Progress (Value-Added)

The corpus acquired the Index first and immediately found that 71.6% of its cross-district variance is the economically disadvantaged share — so most of what it appears to say about schools is a statement about who attends them. Progress exists to strip that out. It does not fully succeed: it still tracks disadvantage at −0.325, about 10.6% of its variance. verified But an order of magnitude less is a different measure, not a refinement of the same one, and the two correlate with each other at only +0.375.

The denominator question does not arise here # ↑ Contents

The finding that dominates expenditure per equivalent pupil — that dividing by a need-weighted pupil count moves the headline from −0.337 to −0.015 — has no counterpart on this measure. Against growth, the two divisors give −0.039 and +0.029: indistinguishable. verified That is the strongest available confirmation of why the denominator mattered in the first place. It mattered because the outcome was composition-driven and the divisor was a composition proxy. Take the composition out of the outcome and the divisor stops doing anything.

The department publishes two of this measure # ↑ Contents

Value-added is reported both as a three-year average — the headline figure, and the one every claim here rests on — and over a single year. The second was extracted into project::outcomes from the beginning and reached no consumer until now. verified

What this repository computed # ↑ Contents

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

The sign of the spending relationship depends on which outcome is asked # ↑ Contents

verified computed; see crates/xcheck/tests/report_card_2425.rs

per-pupil spending (headcount denominator) against...
                                    raw      holding disadvantage constant
  Performance Index  (level)      -0.337              -0.125
  Progress effect size (growth)   +0.018              +0.146

Same spending variable, same 605 districts, same single control. Against attainment level the association is negative; against growth it is positive. Neither coefficient is large — the partial correlations are worth about 1.6% and 2.1% of variance respectively — and neither is causal. What they establish is narrower and firmer: the negative association this corpus recorded against the Performance Index is a property of measuring level, and it does not survive the switch to a measure of gain.

The raw growth figure is near zero for a legible reason rather than because nothing is there. Poorer districts spend more per headcount pupil (+0.345) and show lower growth (−0.325), so the two effects cancel. Removing the confounder does not weaken the relationship; it uncovers one. inference

The need-adjusted model, and the sign difference surviving it # ↑ Contents

Controlling for economic disadvantage, English-learner and disability shares, log enrollment, and assessed valuation per pupil — the specification OCG White Paper 013 named as its own priority next step — the standardized spending coefficient is: verified

on the Performance Index (level)      -0.073   t = -2.84   model R2 0.752
on Progress effect size  (growth)     +0.209   t = +4.60   model R2 0.227

The growth coefficient strengthens at every step as controls are added — +0.016 with spending alone, +0.147 adding disadvantage, +0.209 with the full set. That is the signature of a relationship being uncovered by controls rather than manufactured by them; a coefficient that appears only in the saturated model is the pattern to distrust, and this is its opposite. inference

The published measure is a three-year average, so the year-matched check matters. The overview effect size is identical to the three-year gain (r = 1.0000), which pairs three years of growth with one year of spending. Re-running the full model on the one-year gain gives a standardized coefficient of +0.269, t = +5.75 — same sign, slightly larger. The mismatch was not carrying the result. verified

And wealth separates the two measures cleanly. In the same model, assessed valuation per pupil predicts attainment level (t = +3.12) and does not predict growth at all (t = +0.06). verified Property wealth tells you where a district’s students are, not how far they moved — which is close to the sharpest statement of the level/gain distinction this data supports, and bears directly on equity.

What this does not license. A standardized coefficient of +0.21 in a single-year cross-section is not evidence that spending raises achievement, and six controls do not make it one. Adding controls removes the part of an association the controls explain; it does not turn the remainder into an effect. What remains unmodeled includes district typology and regional labor cost. inference

What the money was spent on, which used to head that list # ↑ Contents

The largest of those omissions was recorded as “what the money was actually spent on, which no column in this workspace measures. A district that spends more on instruction and one that spends more on transportation are the same observation here.” The department publishes that column and this repository has held it since expenditure-functions-fy25 was committed: the same FY2025 operating total, split into the eleven functions it is reported under, per pupil. It had been used to compare two named districts and never entered into a model. verified

The growth relationship is the classroom half of the money and nothing else. Replacing the total with the department’s own two-way split, on the same 606 districts and the same five controls: verified crates/dispersion

against the three-year Progress effect size, standardized
  classroom instruction per pupil          +0.243    t +5.11
  everything else per pupil                −0.032    t −0.61

The +0.209 above is those two averaged under a constraint that they move together. Released from it, one of them is the whole result and the other is indistinguishable from zero — so the two districts in the sentence are not the same observation, and the difference between them is the finding rather than a caveat. verified

And composition carries information the level of spending does not. Entering total spending and the classroom share together asks whether how a district spends predicts anything once how much is held fixed:

outcome                total operating        classroom share
three-year growth      +0.2309  t +5.04       +0.1215  t +2.89
attainment level       −0.0562  t −2.18       +0.0948  t +4.01

On level the two signs differ in the same model, on the same districts: spending more associates with lower attainment and spending a larger share of it in the classroom associates with higher attainment. On level, composition is the firmer of the two. verified

Two things this is not. It is not a lever. The classroom share correlates +0.38 with log enrollment and −0.40 with transportation spending per pupil — a small rural district buses children a long way and cannot spend that money twice — so part of the share is geography. Enrollment is already a control and entering transportation spending directly leaves the share coefficient at +0.112, t +2.48, which is the right check and still a description. verified And it is not equally firm everywhere: on the 303 districts whose English-learner share the department did not suppress, the coefficient falls to +0.087 and t to +1.56. The estimate moves little while the sample halves, which reads as power rather than contradiction — and it is below the conventional bar. verified

The example the corpus chose was the wrong one. Entered function by function against growth, pupil transportation is −0.035 and not distinguishable from zero. The function that is reliably negative is operations and maintenance of plant, at −0.126, t −2.97; instruction is +0.1944, t +4.52. Transportation spending is not what separates these districts and building maintenance is. verified

Composition also varies far less than level does — the median district spends 67.0% of operating money in the classroom, and that share’s coefficient of variation is under half the spending total’s. Two facts sit beside it. Property wealth does not predict the split at all, which pairs with wealth predicting level and not growth above. And districts that spend more spend a smaller fraction in the classroom, at −0.337, which is why the total understates the classroom coefficient rather than overstating it. verified

The honest statement is unchanged in substance and firmer in support: on Ohio’s growth measure the relationship is not negative, and it survives the controls that were supposed to explain it away. That corrects a widely quoted framing. It does not settle whether money works. inference

A cautionary property that is easy to miss. Overall Composite is a precision-scaled statistic — a gain divided by its standard error — so it rises with the number of tested students and correlates with enrollment at +0.244. Overall Effect Size is the standardized gain and correlates at +0.155. The two correlate with each other at +0.917, which makes the choice look immaterial. It is not: ranking districts on the composite ranks them partly by size. Every figure in this corpus uses the effect size. verified

How far the two published forms agree # ↑ Contents

They agree wherever agreement carries information. Of the 534 districts printing a non-zero value on both, 44 point opposite ways, and of those 44 exactly 0 have both magnitudes above 0.05: each is within 0.04 of zero on one of the two, which is no measured growth there and a sign that is arbitrary. The correlation across those 534 is 0.903. verified computed in crates/bundle from the 2024-25 report card

So the smoothing choice is safe for any district with real movement, and the three-year figure remains the right default: a single year of value-added is noisy enough that the department smooths it deliberately. What the choice does decide is the sign for districts sitting on zero, and the sign is what gets reported about them. inference

  mean    -0.003
  median  +0.000
  sd       0.083
  range   -0.29 to +0.29

  economically disadvantaged share (FY24)        -0.325
  Performance Index 2024-25                      +0.375
  assessed valuation per pupil (FY23)            +0.205
  federal expenditure per equivalent pupil       -0.206
  spending per unweighted pupil                  -0.039
  spending per weighted pupil                    +0.029
  prior-year (FY24) spending per pupil           +0.029

  published overview effect size  ==  three-year gain   (r = 1.0000, sd 0.0829)
  two-year gain                       r +0.968 with it   sd 0.0899
  one-year gain                       r +0.904 with it   sd 0.0989