Summary
A medRxiv preprint reports that trial eligibility criteria can create apparent differences in lipid-lowering treatment response by polygenic risk score, even when the true genetic interaction is set to zero. Its cohort analyses found little evidence that coronary artery disease risk scores modified statin effects.
A medRxiv preprint reports that clinical-trial eligibility criteria can make a polygenic risk score appear to predict who benefits more from treatment, even when the underlying genetic difference in response is set to zero. The authors tested this possibility using simulations and analyses of two population cohorts, focusing on statins and other lipid-lowering treatments.
The finding matters because recent reanalyses of major lipid-lowering trials have reported greater treatment benefit among people with higher polygenic risk scores (PRSs) for coronary artery disease (CAD). A PRS combines information from many genetic variants to estimate a person’s inherited predisposition to a trait. It is a measure of risk, not a direct measurement of how a drug will work.
How eligibility can distort a subgroup result
Collider bias is a statistical distortion that can arise when an analysis selects people based on a factor influenced by multiple characteristics. Conditioning on that selection can create an association between characteristics that were not associated in the wider population. In a clinical trial, eligibility rules determine who can enter the study before treatment is randomly assigned. The preprint argues that this selection can make treatment effects appear to differ by PRS.
The distinction is important: randomisation helps balance treatment assignment among people in a trial, but it does not undo biases that may arise from selecting the trial population in the first place.
What the simulations and cohorts found
In simulations where the true PRS-by-treatment interaction was set to zero, applying trial eligibility criteria produced apparent PRS-dependent treatment effects resembling those reported in reanalyses of three trials. One simulated result had a FOURIER-like hazard ratio for lipid-lowering treatment versus placebo of 0.77 in the high-CAD-PRS group and 0.95 in the low-PRS group. These are simulation results, not estimates newly measured in the FOURIER trial.
The authors also analysed the Helix Research Network cohort (175,106 participants) and UK Biobank (92,859 participants) without applying the trial eligibility criteria. They found little evidence that CAD PRS modified the effect of statin use on acute coronary events.
As a better-powered test of the proposed bias, the researchers examined LDL cholesterol response. They found no evidence that an LDL PRS modified statin response in the population cohorts. But stratifying participants by baseline LDL cholesterol produced apparent PRS-dependent response patterns that matched the simulations.
The results show how an eligibility-based selection process can generate an apparent genetic treatment interaction. They do not establish that every PRS-associated difference reported in a particular trial is an artefact; the extent to which this mechanism explains individual trial results remains unresolved. The paper is a preprint and has not yet undergone peer review.
The authors disclosed industry ties, including J Brent Richards’s leadership and equity in 5 Prime Sciences, where Matthew Tudball is employed and holds shares, and employment affiliations for Michael Holmes at Genentech and Elizabeth Cirulli at Helix.