entered social policy from medicine and brought their central virtue with them.
Assigning participants by chance removes the that makes most policy evaluation uninterpretable.
Without it, a programme that helps motivated applicants looks effective when it has merely selected well.
The method has settled several long arguments, particularly about and about schooling inputs.
Its limits are equally well established and are less often stated by those who commission trials.
A trial estimates the effect of a specific intervention in a specific place at a specific time.
Whether that estimate travels to another country depends on the , which the trial itself does not reveal.
Two programmes with identical results may work through entirely different channels, and only one may survive a change of context.
Scale introduces a second problem that no can detect.
A training scheme that places its graduates when it is small may simply move the queue when it is large.
Effects that operate through vanish once everybody receives the treatment.
There is also a question of what can be randomised at all.
Central bank independence, a constitutional provision or a trade agreement cannot be assigned by lottery.
A that ranks trials above all else therefore quietly narrows the policy agenda to what fits the method.
Well-run programmes now combine trials with and with administrative data over longer periods.
The trial answers whether something worked; the other two suggest why and for how long.
Treating the first question as the whole of evaluation is the most common error in the field.
It is also the error most likely to be rewarded, since a single number is what a minister can use.