Identification and Estimation of Local Average Treatment Effects

Joshua D. Angrist, Guido W. Imbens

Abstract

In their seminal 1991 working paper (later published in Econometrica), Joshua Angrist and Guido Imbens address a fundamental challenge in causal inference: how to interpret instrumental variable (IV) estimates when treatment effects vary across individuals. They demonstrate that under specific conditions—most notably "monotonicity"—IV identifies the Local Average Treatment Effect (LATE), which is the average effect for the subpopulation of "compliers" who are induced to change their behavior by the instrument.

Introduction: Beyond Constant Treatment Effects

Before the work of Angrist and Imbens, the standard application of instrumental variables often relied on the assumption of a constant treatment effect—the idea that a program or intervention affects everyone in exactly the same way. In reality, people are heterogeneous; a job training program might significantly help one person while having no impact on another.

When treatment effects are heterogeneous, it is no longer clear what a standard IV estimator represents. Does it measure the average effect for the whole population? For those who took the treatment? Angrist and Imbens provided the mathematical framework to answer this, fundamentally shifting the focus of econometrics toward the "Potential Outcomes" framework (often called the Rubin Causal Model).

The Core Assumptions

To identify a causal effect using an instrument ($Z$) for a treatment ($D$), the authors propose two primary conditions:

  1. Independence and Relevance: The instrument must be as good as randomly assigned (independent of potential outcomes and treatment assignments) and must be correlated with the treatment ($D$).
  2. Monotonicity: This is the critical innovation of the paper. Monotonicity assumes that the instrument affects everyone in the same direction. For example, if a voucher for a private school (the instrument) makes some students more likely to attend, it should not make any student less likely to attend.

Defining the Sub-populations

The LATE framework categorizes the population into four latent groups based on their response to an instrument:

  • Compliers: Those who take the treatment if encouraged by the instrument, but not otherwise.
  • Always-takers: Those who take the treatment regardless of the instrument.
  • Never-takers: Those who never take the treatment regardless of the instrument.
  • Defiers: Those who do the exact opposite of what the instrument encourages.

The authors show that if the Monotonicity assumption holds, "Defiers" are assumed not to exist. In this case, the IV estimator identifies the effect specifically for the Compliers.

Key Mathematical Results

The paper establishes two major theorems that define the behavior of IV estimators in the presence of heterogeneity.

The Wald Estimator and LATE

For a binary instrument, the Local Average Treatment Effect ($\alpha_{LATE}$) is identified by the ratio of the instrument's effect on the outcome to its effect on treatment participation. This is often referred to as the Wald formula:

$$\alpha_{LATE} = \frac{E[Y|Z=1] - E[Y|Z=0]}{E[D|Z=1] - E[D|Z=0]}$$

This result proves that the IV estimate is a valid causal effect for the group of people whose treatment status is changed by the instrument.

IV as a Weighted Average

In cases where the instrument is not binary (e.g., a continuous variable or multiple categories), the authors prove that the standard Two-Stage Least Squares (2SLS) estimator does not identify a single LATE. Instead, it converges to a weighted average of LATEs at different levels of the instrument. The weights are determined by the degree to which the instrument shifts the probability of treatment at each point.

Applications and Legacy

While the paper is theoretical, it provides a lens through which to view major empirical studies. For instance, in the Vietnam-era draft lottery, the "instrument" (the lottery number) only affects the earnings of those who served in the military because they were drafted (the compliers). It tells us nothing about the effect of service on those who would have volunteered anyway or those who would have avoided service regardless of their number.

The "LATE" interpretation has become the standard way to report IV results in social science. Its importance was formally recognized in 2021 when Angrist and Imbens were awarded the Nobel Prize in Economic Sciences for this contribution.

Limitations

Despite its power, the LATE framework has limitations:

  • External Validity: Because the effect is "local" to compliers, it may not generalize to the rest of the population (the "Always-takers" or "Never-takers").
  • Untestability: The Monotonicity assumption cannot be directly tested with data; it requires a strong theoretical justification based on the context of the study.
  • Weak Instruments: If the instrument only affects a very small number of compliers, the resulting estimates can be highly unstable and sensitive to small fluctuations in the data.

Summary updated

Abstract

We investigate conditions sufficient for identification of average treatment effects using instrumental variables. First we show that the existence of valid instruments is not sufficient to identify any meaningful average treatment effect. We then establish that the combination of an instrument and a condition on the relation between the instrument and the participation status is sufficient for identification of a local average treatment effect for those who can be induced to change their participation status by changing the value of the instrument. Finally we derive the probability limit of the standard IV estimator under these conditions. It is seen to be a weighted average of local average treatment effects.
BibTeX
@inproceedings{angrist1994identification,
  author = {Joshua D. Angrist and Guido W. Imbens},
  title = {{Identification and Estimation of Local Average Treatment Effects}},
  booktitle = {Econometrica},
  year = {1994},
  doi = {10.2307/2951620},
}
Figure