TL;DR
While Mendelian Randomization (MR) is a powerful tool for uncovering causal relationships, extending it to study gene-environment (GxE) interactions reveals a stark divide: in linear models, standard techniques work beautifully, but in logistic models for binary outcomes, "non-collapsibility" introduces persistent biases that challenge even the best statistical methods.
What Problem Does This Paper Address?
In epidemiology, we often want to know how a specific environment (like BMI or smoking) interacts with a person's genetic makeup to influence disease risk. However, environmental factors are notoriously "confounded"—they are tangled up with unmeasured variables like lifestyle choices or socioeconomic status.
Mendelian Randomization (MR) solves this by using genetic variants as "instrumental variables" (IVs). Because our genotypes are randomly assigned at conception (a "natural experiment"), they aren't affected by the confounders that plague observational data. While MR is well-established for studying the direct effect of an exposure on an outcome, using it to estimate the interaction between genes and environment is much less understood. This paper investigates whether our standard MR "toolbelt" is up to the task of measuring these complex interplays.
Methods
The authors focus on two workhorse approaches in the IV world:
- 2-Stage Predictor Substitution (2SPS): You first predict the environmental exposure using genetic instruments, then plug those predicted values into your final disease model.
- 2-Stage Residual Inclusion (2SRI): Instead of substituting the exposure, you keep the original (confounded) exposure in the model but add the "residuals" (the error) from the first-stage prediction as a separate covariate to "soak up" the confounding.
The researchers extended these methods to include interaction terms ($G \times E$) and tested them across two settings: Linear Regression (for continuous outcomes like blood pressure) and Logistic Regression (for binary outcomes like "has cancer" vs. "does not"). They used both rigorous analytical derivations and extensive computer simulations to see which method stayed closest to the truth.
Datasets
- Simulated Data: The authors ran thousands of simulations (500–1000 repetitions per setting) varying the strength of confounding, the instrument's power, and the type of interaction.
- Real-World Application: They applied their methods to a massive dataset investigating the interaction between BMI and genetic variants on Colorectal Cancer (CRC) risk. This pooled analysis included 44,500 cases and 52,235 controls from major consortia, including the UK Biobank and the Genetics and Epidemiology of Colorectal Cancer Consortium (GECCO).
Key Insights and Ideas
- The Linear "Easy Mode": In linear models, both 2SPS and 2SRI are "consistent"—meaning with enough data, they will give you the exactly correct answer. 2SRI, however, emerged as the winner for efficiency (smaller error bars).
- The Logistic "Hard Mode": For binary outcomes, the math breaks down. Due to a property called non-collapsibility, the "odds ratio" measured in a model with a confounder isn't the same as the one measured without it, even if the confounder is perfectly accounted for.
- 2SRI > 2SPS: Even in the messy logistic setting, 2SRI generally outperformed 2SPS. While 2SPS showed notable bias even in simple scenarios, 2SRI was often "less wrong" and maintained better control over Type-I error (false positives).
- The "Naive" Trap: The paper shows that "naive" estimators (which ignore confounding) are surprisingly okay at testing if an interaction exists in linear settings, but they fail miserably at estimating the size of the main effects, making them dangerous for clinical interpretation.
Key Results
In the Colorectal Cancer study, the researchers found:
- No GxBMI Interaction: Across hundreds of thousands of genetic markers, there was no widespread evidence that BMI interacts with specific genes to change CRC risk.
- The Causal Gap: Interestingly, the MR methods estimated the causal effect of BMI on CRC risk to be significantly higher than traditional observational studies. The MR log-odds ratio was approximately 0.23 per 5-unit increase in BMI, compared to only 0.14 for the naive observational estimate.
- Why the difference? This suggests that traditional studies might be underestimating the risk of BMI due to measurement errors or "reverse causation" (where early undiagnosed disease might cause weight loss).
Limitations and Future Work
The authors acknowledge that their study used a relatively simple model for unmeasured confounding. In the real world, confounders might interact with genes or the environment in ways not captured here.
Furthermore, the "non-collapsibility" issue in logistic regression remains an "unresolved challenge." The authors conclude that while 2SRI is currently the best practical choice, the field urgently needs new, more robust statistical frameworks specifically designed for non-linear MR to handle binary disease outcomes without the inherent biases of the logit link.
Summary updated