Clustering at the school, not the child, showed deworming's payoff on untreated neighbors

the problem

Individual trials found tiny deworming effects and the cheap pill looked worth the effort, so donors under-invested

background

what everyone would do

The standard move is randomized treatment at the individual level, and it is the textbook answer — but it pits two children against the same worm reservoir, so the treated child is reinfected by the untreated one and the control child benefits from the treated one. Individual randomization actively erases the effect it exists to find.

what they saw

An externality has a carrier, and you see it only by randomizing that carrier. The worm travels through the shared school pool, so the cluster to randomize was the school, not the child.

the move

Miguel and Kremer ran the evaluation the other way: instead of randomizing the drug to single children inside each school, they randomly phased deworming across whole schools. This preserved the transmission channel, so treated and untreated schools could be compared while the worm pool was actually broken at the school boundary.

why it works

Deworming works through a double channel that individual trials split in half. Treated children shed fewer eggs, so children around them are infected less; clustering treatment at the school boundary let the worm pool drop below the reinfection level, which is precisely when mass treatment flips from helping the dosed child to protecting everyone in the pool. Randomizing between schools measured both effects at once — the direct dose effect and the transmission spillover — while individual trials cancelled the second against itself and reported only the weak remainder. The same design that made the effect hard to see was the design that, inverted, made it huge.

the payoff

Deworming cut absenteeism by a quarter; untreated neighbors gained because transmission fell. The externalities justified full subsidy.

where it breaks

It fails when there is no shared coupling to exploit — for interventions whose benefit is purely individual, school-level clustering buys nothing. It needs the off-target spillover to be large enough to dominate the direct effect, and it needs a boundary (a school, a neighborhood) the transmission respects; where people cross boundaries freely the externality leaks away. And the subsidy logic depended on the untreated neighbors being reachable for free — where the disease reservoir lies outside the treated pool entirely, treating inside it cannot break transmission.

what came after

The 2004 Econometrica paper became the empirical backstop for mass deworming worldwide — WHO school-based deworming programs and GiveWell's top-rated charities rest on it — and a ten-year follow-up in the QJE found long-run impacts on earnings, with a conservative 32% internal rate of return.

references

1. [1]NBER Working Paper 8481 — Worms: Education and Health Externalities in KenyaNational Bureau of Economic Research, 2001nber.org
2. [2]AEA RCT Registry — Worms: Identifying Impacts on Education and Health in the Presence of Treatment ExternalitiesAmerican Economic Association, 2016socialscienceregistry.org

keep it

same kind of clever

• 2017 · San Diego County Medical Examiner's OfficeDoctors kept prescribing opioids to patients who later overdosed, and never found out — so the county coroner started telling them, one named patient at a time.
• 1086 · William the Conqueror (Domesday Survey)Twenty years after conquering England, William still couldn't tax it fairly — so he sent commissioners to make every landholder's neighbors swear to what he actually owned.
• 2002 · Chicago Public Schools (with researchers Brian Jacob and Steven Levitt)You can't catch a cheating teacher by watching one classroom — but you can catch them by watching the pattern across thousands