NOT_ECON_DIS cells (50-155 students), and Newcomers High School (30Q555, -4.6pp) serves recent immigrants where the non-poverty cell is atypical.
• (5) Gap trend (citywide weighted): poverty gap 14.7 (2018-19) -> 17.3 (2021-22) -> 16.8 -> 18.6 -> 18.7pp (2024-25); STH gap 18.5 (2018-19) -> 15.2 (2021-22) -> 16.9 -> 20.9 -> 19.5pp (2024-25). COVID's immediate effect (2021-22) widened the poverty gap (+2.6pp vs baseline) but NARROWED the STH gap (-3.3pp, because non-STH CA spiked from 24.6 to 38.7). During recovery the advantaged groups improved much faster (NOT_ECON_DIS -9.0pp from 2021-22 peak vs ECON_DIS -7.6pp; non-STH -9.2pp vs STH -4.9pp), so by 2024-25 BOTH gaps sit ABOVE pre-pandemic levels: poverty +4.0pp vs 2018-19, STH +1.0pp. Neither has recovered; the poverty gap is at its 7-year maximum.

Tables

Table 1. Citywide student-weighted chronic absenteeism by subgroup, NYC district schools (default universe), 2024-25 vs 2018-19. Weighted by metric denominator; non-suppressed cells only.

Table 2. Within-school subgroup CA gap distributions, 2024-25 (gap = disadvantaged minus advantaged subgroup, same school; both cells non-suppressed with denominator >= 50).

Table 3. Top 10 schools where economically disadvantaged students attend BETTER than non-disadvantaged peers (poverty gap < -2pp, both N>=50, 2024-25; 24 such schools total).

Table 4. Citywide student-weighted poverty and student-in-temporary-housing (STH) CA gaps by year. 2019-20 (truncated COVID year) and 2020-21 (remote attendance) shown for completeness but excluded from trend interpretation.

*Anomalous COVID-measurement years.

Caveats

• Universe is NYC district schools with include_in_default_comparisons=true (no charters, no D75/D79/alt), aggregated from school-level cells weighted by denominator — so 'citywide' figures will differ ~1pp from DOE-published citywide rates (e.g., our STH 49.0 vs published ~48.7, non-STH 29.5 vs ~30.7); cells suppressed at the school level are also missing from the weighted sums.
• Per-year gaps are computed from cells where both subgroups happen to be non-suppressed at each school independently; the school sets behind ECON_DIS and NOT_ECON_DIS columns in Table 4 are not forced to be identical pairs, though coverage is near-universal (1,385-1,447 schools per cell in 2024-25) so the effect is small.
• Within-school gap cutoff is denominator >= 50 in BOTH cells. This restricts the sample non-randomly: the poverty-gap sample (795 schools) drops very-high-poverty schools lacking 50 non-poor students, and the STH sample (976) drops low-STH schools — within-school means are therefore not from the full universe, and the within- vs citywide-gap decomposition is approximate.
• Reversed-gap list (Table 3) is descriptive, not causal: several schools have small NOT_ECON_DIS cells (50-155 students) where composition quirks (e.g., Newcomers HS's recent-immigrant population, schools near the poverty-flag boundary) can drive the reversal; cause of any individual reversal is not established.
• Gender row NON_BINARY (2024-25: 25.0%, N=8 students, 1 school) is omitted from findings as too small to interpret; it does not exist in 2018-19 data.
• Subgroup rates come straight from the loaded school_year_metrics values; the complement subgroups (NOT_ECON_DIS, NOT_ELL, NOT_TEMP_HOUSING, OTHER_ETHNICITY) were only recently recovered by a loader fix and have not been cross-validated against any external published source (no published school-level complement figures exist to check against).