Interpretation and Potential Biases of Mendelian Randomization Estimates With Time-Varying Exposures

• Core Problem: This study addresses the difficulty of interpreting Mendelian Randomization (MR) results when the exposure (e.g., smoking) is time-varying and changes over an individual’s life course.
• Interpretation: Standard MR estimates for a time-varying exposure should be interpreted as the causal effect of an intervention on the long-term average or cumulative lifetime exposure, not the effect of intervening at a specific moment (e.g., quitting at age 40).
• Recommendation: Researchers must be explicit about this interpretation, as the estimate may not align with the causal effect relevant to clinical or public health policy interventions, necessitating the development of MR methods based on g-methods for time-varying treatments.

PubMed: 30239571 DOI: 10.1093/aje/kwy204 Overview generated by: Gemini 2.5 Flash, 28/11/2025

Key Findings: MR and Time-Varying Exposures

This paper explores the interpretation and potential biases of Mendelian Randomization (MR) estimates when the exposure of interest is time-varying (changes over an individual’s life course). The often-cited advantage of MR is that it estimates a “lifetime effect,” but this term is frequently vague. The authors clarify that standard MR estimates typically reflect the effect of a time-fixed intervention—specifically, an intervention on the long-term average of the exposure—rather than the effect of an intervention at a specific time point. This interpretation is often inconsistent with the target of clinical or public health interventions.

Methods and Study Design

The study used an empirical example with data from the Rotterdam Study to illustrate the problem. They applied standard MR to examine the effect of smoking (a classic time-varying exposure) on blood pressure.

• Standard IV/MR Model: The traditional Instrumental Variable (IV) and MR framework is designed for estimating the effect of a point or time-fixed exposure, not a dynamic, time-varying exposure.
• Interpretation of Standard MR: When applied to a time-varying exposure (like smoking duration or intensity), the resulting MR estimate is interpreted as the causal effect of an intervention that fixes the exposure status to be constantly high (or low) from birth. This is equivalent to estimating the effect of an intervention on the cumulative or long-term average of the exposure.

Potential Biases

The authors identify two key sources of bias when standard MR methods are applied to time-varying exposures without proper consideration:

1. Misalignment with Target Causal Effect

The primary issue is the misalignment between what standard MR estimates and what is clinically relevant. * Clinical Target: Policymakers and clinicians are usually interested in the causal effect of intervening at a specific time (e.g., quitting smoking at age 40) or intervening on the intensity (e.g., reducing alcohol intake by one drink per day). * MR Estimate: Standard MR estimates the effect of a lifetime difference in exposure average, which may not be the relevant policy parameter, potentially leading to misleading conclusions about the impact of a feasible intervention.

2. Time-Varying Confounding

If there are time-varying confounders that are themselves affected by prior exposure, standard MR estimates can be biased. For example, a genetic variant associated with smoking might also be associated with factors that influence both later-life smoking and the outcome (blood pressure). The standard MR assumptions do not adequately account for this complex, time-dependent structure.

Conclusions and Recommendations

The paper concludes that while MR remains a powerful tool, researchers must be more explicit and cautious when applying it to time-varying exposures.

• Explicit Interpretation: Researchers should clearly state that standard MR estimates for time-varying exposures reflect the causal effect of an intervention on the lifetime average or cumulative exposure, not the effect of an intervention at a specific point in time.
• Causal Models: Future methodological work should focus on leveraging the causal inference tools developed for time-varying treatments (e.g., g-methods) to create MR-based methods that can estimate more policy-relevant parameters, such as the effect of initiating or stopping an exposure at a specific age.