The team set out to quantify, systematically and at genome-wide scale, how well the two dominant study designs converge. Cross-sectional designs compare genetic associations between individuals of different ages at a single time point, asking whether a variant’s effect differs between a 45-year-old and a 65-year-old sampled simultaneously. Longitudinal designs instead track the same individuals over time, modeling within-person change directly. In principle, if aging is the only force at work, both approaches should estimate the same quantity. In practice, they rest on strong assumptions: that genetic effects do not vary by birth cohort, that sampling is comparable, and that participation in the study does not itself depend on age and phenotype.

Using the UK Biobank, the researchers analyzed 31 non-binary health-related phenotypes spanning cognition, physiology, clinical indicators, anthropometrics and lifestyle behaviors. The cross-sectional arm drew on up to 498,845 participants aged 40 to 69 at baseline; the longitudinal arm followed up to 99,459 of them across repeat assessments conducted between 2012 and 2024, with a mean follow-up of roughly 11 years. At the phenotypic level, cross-sectional age effects explained about 70.6 percent of the variance in longitudinal estimates—substantial, but far from the near-perfect agreement that equivalence would demand. The largest discrepancies appeared in behavioral traits. For the number of medications taken, the cross-sectional model estimated an age effect three times larger than the longitudinal one. For smoking and alcohol use, the two designs even disagreed about the sign: traits that decline with age within individuals falsely appeared to increase with age when compared across individuals.

The culprit, the authors show, is confounding by birth cohort. In cross-sectional data, age and year of birth are almost perfectly inversely correlated—age equals period minus cohort—so differences between generations can masquerade as aging. Because date of birth and cross-sectional age correlated at roughly −0.99 in the UK Biobank, negative cohort effects systematically inflate cross-sectional age estimates and can even reverse their direction. Significant cohort effects emerged for 28 of the 31 traits, and cohort confounding accounted for a striking 90.8 percent of the variance in the discrepancy between longitudinal and cross-sectional age estimates. The historical and social contexts in which people were born—educational reforms, economic upheavals, public health policies such as the 1965 UK ban on cigarette advertising—leave lasting marks on trait levels that a cross-sectional comparison misreads as aging.

The researchers then turned to the genetic level, implementing a two-step genome-wide strategy that combined scans for marginal genetic effects, cross-sectional gene-by-age interactions and genetic effects on longitudinal change. They identified 57 linkage-disequilibrium-independent variants with significant age-varying effects, most of them detected in cross-sectional analyses (44 variants, in up to 406,226 individuals) and fewer in longitudinal analyses (14 variants, in up to 83,579 individuals). The variants clustered around anthropometric traits such as weight, body fat-free mass and body mass index, metabolic indicators such as basal metabolic rate, and clinical outcomes including medication count and number of cancers. Self-reported health behaviors—physical activity, smoking, sleep—showed few or no significant age-varying genetic effects.

Encouragingly, direction was largely preserved across designs: 84.21 percent of the 57 variants showed consistent interaction directions whether tested cross-sectionally or longitudinally. Among the concordant variants, both attenuation and intensification of genetic effects with age were common, accounting for 50 percent and 35.42 percent of cases respectively. The pattern was trait-specific. Genetic effects on obesogenic traits tended to weaken with age, consistent with earlier reports of declining heritability for body mass index and depression in adulthood—plausibly because accumulated environmental exposures, medication, dietary change and shifting occupational circumstances gradually swamp genetic predisposition. Conversely, traits adversely affected by aging itself, such as cancer count, medication burden and reaction time, more often showed intensifying genetic influence over time. Crossover effects, in which the direction of a genetic association reverses across the lifespan, were rare, appearing in just seven variants with no detectable marginal effect.

Magnitude was another matter. Although cross-sectional and longitudinal estimates of age-varying genetic effects were linearly related, the regression slope was significantly below one, indicating that cross-sectional estimates were systematically larger. Decomposing the discrepancies revealed a familiar hierarchy: gene-by-birth-year interactions explained 70.8 percent of the variance in effect-size differences across variants, selective participation accounted for an additional 11.6 percent, and unmodeled nonlinear age trajectories contributed only 4.2 percent. For several variants—including rs2597355 on depression, rs56299829 on fruit intake and rs17362578 on walking pace—cohort confounding was strong enough to reverse the estimated direction of the age-varying genetic effect under a cross-sectional model. Suggestive gene-by-cohort effects were observed for 10 of the 57 variants, and the authors note that such effects are not mere noise: they capture how societal change reshapes genetic associations across generations, as documented in studies of genetic influences on social outcomes before and after the collapse of the Soviet Union and in response to educational policy reforms.

Participation bias, the second-largest contributor, operates differently in each design. Cross-sectional estimates are vulnerable to selective volunteering at baseline—the UK Biobank’s well-documented healthy volunteer effect—whereas longitudinal models adjust for time-invariant selection but assume that continued participation is independent of changes in phenotype. Because the pressures shaping initial recruitment differ from those governing retention, the two designs absorb different slices of selection bias. The scale of loss to follow-up is considerable: of the 399,661 baseline participants who did not take part in follow-up research, 42,001—about 11 percent—died before the first longitudinal assessment, though most attrition stemmed from other causes. Reweighting the samples to improve representativeness shifted cross-sectional estimates more than longitudinal ones, with sign reversals observed for four traits, and sample representativeness explained 2.8 percent of the discrepancy between designs—small next to the 90.8 percent attributable to cohort effects at the phenotypic level.

Nonlinear aging, by contrast, proved a minor player. Significant quadratic age effects were detected for 17 traits, meaning some aging trajectories genuinely bend rather than run straight, and such curvature can make estimated slopes depend on the age range sampled—the cross-sectional sample tops out at 69 years of baseline age while follow-up reaches 86. Yet deviations from linearity explained under 5 percent of the variance in design discrepancies, and the authors caution that their small-subset estimates are diluted by measurement error, with a dilution ratio of −0.43 for the nonlinear component. The findings replicated at the polygenic level: polygenic score effects on anthropometric and metabolic traits showed consistent directions but divergent magnitudes across designs, with cohort confounding again dominant at 85.2 percent of explained variance—a result with practical implications for anyone deploying polygenic risk scores across age groups or generations.

Where does this leave the field? The authors argue that neither design is immune to its own pathologies, and that robust inference demands integrating both. Cross-sectional analyses offer unmatched statistical power and age range, maximizing discovery of candidate age-varying variants; longitudinal analyses provide direct modeling of within-person change, control against time-invariant confounding and the ability to detect gene-by-cohort effects and nonlinear trajectories that cross-sectional data cannot in principle deliver. The team also illustrates how longitudinally derived genetic instruments could feed into Mendelian randomization, recovering the well-established adverse effect of body mass index on systolic blood pressure, albeit with wide uncertainty given the scarcity of strong instruments for change. As prospective biobanks, electronic health records and whole-genome sequencing expand, the framework will extend to rare variants and richer repeated measures. For now, the message is clear: before declaring that a gene’s influence grows or fades with age, researchers must ask whether they are measuring aging—or merely the echo of the century in which their participants were born.

Subject of Research: Age-varying genetic effects on complex traits and how study design, cohort confounding and participation bias shape their inference

Article Title: Design and model choices shape inference of age-varying genetic effects on complex traits

Article References: Schoeler, T., Wiegrebe, S., Winkler, T. W., & Kutalik, Z. (2026). Design and model choices shape inference of age-varying genetic effects on complex traits. Nature Aging. https://doi.org/10.1038/s43587-026-01232-w

Image Credits: AI Generated

DOI: 10.1038/s43587-026-01232-w

Keywords: genetic epidemiology, age-varying genetic effects, UK Biobank, birth cohort effects, participation bias, genome-wide association study, gene-by-age interaction, longitudinal analysis, cross-sectional design, polygenic scores, Mendelian randomization, complex traits