Gene-environment interaction: why nutrigenomic effect sizes keep coming out small

What an interaction claim actually asserts

A gene-diet interaction claims that the effect of a dietary factor differs depending on which genetic variant a person carries. This is a stronger claim than saying either factor matters on its own, because it concerns the difference between two differences. Estimating a difference between differences uses the information in the data less efficiently, which is the root of the statistical difficulty.

As a rough guide, detecting an interaction reliably requires a very substantially larger sample than detecting a comparable main effect. Studies powered adequately for main effects are therefore routinely underpowered for the interaction analyses they go on to report as secondary findings. None of this is a criticism of the researchers involved, since the sample sizes required are frequently beyond what any single cohort can supply.

Why underpowered studies produce dramatic findings

When a study is underpowered, the effects that happen to reach significance are the ones that were overestimated by chance. This means the published effect from a small study is systematically larger than the true effect it was trying to measure. Replication in a larger sample then produces a smaller estimate, which reads publicly as a reversal rather than as expected behaviour.

The pattern is not specific to nutrigenomics and appears wherever small studies search across many possible comparisons. Recognising it turns an apparently chaotic literature into a predictable one, which is more useful than treating each reversal as news.

Diet is measured badly, and that matters

Dietary intake in large cohorts is usually assessed by questionnaire, which relies on recall and on estimating quantities people did not weigh. The resulting measurement error is substantial and it attenuates any association, pushing estimated effects toward zero. Attenuation affects interaction estimates more severely than main effects, because the error enters the analysis at multiple points.

Better dietary assessment through biomarkers exists for some nutrients and is expensive to apply at the scale interactions require. The field is therefore caught between needing very large samples and needing measurement quality that large samples cannot afford.

The multiple comparisons problem

A dataset containing many genetic variants and many dietary factors permits an enormous number of possible interaction tests. If those tests are run without correction, findings reaching conventional significance thresholds will appear by chance alone in quantity. Analyses declared in advance address this, and analyses conducted after inspecting the data are much harder to evaluate.

Whether a specific comparison was planned or selected afterwards is frequently not stated clearly in published reports. This is probably the single most useful question to ask about any surprising interaction result reported anywhere in this literature. Pre-registration has improved matters in the parts of the field that have adopted it, and adoption remains partial rather than universal.

What survives and what it supports

Some interactions have replicated consistently, generally involving variants with large effects on a well-understood metabolic pathway. Those cases tend to concern specific conditions rather than general dietary optimisation for people without any diagnosis. The broader promise of tailoring an everyday diet to a genotype has not been supported by the trials conducted so far.

That may change as datasets grow, and the reasonable position is to describe the current state rather than the anticipated one. Dietary changes made for a diagnosed condition belong under the guidance of a clinician or dietitian rather than a genetic report.