New Family-Trio Genetic Method Exposes Hidden Bias in Causal Gene Studies

One of the most powerful tools in modern genetics may be quietly misleading researchers, and a new statistical framework claims to know exactly why. In a study published in BMC Bioinformatics, a team of researchers led by Shun Zhang and Jia-Hao Mai of Southern Medical University introduced FT-SEM, a structural equation modeling framework designed to run genome-wide association studies, or GWAS, on family trio data while simultaneously exploiting parental phenotypes that earlier methods simply threw away. The framework’s central promise is ambitious: to strip out a subtle form of inherited confounding known as dynastic effects, which can contaminate both genetic association estimates and the causal inferences built on top of them. When the authors applied their method to real data from the UK Biobank, a striking pattern emerged. A widely reported apparent causal link between systolic blood pressure and obesity largely dissolved once familial confounding was accounted for, while the effect of high-density lipoprotein cholesterol on obesity remained directionally consistent, suggesting a genuine biological relationship.

To understand why this matters, it helps to revisit how Mendelian randomization works. The technique uses genetic variants, typically single nucleotide polymorphisms, as instrumental variables to estimate whether an exposure truly causes an outcome. Its validity rests on three assumptions: the instrument must be strongly associated with the exposure, it must be independent of confounders, and it must influence the outcome only through the exposure. In practice, the second and third assumptions are fragile. Horizontal pleiotropy, in which a variant affects the outcome through pathways unrelated to the exposure, violates the exclusion restriction and has spawned an entire industry of robust correction methods. But dynastic effects present a different and arguably more insidious problem, one that cannot be fixed by looking harder at population-level summary statistics alone.

Dynastic effects arise specifically within families. In a trio consisting of two parents and an offspring, parental genotypes can influence offspring traits through unobserved parental phenotypes or other intergenerational mechanisms, such as the family environment parents create or lifestyle behaviors passed down across generations. These pathways create a persistent confounding structure in the association between a SNP and an outcome, violating the independence assumption that Mendelian randomization depends on. Crucially, the bias does not average out with large sample sizes; it simply persists. The only reliable way to block these non-causal pathways is to bring family-level data into the analysis, which is precisely what the authors set out to do in a more comprehensive way than previous designs allowed.

Existing family-based approaches each come with trade-offs. A standard trio-based method fits a linear model regressing the offspring’s phenotype on both offspring and parental genotypes, which effectively controls for dynastic bias but discards parental phenotypic information entirely, limiting statistical power. Mother-offspring pair designs use structural equation models to separate maternal and fetal effects, but they ignore paternal contributions and discard offspring genotype data, confining their use largely to birth-weight research. Sibling-pair designs, increasingly popular as biobanks accumulate sibling data, exploit differences between siblings to cancel out dynastic effects, but they cannot be applied to trios with an only child. Perhaps most importantly, none of these frameworks had been extended to multivariate analysis, leaving correlated traits to be examined one at a time.

FT-SEM fills that gap by modeling multiple correlated phenotypes through a shared latent factor. In the framework, the genetic effect at each SNP is partitioned into paternal, maternal, and offspring components, and the correlations among several traits are assumed to be driven by a common unobserved construct. For the obesity application, that latent factor was built from body mass index, body fat percentage, and waist-to-hip ratio, three moderately correlated measures that each capture a different dimension of adiposity. By estimating SNP effects on the shared factor rather than on each trait separately, the method reduces phenotype-specific measurement error and biological heterogeneity. Latent grandparental genotypes, which are never actually observed, are incorporated as latent variables with variances constrained to match observed genotypes under Hardy-Weinberg equilibrium, and Mendelian transmission paths are fixed at 0.5 to identify the model.

One technical obstacle the authors had to overcome is that the genetic effects of interest are not directly identifiable when the latent factor is estimated jointly with the SNP effect. Their solution is a two-stage procedure: first, a null model excluding all genetic variant components is fitted to obtain reference factor loadings; then the full model is fitted, and a correction factor computed from the ratio of loadings between the two fits is applied to recover accurate estimates of the offspring, paternal, and maternal effects. The offspring effect is the primary target for downstream interpretation and causal inference, because it represents the direct genetic effect on the offspring’s phenotype, while the parental effect estimates serve as diagnostic quantities that characterize and quantify dynastic effects themselves.

The simulation evidence is extensive. Across 10,000 replicates per setting, with sample sizes ranging from 1,000 to 3,000 trios and dynastic effect magnitudes of zero, 0.2 percent, and 0.4 percent of phenotypic variance, FT-SEM produced unbiased effect estimates with well-controlled type I error rates and confidence interval coverage near the nominal 95 percent level. A population-based comparator, IndSEM, was more efficient when no dynastic effects were present, achieving the lowest root mean squared error and the narrowest intervals, but its performance collapsed as dynastic effects increased, producing biased estimates and inflated false positive rates. FT-SEM consistently outperformed PSEM, a multivariate extension of the conventional trio-based linear model, by exhibiting lower root mean squared error and higher statistical power, a direct benefit of incorporating parental phenotypic data. Additional simulations showed the framework remained robust even when the assumed latent-factor structure or cross-generational measurement invariance was deliberately misspecified.

The real-data application turned on 778 complete family trios from the UK Biobank, identified through stringent kinship-based filtering and restricted to individuals of European ancestry to minimize population stratification. After quality control on genotypes and phenotypes, the authors ran genome-wide scans under FT-SEM, PSEM, and IndSEM. No locus reached genome-wide significance, which is unsurprising given the modest trio sample size, but that was not the point. The family-corrected summary statistics for the latent obesity factor served as an unbiased outcome base for downstream two-sample Mendelian randomization, paired with sibling-based exposure data covering 99,998 individuals for the within-family analyses and independent-individual data of 315,133 and 757,601 participants for the comparison arm. Because complete trios are typically excluded from conventional population-based GWAS, the two-sample design involved no sample overlap between exposure and outcome datasets.

The results were revealing. For high-density lipoprotein cholesterol, the population-based IndSEM framework detected significant causal associations under weighted median, inverse-variance weighting, and MRcML methods, with p values around 0.041 to 0.042, and the family-based methods, though underpowered, pointed in the same direction. For systolic blood pressure, the picture was different: IndSEM flagged a significant association under MR-Egger regression, but the family-based FT-SEM and PSEM estimates were substantially attenuated toward zero and nowhere near significance. A sensitivity analysis using a harmonized instrument set drawn from sibling-based GWAS found no significant blood pressure association either, though the authors caution that weaker sibling-based instruments and reduced statistical power likely contributed to that attenuation. Taken together, the pattern suggests the apparent blood pressure to obesity link reported in conventional studies may be spurious, driven largely by familial confounding rather than genuine causation.

The authors are candid about limitations. The trio sample is small and underpowered for discovery, the framework assumes multivariate normality, additive genetic effects, and a constant genetic architecture across generations, and it is currently restricted to trios with a single offspring and does not model gene-environment or gene-gene interactions. The structural equation model is also computationally heavier than standard linear models, since it must be re-estimated for every SNP, though the authors argue the burden is manageable on modern parallel computing platforms. Still, the broader message is hard to ignore: as Mendelian randomization grows in popularity, susceptibility to dynastic effects has become a critical concern, and discrepancies between population-based and family-based estimates should generally be resolved in favor of the family-based result. FT-SEM offers researchers a way to have both rigor and efficiency, blocking non-causal intergenerational pathways while squeezing maximal information from every member of the family trio, and providing a template for future pedigree-wide analyses that could dissect direct genetic effects, assortative mating, and other intergenerational mechanisms across entire family trees.

Subject of Research: A multivariate GWAS framework using family trio data and parental phenotypes to control dynastic effects in genetic association and Mendelian randomization analyses

Article Title: Multivariate GWAS framework for family trios with parental phenotypes to control dynastic effects

Article References: Zhang, S., Mai, J.-H., Zhong, Q., Zhu, Q.-W., Li, Y.-S., Wu, X.-B., & Zhou, J.-Y. (2026). Multivariate GWAS framework for family trios with parental phenotypes to control dynastic effects. BMC Bioinformatics, 27(1), Article 222. https://doi.org/10.1186/s12859-026-06597-8

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06597-8

Keywords: GWAS, Mendelian randomization, family trios, dynastic effects, structural equation modeling, UK Biobank, obesity, direct genetic effects, parental phenotypes, latent factor, pleiotropy, BMC Bioinformatics