Methodological Considerations in Sibling Analyses of Prenatal Acetaminophen

This study demonstrates that divergent results in bi-directional sibling analyses of prenatal acetaminophen are often predictable artifacts of unadjustable parity confounding rather than evidence of carryover effects, suggesting that total sibling models adjusted for parity remain robust.

The Gist

Imagine you are trying to figure out if eating a specific type of candy causes a child to become hyperactive. To get a truly fair answer, you decide to compare two siblings from the same family. One sibling ate the candy before school, and the other didn't. Since they share the same parents, genes, and home environment, any difference in their behavior should be due to the candy, right?

This is the basic idea behind sibling studies, a popular method scientists use to find the truth about health risks.

However, a recent study by Ahlqvist and colleagues found a hidden trap in how some scientists analyze these sibling comparisons. They discovered that a common "safety check" scientists use can actually create fake results that look scary but aren't real.

Here is the story of their discovery, explained simply:

The "Safety Check" That Backfired

Scientists often worry that the order in which children are born matters. Maybe the first child gets more attention, or maybe the parents are more tired with the second child. These are called "parity" or birth order effects.

To make sure birth order wasn't messing up their results, some researchers do a "safety check" called a bi-directional analysis. They split the data into two separate groups:

1. Group A: Families where the older sibling was exposed (ate the candy) and the younger wasn't.
2. Group B: Families where the younger sibling was exposed and the older wasn't.

If the results are the same in both groups, the scientists feel confident. But if the results are totally different (e.g., Group A shows a huge risk, but Group B shows no risk), they often panic and say, "Our study is flawed! There must be some hidden bias we didn't catch!"

The "Perfect Match" Problem

The authors of this paper realized that this safety check has a fatal flaw. It's like trying to weigh a person on a scale that is already broken because the person is standing on a box that is the exact same size as the scale's platform.

In these split groups:

• If you are in Group A, you are always the older sibling.
• If you are in Group B, you are always the younger sibling.

Because "being older" and "being exposed" are now perfectly locked together, you cannot mathematically separate them. You can't adjust for birth order because birth order is the exposure in that specific group.

The Candy Experiment (The Simulation)

To prove this, the authors ran a computer simulation. They created a fake world with 500,000 families.

• The Truth: In this fake world, the candy (acetaminophen) had zero effect on behavior. It was completely harmless.
• The Twist: They made the older siblings slightly more likely to get a diagnosis (like ADHD) just because they were older, and slightly more likely to eat the candy just because they were older.

When they analyzed the whole group together and adjusted for birth order, the result was perfect: No effect. (Just like the truth).

But when they did the "safety check" (splitting into older-exposed vs. younger-exposed), the results went crazy:

• Older exposed: The computer said the candy increased risk by 68%.
• Younger exposed: The computer said the candy decreased risk by 40%.

The Big Takeaway

The authors found that this "divergence" (where the two groups give opposite answers) wasn't because the study was bad or because there was a hidden bias. It was just a mathematical trick caused by the way the groups were split.

Think of it like this:
Imagine you are testing if running shoes make you faster.

• You compare your older brother (who runs fast naturally) wearing the shoes, to your younger brother (who runs slow naturally) who doesn't.
• Then you compare your younger brother (who runs slow) wearing the shoes, to your older brother (who runs fast) who doesn't.

If you look at these two comparisons separately, the results will look totally different because of the natural speed difference between the brothers. But if you look at the whole picture and account for their natural speed, you see the shoes do nothing.

Why This Matters

This paper is a wake-up call for researchers studying prenatal acetaminophen (a common painkiller) and neurodevelopmental issues like autism or ADHD.

A famous study by Lee et al. previously found that acetaminophen had no effect when looking at siblings, but they got scared because their "safety check" showed weird, opposite results. They thought their main study might be wrong.

This new paper says: "Don't panic."
The weird results in the safety check were likely just a mathematical illusion caused by birth order. The main study (which adjusted for birth order correctly) is probably right: Acetaminophen likely has no causal link to these conditions.

The lesson? Just because a safety check gives weird, split results, it doesn't mean the main finding is broken. Sometimes, the safety check itself is the one that's broken.

Technical Summary

1. Problem Statement

Sibling-matched designs are widely used in epidemiology to control for shared familial confounders (e.g., genetics, parental socioeconomic status). However, these designs remain vulnerable to non-shared confounders. A common sensitivity analysis used to validate sibling designs is the bi-directional analysis, which stratifies families based on whether the older or younger sibling was exposed.

The authors address a specific controversy arising from a recent study by Lee et al. regarding prenatal acetaminophen use and neurodevelopmental outcomes (ASD/ADHD). While Lee et al. found a null association in their main sibling analysis, their bi-directional sensitivity analysis showed divergent results: the risk appeared elevated when the older sibling was exposed (HR ~1.33–1.75) but protective when the younger sibling was exposed (HR ~0.74–0.75). Lee et al. interpreted this divergence as evidence of unaddressed bias (potentially "carryover effects") that invalidated their main null finding.

The Core Question: Is this divergence evidence of a flawed main analysis, or is it a predictable statistical artifact inherent to the bi-directional sensitivity analysis itself?

2. Methodology

The authors employed a simulation study to test whether parity (birth order) alone could generate the divergent signals observed in Lee et al., assuming the true causal effect of acetaminophen was strictly null.

• Simulation Design:
  • Cohort: 500 simulated cohorts of 525,000 two-sibling families.
  • True Causal Effect: Set to null (Odds Ratio = 1.00) for all individuals.
  • Confounding Structure: A latent familial factor simulated shared confounding. Crucially, parity-related confounding was introduced by varying baseline probabilities by birth order:
    • Exposure: 48% for the firstborn, 50% for the second-born.
    • Outcomes: ASD probability (1.25% vs. 0.75%) and ADHD probability (6% vs. 4.5%) were higher for the firstborn, reflecting real-world birth-order effects.
  • Analysis Models: The authors restricted the analysis to exposure-discordant families and applied conditional logistic regression under four specifications:
    1. Total sibling sample (unadjusted for parity).
    2. Total sibling sample (adjusted for parity).
    3. Bi-directional subset: Older sibling exposed (younger unexposed).
    4. Bi-directional subset: Younger sibling exposed (older unexposed).

3. Key Contributions

• Identification of Perfect Collinearity: The paper demonstrates that in bi-directional stratified models, parity is perfectly collinear with exposure. In the subset where the "older sibling is exposed," the exposure variable is identical to the "firstborn" variable. Consequently, parity cannot be adjusted for in these specific models, unlike in the main sibling analysis where it can be included as a covariate.
• Reinterpretation of Divergence: The authors argue that divergent estimates in bi-directional analyses do not necessarily indicate invalid unmeasured bias in the main study. Instead, they are a predictable artifact of parity confounding that is unavoidable in the stratified sensitivity analysis.
• Distinction between Birth Year and Parity: The paper clarifies that adjusting for "birth year" (to control for secular trends) does not resolve parity confounding, as birth year and birth order are distinct non-shared confounders.

4. Results

The simulation results mirrored the real-world findings of Lee et al., despite the true causal effect being zero:

• Total Sibling Models:
  • When adjusted for parity, the estimated Odds Ratio (OR) was 1.00 (95% CI: 0.95–1.05), correctly recovering the null effect.
  • Even without parity adjustment, the bias was minimal (OR = 0.98).
• Bi-Directional Models (Stratified):
  • Because parity could not be adjusted for, the models exhibited severe artifactual divergence:
    • Older sibling exposed: ASD OR = 1.68 (95% CI: 1.56–1.81).
    • Younger sibling exposed: ASD OR = 0.60 (95% CI: 0.56–0.64).
  • Similar divergence was observed for ADHD (ORs of 1.36 vs. 0.74).
• Comparison to Lee et al.: The simulated ORs (1.68 and 0.60) closely matched the hazard ratios reported by Lee et al. (1.75 and 0.74), confirming that the observed divergence in the real data could be entirely explained by parity confounding.

5. Significance and Conclusion

• Methodological Robustness: The study suggests that overall sibling models adjusting for parity are more robust than the bi-directional sensitivity analyses. The divergence in stratified models is a mathematical inevitability when birth order effects exist, not necessarily proof of bias in the primary null finding.
• Implications for Interpretation: Researchers should not automatically dismiss a sibling study's main null result simply because bi-directional analyses show divergent signals. Such divergence may reflect the inability to control for birth order within the stratified subsets rather than a failure of the sibling design itself.
• Future Directions: The paper calls for a re-evaluation of how sensitivity analyses are interpreted in sibling studies, emphasizing that the "validity" of a sibling design should not be judged solely by the consistency of bi-directional stratified results when parity effects are present.

In summary, Ahlqvist et al. provide a critical methodological correction, demonstrating that the "flip" in risk estimates observed in previous acetaminophen studies is likely a statistical artifact of parity confounding, thereby reinforcing the validity of the original null findings regarding prenatal acetaminophen and neurodevelopmental outcomes.