Behavioral Science Dictionary

Local average treatment effect

Also known as: Complier average causal effect

Methods & Evidence

The causal effect estimated only for those whose behavior responds to the nudge or instrument.

What it means

The local average treatment effect is the causal effect of a treatment estimated specifically for the subgroup of people whose treatment status is actually changed by an instrument or encouragement—the 'compliers'—rather than for everyone in the study. It arises in instrumental-variable and encouragement designs, where some assigned-to-treatment people would not take it and some assigned-to-control would get it anyway, so randomization affects only those on the margin who do what their assignment tells them. Under assumptions including a valid instrument and 'monotonicity' (no one does the opposite of their assignment), the instrumental-variable estimate recovers the average effect for these compliers, not the population-wide average treatment effect. The crucial nuance is interpretive: the LATE may differ from the effect on always-takers, never-takers, or the whole population, and the complier group is defined by the particular instrument, so a different instrument can yield a different LATE. This limits generalizability even as it secures internal validity. It matters because much credible causal evidence from natural experiments and imperfect-compliance trials estimates a LATE, and mistaking it for the overall effect can mislead policy.

Examples

If a mailed voucher induces only some families to enroll their children in a program, the instrument identifies the program's effect on those induced families—the compliers—not on families who would never or always have enrolled.

A text nudging patients to book a screening measures the screening's effect only on those who booked because of the text, not on people who would have gone regardless.

Randomly offering a free coding course estimates its effect on people who enrol only when offered; the keen who would have paid and the uninterested who never enrol stay unmeasured.

First described in Guido Imbens & Joshua Angrist (1994).

← All 1001 terms