The challenging interpretation of instrumental variable estimates under monotonicity

causal inference
epidemiology
instrumental variables
mendelian randomization
statistical methods
  • Core Argument: The paper analyzes the interpretation difficulties of Instrumental Variable (IV) estimates that rely on the monotonicity assumption, particularly when the instrument is non-causal (e.g., in many Mendelian Randomization studies).
  • Effect Identified: IV methods typically identify the Local Average Treatment Effect (LATE), which applies only to the subgroup of compliers—those whose treatment status is modified by the instrument.
  • Policy Implications: The LATE is often not ideal for informing clinical or policy decision-making because the complier subgroup is often ill-defined or not the target of the intervention, requiring researchers to be cautious about generalization.
Published

23 January 2026

PubMed: 28379526 DOI: 10.1093/ije/dyx038 Overview generated by: Gemini 2.5 Flash, 28/11/2025

Background: The Local Average Treatment Effect (LATE)

Instrumental Variable (IV) methods are widely used in epidemiology and causal inference to estimate causal effects in the presence of unmeasured confounding. When combined with the necessary assumptions, including the monotonicity assumption (the instrument affects the treatment in only one direction), IV methods typically identify the Local Average Treatment Effect (LATE). The LATE represents the average causal effect of treatment only in the specific subgroup called ‘compliers’—individuals whose treatment status is changed by the instrument.

Methods: Non-Causal Instruments

This paper provides a detailed framework for interpreting LATE under monotonicity, paying particular attention to instruments that are non-causal for the exposure.

  • Non-Causal Instrument Definition: A non-causal instrument is one that affects the outcome only through the exposure, but does not itself cause the exposure. Examples often include proxies for physician preference or certain genetic variants used in Mendelian Randomization (MR) studies.
  • The Problem: When the instrument is non-causal, the interpretation of the complier population—the subgroup to which the LATE applies—becomes highly complex. Little attention has historically been paid to the difficulty of interpreting the LATE when it is defined by adherence to an instrument that doesn’t actually cause the exposure.

Results: Challenges to Interpretation

The paper clarifies the difficulties arising from the LATE’s reliance on the non-causal instrument:

  • Limited Generalizability: Since the LATE applies only to the compliers, it represents a ‘local’ effect that may not be directly transferable to the entire population. This limits its utility for informing broad clinical or public health policy decisions.
  • Ambiguous Subgroup: When the instrument is non-causal, the defining characteristics of the complier population are often unknown or difficult to characterize in a clinically meaningful way. This ambiguity makes the LATE difficult to translate into practical recommendations.
  • Risk of Misleading Inference: Failure to fully account for the LATE’s limited applicability, especially when using a non-causal instrument, risks drawing misleading causal conclusions that overestimate the scope of the treatment effect. This is particularly relevant in MR studies where genetic variants are used as non-causal instruments for environmental or behavioral exposures.

Conclusions and Recommendations

The authors urge researchers using IV methods to exercise greater caution in the interpretation of their estimates, especially in observational settings where instruments are often non-causal. The primary recommendation is to explicitly define the complier population and clearly state the limited scope of the LATE estimate. The paper emphasizes the need for future research to develop methods that can identify causal effects that are more broadly relevant to clinical practice and policy.