Local Average Treatment Effects under Imperfect Compliance

15 concepts11 papers1 proposal

In an ideal randomized controlled trial, participants do exactly what they are told. If they are assigned to the treatment group, they take the treatment; if they are assigned to the control group, they do not. In this perfect world, calculating the causal effect of an intervention is simple: you just subtract the average outcome of the control group from the average outcome of the treatment group.

But human behavior is rarely so cooperative. In real-world experiments, researchers frequently encounter imperfect compliance. Some individuals assigned to the treatment group drop out or refuse to participate (non-compliers), while others assigned to the control group manage to obtain the treatment anyway.

If we simply compare those who actually took the treatment against those who did not, we destroy the benefits of randomization, reintroducing selection bias. If we instead compare people based strictly on their original assignment—an approach known as Intention-to-Treat (ITT) analysis—we dilute the estimated effect because many people didn't actually experience the intervention. How do we recover a meaningful, unconfounded causal effect of the treatment itself when compliance is imperfect?

The solution lies in a foundational framework that leverages random assignment as an instrumental variable to isolate a specific, clean causal signal: the Local Average Treatment Effect (LATE), also known as the Complier Average Causal Effect (CACE).

The Core Intuition: Partitioning the Population

To understand why this approach works, we must look at how a population responds to an external nudge or assignment. In their foundational work, Identification and Estimation of Local Average Treatment Effects, Joshua D. Angrist established that we can conceptually divide a population into four distinct latent subgroups based on how their behavior changes in response to an instrument (such as a randomized encouragement or assignment):

  1. Compliers: Individuals who take the treatment if they are assigned to it, but do not take it if they are not. Their behavior is directly altered by the instrument.
  2. Always-Takers: Individuals who will find a way to take the treatment regardless of their assignment.
  3. Never-Takers: Individuals who will refuse the treatment regardless of their assignment.
  4. Defiers: Individuals who do the exact opposite of what they are assigned to do.

To make causal inference possible, researchers typically invoke a monotonicity assumption, which rules out the existence of defiers.

Now, consider what happens when we randomize assignment. Because assignment is random, the proportion of always-takers, never-takers, and compliers will be identical (on average) in both the assigned and unassigned groups.

When we shift the assignment from "not assigned" to "assigned," what changes?

  • The never-takers do not change their behavior; they take nothing in both states.
  • The always-takers do not change their behavior; they take the treatment in both states.
  • Only the compliers change their behavior—moving from untreated to treated.

Because the behavior of always-takers and never-takers remains completely static across both groups, any difference in the average outcome between the assigned and unassigned groups must be driven entirely by the compliers.

By taking the overall difference in outcomes (the ITT effect) and dividing it by the difference in treatment take-up rates (the proportion of compliers), we scale the effect back up to discover the true causal impact of the treatment on the subpopulation of compliers. This is the Local Average Treatment Effect.

LATE in Practice: From AI to Public Health

This mathematical machinery allows researchers to maintain rigorous causal boundaries across diverse empirical settings where compliance cannot be strictly enforced.

For instance, in a large-scale field experiment evaluating the integration of generative AI into corporate workflows, Eleanor Wiske Dillon et al. utilized an instrumental variables approach to calculate LATE. Because giving employees access to an AI tool does not guarantee they will actually use it, evaluating the raw assignment would understate the tool's true impact. By focusing on the compliers, the authors isolated the specific productivity shifts caused by actual tool utilization.

Similarly, Cameron Raymond et al. deployed both ITT and LATE analyses to evaluate automated versus manual pandemic contact tracing. While ITT provided policy-relevant estimates of deploying the system at scale, the LATE framework revealed the precise epidemiological impact on the subpopulation that actually complied with the automated surveillance prompts.

The same principle underpins clinical and social policy trials. When evaluating complex interventions like chronic pain self-management courses (Brennan C Kahan et al.) or state-wide food assistance outreach programs (Elizabeth Stuart et al.), participants frequently drop out or fail to engage. By treating the initial random assignment as an instrument, these studies successfully recover the CACE, stripping away the confounding bias introduced by participant non-compliance. Even when data goes missing alongside non-compliance, specialized moment and maximum likelihood estimators developed by Sierra M. Li can preserve the integrity of these subpopulation estimates.

The Trade-off: Internal vs. External Validity

The precision of LATE comes with a significant conceptual trade-off: it is explicitly local. It identifies the causal effect only for individuals whose behavior can be modified by the specific instrument being used. It tells us nothing about the treatment effect for always-takers or never-takers, for whom the instrument is powerless.

As synthesized by Jeffrey M. Wooldridge, this realization was a cornerstone of the "Credibility Revolution" in econometrics, shifting the field away from making sweeping, unprovable assumptions about population-wide effects toward prioritizing rock-solid internal validity.

This local focus has drawn critiques from researchers who argue that policy-makers need to know the Average Treatment Effect (ATE) for the entire population, not just a narrow slice of compliers. Defending the framework in Better LATE Than Nothing and a subsequent Rejoinder, Guido Imbens argued that prioritizing a well-defined, internally valid estimate for a specific subpopulation is fundamentally more valuable than relying on untestable assumptions to force a population-wide estimate that may be entirely biased. The specificity of LATE is not a bug; it is a mathematical guarantee of clarity.

Modern Frontiers: Continuous Margins and Full Distributions

Modern causal inference has expanded the LATE framework beyond simple binary assignments and simple averages.

When treatment selection is influenced by continuous factors rather than a binary switch, researchers use the Local Instrumental Variable (LIV) approach. As demonstrated by David G. Lugo‐Palacios et al., the LIV method allows researchers to estimate Marginal Treatment Effects (MTE). By evaluating individuals who sit exactly at the margin of choosing a treatment, this method maps out how treatment effects vary continuously across a population, even when unobserved factors influence both selection and outcomes.

[Binary Instrument]  ──► Isolates Compliers ──────────► Local Average Treatment Effect (LATE)
[Continuous Margin]  ──► Isolates Marginal Choices ──► Marginal Treatment Effect (MTE) via LIV

Furthermore, researchers are no longer content with just estimating the average effect for compliers. In many scenarios, a policy might leave the average outcome unchanged while radically altering inequality or variance. To capture this, Undral Byambadalai et al. introduced a framework for estimating Local Distributional Treatment Effects (LDTE) under imperfect compliance. By marrying instrumental variables with machine learning-based regression adjustments, this approach reconstructs how an intervention shifts the entire probability distribution of an outcome for the complier subpopulation.

By focusing squarely on the subpopulation moved by an instrument, the LATE framework and its modern extensions turn the problem of human non-compliance into an asset, providing a mathematically rigorous foundation for discovering what works, for whom, and to what extent.

Key Foundational Work

Identification and Estimation of Local Average Treatment Effects (1995) — contributed 3 concept elements to this direction.
ITT analysis of randomized encouragement design studies with missing data (2005) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_local_average_treatment_effect_16fe53cb,
  title = {local average treatment effect},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/local-average-treatment-effect/rdr_16fe53cb53bd4d14ae286498cbdb9fc6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_local_average_treatment_effect_16fe53cb_v1,
  title = {local average treatment effect},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/local-average-treatment-effect/rdr_16fe53cb53bd4d14ae286498cbdb9fc6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake