Class-size cutoff rules

Why can enrolling one extra pupil suddenly cut a class almost in half?

A school year group has forty pupils in one room. A family moves into the catchment in August and the child is enrolled. In September there are two classes of twenty and twenty-one.

One child arrived and the class each child sits in halved. Nobody decided that this cohort deserved better teaching; nobody weighed the cost against the benefit. A counting rule did it. What kind of rule turns a smooth quantity — how many children live nearby — into a step, and what does that step let us learn that ordinary school data cannot?

Begin with the arithmetic, because the whole shape follows from it. A maximum class size of forty means the number of classes must be the enrolment divided by forty, rounded up. Predicted class size is therefore the enrolment divided by that count. Walk it upward: forty pupils, one class of forty. Forty-one pupils, two classes averaging 20.5. Then class size climbs again as the year group fills — eighty pupils, two classes of forty — and eighty-one pupils drops it to three classes of exactly twenty-seven. The rule produces a sawtooth: long slow climbs punctuated by cliffs at every multiple of forty.

This is Maimonides' rule, so called because the maximum-of-forty formulation is found in the twelfth-century commentary, and it has governed Israeli class formation in the modern period. Joshua Angrist and Victor Lavy used it in 1999 in a way worth reconstructing, because the reasoning is the interesting part.

The ordinary problem with studying class size is that class size is not handed out at random. Small classes cluster in wealthy schools, or in schools that concentrate resources on struggling cohorts — so a raw correlation between class size and attainment is measuring the school, the neighbourhood and the policy, in unknown proportion. What would you need in order to strip all that out?

You would need two groups of pupils alike in every respect except the size of their class. The rule supplies almost exactly that. A year group of thirty-nine and one of forty-one are not meaningfully different populations — one more family moved in, which is close to a coin toss — yet the first sits in classes of thirty-nine and the second in classes of about twenty. Compare their attainment and you have held constant everything that varies smoothly with enrolment: school wealth, neighbourhood, parental choice. That is a regression discontinuity design, and border-prosperity-discontinuity.md is the same logic applied to a line on the ground rather than one in an enrolment register.

But notice what the design quietly assumes, and how it could fail. It needs the enrolment count near the threshold to be unmanipulated. If an ambitious headteacher, knowing the rule, admits one extra pupil to trigger a split, then schools just above forty differ from schools just below by the ambition of their head — and that is not something that varies smoothly. Equally, if parents who understand the rule move their children into the smaller classes after the split, the two groups stop being alike.

Both failures leave fingerprints, which is what makes this testable rather than merely arguable. Manipulation piles enrolments up on the favourable side of the threshold, so plot the number of year groups at each enrolment count and look for a spike just above forty against a hole just below — the standard density check. Sorting shows up as a jump at the cutoff in things class size could not possibly have caused: prior attainment, family background. If those move at the threshold, the estimate is not causal, and no statistical sophistication downstream will rescue it.

Two clarifications keep this honest.

The first is about what such a study can and cannot tell you. A discontinuity identifies the effect at the threshold, for the schools and cohorts sitting near it. It says what happens when a class of forty becomes two of twenty. It does not license a claim about cutting classes from thirty to twenty-five everywhere, and it is silent about what happens when a whole system reduces class sizes at once and must staff the new rooms. tutoring-versus-class-size.md takes up that second question — why the class-size lever is so lumpy, and why small cuts change nothing about how a teacher actually teaches. The fixed point of difference is that the neighbour asks whether the input matters; this piece asks what an administrative rule does, and how a rule can be borrowed as an experiment nobody ran.

The second is that the original finding did not stand as firmly as its fame suggests. Angrist and Lavy reported that the smaller classes created by the rule raised attainment, particularly for disadvantaged pupils. A later re-examination by Angrist with Lavy and colleagues, using more recent Israeli data — the authors' own revisiting, which is the strongest form this correction can take — found the effect much weaker or absent. I am recalling the shape of that literature rather than quoting it, and I will not attach a number to either estimate; the point that survives is that the design is sound and its result proved fragile, and those are separate things.

• How bunching tests distinguish a threshold that is administered from one that is gamed.
• Why an effect measured at a cliff can be real and still be a poor guide to a system-wide policy.

Every term the collection defines is gathered in the glossary.