Triangulation of evidence to examine selection bias in lifecourse Mendelian randomization studies: an example using early life adiposity and breast cancer.

This study employs a triangulation framework combining empirical analyses and simulations to demonstrate that plausible selection bias mechanisms are unlikely to fully explain the observed inverse causal effect of early-life adiposity on breast cancer risk, thereby supporting the validity of the protective association.

The Gist

The Big Question: Is the "Fat Kid" Clue Real or a Trick?

Imagine scientists have been studying a strange pattern for years: Children who were heavier as kids seem to have a lower risk of getting breast cancer when they grow up.

This sounds counterintuitive. Usually, we think being overweight is bad for health. But the data keeps showing this "protective" effect. To prove it wasn't just a fluke or caused by bad habits (confounding), researchers used a clever tool called Mendelian Randomization (MR). Think of MR as using a person's genes as a "fate coin" to see if childhood weight causes the lower cancer risk later in life.

The Skeptics' Argument:
Some critics said, "Wait a minute! This isn't a real cause-and-effect. It's a trick of the light caused by Selection Bias."

They argued that the studies were like a party where only certain people showed up. If the people who survived to be recruited into the study were a specific type of person, the data could look like a protective effect even if none existed. It's like looking at a photo of a marathon and only counting the runners who finished; if you ignore the ones who dropped out, you might think the race was easier than it really was.

The Paper's Mission: The "Triangulation" Detective Work

The authors of this paper decided to play detective. Instead of just arguing, they used a method called Triangulation. Imagine trying to find a hidden treasure. If you only look from one spot, you might be wrong. But if you look from three different angles (using three different methods), and they all point to the same spot, you can be sure the treasure is there.

They used three different "flashlights" to see if the "Selection Bias" monster was actually hiding the truth.

Flashlight 1: The Family Photo Album (Proxy-Genotype Analysis)

The Theory: Critics said that when we look at a person's own data, the "protective" effect looks strong, but when we look at their siblings or mothers, the effect gets weaker. They claimed this meant the main study was biased.
The Test: The researchers ran a computer simulation (a "virtual family") where they knew for a fact there was no bias. They simply used the rules of genetics (how genes are passed down) to see what would happen.
The Result: Even without any bias, the "protective" effect looked weaker in the virtual siblings and mothers, just like in the real data.
The Analogy: It's like looking at a photo of yourself in a mirror versus a photo of your brother. The mirror image looks slightly different just because of how the reflection works, not because the photo is fake. The "weakening" effect was just a natural result of how genes are shared, not a sign of a trick.

Flashlight 2: The Survival Test (Parental Longevity)

The Theory: Critics suggested that maybe heavy children died young before they could be recruited into the study, leaving only the "healthy" heavy kids behind. This would make it look like being heavy is good for survival.
The Test: The researchers looked at the genes of children to see if they predicted how long their parents lived. If being a "heavy kid" meant you were more likely to die young, we should see that in the parents' lifespans.
The Result: They found that being a heavy adult definitely shortened life. But being a heavy child did not seem to shorten life once you accounted for how heavy the person became as an adult.
The Analogy: Imagine a forest where trees that grow too fast as saplings (children) don't necessarily die young. It's only when they get huge as adults that they might struggle. Since the "heavy kid" factor didn't kill the parents, it's unlikely that "heavy kids" were dying off before the study started.

Flashlight 3: The Stress-Test Simulation (The "What If" Game)

The Theory: What if the selection bias was so extreme and weird that it could create this fake protective effect?
The Test: The researchers built a massive computer simulation. They created a world where childhood weight had zero effect on breast cancer (a "Null Model"). Then, they introduced the most extreme, crazy selection biases they could imagine—like a study where only the thinnest children who also happened to get cancer were allowed to join.
The Result: Even with these "super-villain" levels of bias, the computer could not recreate the strong protective effect seen in the real studies. The bias was too weak to pull off the magic trick.
The Analogy: Imagine trying to make a fake gold coin look real by painting it. The researchers tried every shade of paint, every thickness of gold leaf, and every lighting trick. No matter how hard they tried, the fake coin still looked like a fake coin. It couldn't fool the experts.

The Verdict: It's Real!

After shining all three flashlights, the authors concluded:

1. The "weakening" effect in family studies is just normal genetics, not a bias.
2. Heavy kids aren't dying off before the study starts, so survival bias isn't the culprit.
3. Even the most extreme, unrealistic selection biases can't explain away the results.

The Bottom Line:
The protective effect of being a heavier child against breast cancer later in life is likely real, not a statistical illusion.

Why does this happen?
The paper suggests that being heavier as a child might change how the body develops (like lowering breast density or changing hormone levels), which actually protects the breast tissue from cancer later in life.

Summary for the Everyday Reader

Think of this study as a court case. The prosecution (the critics) said, "The evidence is a trick caused by who showed up to the party!" The defense (this paper) brought in three expert witnesses (Family Genetics, Survival Data, and Super-Computers). All three testified that the "trick" theory doesn't hold up. The evidence stands: Being heavier as a child really does seem to lower the risk of breast cancer as an adult.

Technical Summary

1. Problem Statement

There is a well-established inverse association between higher adiposity (body fat) in early life and the risk of breast cancer in later life. Mendelian Randomization (MR) studies have supported a causal interpretation of this protective effect. However, critics have argued that these findings may be spurious artifacts of selection bias, specifically collider stratification bias.

The core concern is that in large cohort studies (like the UK Biobank), participation or survival to recruitment is influenced by both the exposure (childhood body size) and the outcome (breast cancer risk). Conditioning on this selection mechanism can induce a non-causal inverse association even if the true causal effect is null. The authors aim to determine whether plausible selection mechanisms are sufficient to explain the observed protective effect of early-life adiposity on breast cancer.

2. Methodology: A Triangulation Framework

The authors employed a "triangulation" approach, combining empirical analyses, family-based simulations, and extensive simulation studies under a null causal model to evaluate the selection bias hypothesis.

A. Re-evaluation of Proxy-Genotype MR

• Context: Previous critiques noted that MR estimates using offspring genotypes as proxies for parental genotypes (proxy-MR) showed weaker associations than self-MR estimates. Critics interpreted this attenuation as evidence of selection bias.
• Approach: The authors re-analyzed summary data for self-MR, sibling-MR, and mother-MR. They then conducted family-based simulations assuming a true causal effect but no selection bias.
• Goal: To determine if the observed attenuation in proxy-MR estimates could be explained solely by Mendelian inheritance dilution and measurement error, without invoking selection bias.

B. Multivariable MR of Parental Survival

• Context: Selection bias could arise if childhood adiposity influences mortality prior to recruitment (survival bias).
• Approach: The authors performed multivariable MR (MVMR) using offspring genotypes as proxies for parental genotypes to assess the effect of childhood vs. adulthood body size on parental survival.
• Goal: To test if survival differences are driven by childhood adiposity (which would support the selection bias hypothesis) or adulthood adiposity.

C. Simulation Studies under a Sharp Null Model

• Design: The authors simulated data where the true causal effect of both childhood and adulthood body size on breast cancer was zero (sharp null).
• Mechanisms Tested:
  • Additive Selection: Selection based on the sum of exposure and outcome.
  • Interaction-Driven Selection: Selection based on the interaction between exposure and outcome.
  • Confounding: Inclusion of unmeasured confounders affecting body size and cancer risk.
• Scenarios: 27 scenarios were tested, varying selection strength from null to extreme levels (exceeding empirical observations in UK Biobank) to perform a "stress test" of the selection hypothesis.
• Analysis: They compared the simulated MR estimates (both univariable and multivariable) against the empirical estimates observed in real data.

3. Key Contributions

• Methodological Rigor: The paper moves beyond theoretical arguments about selection bias to provide quantitative, simulation-based evidence regarding the magnitude of bias required to explain specific empirical findings.
• Decoupling Life-Stage Effects: By using MVMR on parental survival, the study distinguishes between selection mechanisms driven by childhood vs. adulthood adiposity.
• Triangulation: The study integrates three distinct lines of evidence (proxy-MR attenuation, survival MVMR, and null-model simulations) to provide a robust conclusion that selection alone is insufficient to explain the data.

4. Key Results

A. Proxy-Genotype Attenuation

• Finding: The simulations reproduced the pattern of attenuated estimates in proxy-MR (sibling/mother) compared to self-MR without introducing any selection bias.
• Implication: The attenuation observed in previous studies is consistent with standard Mendelian inheritance dilution and measurement error, not necessarily selection bias.

B. Parental Survival Analysis

• Finding: In MVMR analyses of parental survival, higher adulthood body size exerted a robust adverse effect on survival. However, the effect of childhood body size on survival fully attenuated after adjusting for adulthood body size.
• Implication: Survival-related selection is driven primarily by adulthood adiposity, not childhood adiposity. Therefore, survival bias is unlikely to generate a spurious protective effect specifically for childhood body size.

C. Simulation Results (Null Model)

• Additive Selection: Produced minimal bias in childhood MR estimates.
• Interaction-Driven Selection: Generated increasing negative (protective) bias as interaction strength increased. However, even under extreme scenarios (parameters exceeding empirical observations), the simulated bias was insufficient to replicate the magnitude of the observed protective effect (logOR = -0.53).
• Selection via Adulthood: When selection was modeled to operate through adulthood body size, the bias manifested primarily in adulthood estimates, leaving childhood estimates near the null.
• Joint Pattern Failure: Across 781,250 simulation scenarios, none successfully reproduced the joint pattern of empirical results (specifically, the combination of univariable and multivariable estimates for both childhood and adulthood) under a null causal model with selection bias alone.

5. Significance and Conclusion

The study concludes that while selection bias is a valid concern in epidemiology, plausible selection mechanisms are unlikely to fully explain the observed inverse causal effect of early-life adiposity on breast cancer risk.

• Support for Causality: The findings support a causal interpretation where higher early-life adiposity is genuinely protective against breast cancer, potentially mediated by hormonal changes or reduced mammographic density.
• Methodological Guidance: The paper highlights the value of triangulating evidence. Relying on a single line of evidence (e.g., attenuation in proxy-MR) can be misleading; combining family-based designs, survival analyses, and stress-test simulations provides a more robust assessment of bias.
• Future Directions: The authors suggest that future research should focus on clarifying the biological pathways (e.g., sex hormones, breast tissue development) through which early-life adiposity exerts this protective effect, rather than dismissing the association as an artifact of selection.