Expenditure Per Equivalent Pupil

It is not per-pupil operating expenditure. Within the FY2025 Expanded List, where the department publishes the same expenditure total over both divisors, the equivalent-pupil figure is 21% lower at the median: $12,854 against $16,289. verified Quoting one where a reader expects the other understates Ohio school spending by a fifth.

What this repository computed # ↑ Contents

Not what Ohio publishes. Figures derived here from committed fixtures, each one citing the test that reproduces it.

A need-weighted denominator is a need adjustment, and it runs one way # ↑ Contents

This was recorded as an argument. It is now a measurement, and it is larger than the argument predicted. verified computed; see crates/xcheck/tests/report_card_2425.rs

Performance Index against FY2025 operating expenditures
  divided by WEIGHTED ADM      -0.0155    (the published measure)
  divided by UNWEIGHTED ADM    -0.337

One numerator. Two divisors the department publishes side by side in one workbook. The same 607 districts, the same school year, no covariates, no model. The correlation moves from nothing to something because of which column was in the denominator.

This is why OCG White Paper 013 can report Pearson −0.016 and describe the analysis as unadjusted. The word is true of the model — no covariates — and false of the measure. The paper’s stated limitation, “no need-adjusted model in this version,” understates it: there is a need adjustment, it is inside the independent variable, it applies to one side only, and it points in the direction that flattens the result.

The mechanism, measured. The ratio of the two ADM columns is very nearly a poverty index:

weighted ADM / unweighted ADM      median 1.261, range 1.05 to 1.62
  against economically disadvantaged share    +0.800
  against Performance Index                   -0.745

verified Dividing by that ratio is dividing by a proxy for the outcome variable. It removes the thing being measured, and the near-zero coefficient is what is left.

What the corrected relationship is, and is not. Holding the economically disadvantaged share constant, the unweighted-denominator association falls from −0.355 to −0.125, still detectable but small. verified So roughly two-thirds of it is poverty. It is not evidence that spending depresses achievement — the corrected reading points the other way. Ohio spends more per headcount pupil where need is greater, need drives measured attainment, and a negative association is compensation working. The corpus’s standing rule applies: a statistic must name whether its subject is a resource or a compensation, because the direction of “better” is not the same. See skills/dispersion and equity.

The mechanism, confirmed from the other side # ↑ Contents

If the divisor mattered because it is a composition proxy and the outcome was composition-driven, then an outcome with the composition taken out should make the divisor irrelevant. It does. Against Progress (value-added), the two divisors give −0.039 and +0.029 — indistinguishable, where against the Performance Index they give −0.337 and −0.0155. verified The denominator effect is not a general property of the spending measure; it appears only when the outcome is a level.

Only one of the two estimates is stable, and it is not the published one # ↑ Contents

verified crates/xcheck/tests/report_card_2425.rs

scenario                              published    headcount
all districts                          -0.0155      -0.3365
excluding the highest per-pupil        -0.0042      -0.3546
excluding enrollment under 582         +0.0358      -0.3371

verified crates/dispersion

The headcount coefficient moves by 0.018 across every check and never stops saying the same thing. The published coefficient changes sign — the scenarios do not agree on the direction of the association, which is the one thing a coefficient is for — and its magnitude varies by a factor of eight.

Weighting districts by enrollment separates them further: −0.149 published against −0.547 on a headcount denominator. That is a tenfold move in the published figure and a 1.6× move in the other. verified

So “near zero in every case” is stability only in the sense that near-zero numbers stay near zero. The paper’s robustness claim survives as a statement about a quantity too small to have a direction, and not as evidence that its estimate is reliable. This is the sharper objection and the one the corpus should make: not the denominator matters, but one of these two estimates is stable under the paper’s own checks and it is not the one it published.

And a second source agrees # ↑ Contents

The corpus’s independent FY2024 Cupp Report headcount series, a different file for a different year, correlates with the Performance Index at −0.388. verified Two headcount measures from two publications give −0.337 and −0.388. The published equivalent-pupil measure gives −0.016. The odd one out is the one with the weighted divisor.

What the paper’s Table 3 shows # ↑ Contents

Sorting districts by performance rather than spending, the paper reports total spending as “roughly flat” and state-and-local as “moving in the opposite direction, from $12,403 to $13,324.” The second description is wrong in a way this data can correct. The row is:

performance quintile      Q1(low)    Q2      Q3      Q4      Q5(high)
state & local / pupil     $12,403 $12,436 $12,236 $12,340   $13,324
federal / pupil            $1,108    $704    $686    $486      $382

State-and-local is flat within $200 across four quintiles and then steps $984 at the top. Not a gradient, a step — and the corpus names the mechanism: local revenue per pupil tracks assessed valuation per pupil at +0.807. verified “Roughly flat” total spending is a need channel declining with performance plus a wealth step at the summit. Both gradients show directly in the computed data: FY2025 spending per unweighted pupil correlates with the disadvantaged share at +0.345 and with valuation per pupil at +0.281, and the FY2024 Cupp series gives +0.383 and +0.279 independently. verified Ohio’s poorest and wealthiest districts both spend above its middle, for entirely different reasons, and a single coefficient across that shape is two relationships canceling.

And the corrected number does not survive a change of outcome # ↑ Contents

Holding disadvantage constant, per-pupil spending correlates −0.125 with attainment level and +0.146 with growth. verified The negative sign recorded above belongs to the level measure, not to spending. Anyone quoting −0.337 as what Ohio’s spending achieves is making the same category of error this node was written to correct, one step further along.

A concrete case # ↑ Contents

Toledo City divides $440.2 million of FY2025 operating expenditure by 21,160 enrolled pupils or by 30,758 weighted ones: $20,805 or $14,312, a 45% spread on one district in one year. Both numbers have been published about Toledo — the second in White Paper 013, the first three weeks later in fact-check RL-2026-021 — by the same author, with neither noting the other’s basis. verified The abstract argument on this node has a district attached to it.

The fact-check also shows the rule generalizing past this metric. It compares special-education spending per total pupil across two districts whose disability shares are 21.9% and 11.3%, and reads the residual as heavier need in the urban district. Per student with a disability the comparison reverses. A denominator does not have to be need-weighted to be the wrong population. inference

A hypothesis tested and dropped # ↑ Contents

The obvious suspicion was the temporary transitional aid guarantee: 48.3% of districts are funded off-formula against an un-rebased FY2020 baseline, so much of the spending variation being correlated ought to be a frozen historical artifact. It is not the explanation. Splitting the 607 districts on FY2027 guarantee status leaves the denominator gap intact in both halves — guaranteed districts −0.389, on-formula districts −0.299, and both near zero on the weighted divisor. verified The guarantee moves a great deal of money and does not move this. Recorded because a negative result that was expected to be positive is worth more than an untested suspicion.

One guarantee finding does survive, and it corroborates the corpus: guaranteed districts average a Performance Index of 89.8 against 85.7 for on-formula districts. verified The guarantee holds up property-wealthy districts, and property-wealthy districts score higher.

The weight schedule, measured rather than missing # ↑ Contents

This node’s definition records that DEW’s weight schedule “is not held in this corpus”. It still is not. But the report card publishes the weighted count, the headcount and all three shares the weighting is said to answer to, so how much of the schedule is missing is a measurable quantity rather than an unknown one. Three shares explain 91.4% of the cross-district variation in the weight, over the same 606 districts every model here fits. verified crates/xcheck

extra pupils per pupil, per point of share, FY2025
  economically disadvantaged        +0.00206
  English learners                  +0.00442
  students with disabilities        +0.01185
carried by the mean district         0.283

A point of disability is worth 5.7 points of economic disadvantage, and a district at 100% disadvantage carries 0.21 extra pupils per pupil on that account alone. The shape is the one the definition asserts, so the definition’s account of what the weighting responds to is now corroborated rather than merely stated.

The missing 8.6% is not rounding, and that is the interesting part. weighted_adm is published to whole pupils, but dividing the expenditure total by the published per-equivalent-pupil figure recovers it to a fraction of a pupil, and the fit is identical either way. verified The residual is structural: it correlates +0.34 with log enrollment, and the 57 districts above 5,000 pupils are under-predicted by +0.021 against −0.005 for everyone else. So the department’s weighting is not an additive function of these three published rates. verified The likeliest reason is one this corpus already documents elsewhere — disability is priced in six categories that cost different amounts, and the report card publishes a single share — but which of the three rates hides the structure is open.

Half the English-learner column is suppressed, and reading the suppression as zero is doing real work: 303 of the 606 districts carry no published share. Refitting on the 303 that do moves the English-learner coefficient from 0.00442 to 0.00346 — smaller, as imputing a floor would predict, and nowhere near zero — while the other two barely move and the fit improves to 92.5%. verified crates/xcheck The imputation is therefore a bound on that one coefficient, not a load-bearing assumption.