Proportional hazards models have previously been proposed as a method to analyse time-to-event phenotypes in genome-wide association studies. While proportional hazards models have many useful applications, their ability to identify genetic associations under different generative models where ascertainment is present in the analysed data is poorly understood. This includes commonly used study designs such as case-control and case-cohort designs (e.g. the iPSYCH study design) where cases are commonly ascertained. Here we examine how recently proposed and computationally efficient Cox regression for GWAS perform under different generative models with and without ascertainment. We also propose the age-dependent liability threshold model (ADuLT), first introduced as the underlying model for the LT-FH++ method, as an alternative approach for time-to-event GWAS. We then benchmark ADuLT with SPACox and standard case-control GWAS using simulated data with varying degrees of ascertainment. We find Cox regression GWAS to underperform when cases are strongly ascertained (cases are oversampled by a factor larger than 5), regardless of the generative model used. In contrast, we found ADuLT to be robust to case-control ascertainment, and be orders of magnitude more efficient computationally. We then used the methods to conduct GWAS for four psychiatric disorders, ADHD, Autism, Depression, and Schizophrenia in the iPSYCH case-cohort sample, which has a strong case-ascertainment. Summarising across all four mental disorders, ADuLT found 20 independent genome-wide significant associations, while case-control GWAS found 17 and SPACox found 8, consistent with our simulation results. As more genetic data are being linked to electronic health records, robust GWAS methods that can make use of age-of-onset information have the opportunity to increase power in analyses. We find that ADuLT to be a robust time-to-event GWAS method that performs on par with or better than Cox-regression GWAS, both in simulations and real data analyses of four psychiatric disorders. ADuLT has been implemented in an R package called LTFHPlus, and is available on GitHub.

Thyroid disorders are common. Information on long-term progression of morphologic disorders are scarce. The aim of this study was to describe the course of thyroid nodules and volume over a period of up to ten years. Data from the population-based Study of Health in Pomerania were used for longitudinal analysis of 10 years, on average. Billing data from the Association of Statutory Health Insurance Physicians were matched to the data to exclude participants with thyroid surgery, radioiodine therapy and thyroid carcinoma. Changes in number and size of thyroid nodules and thyroid volume were observed using ultrasound. A total of 1270 participants were included (53% female, median age at baseline 51 years). The proportion of subjects with at least one thyroid nodule increased from 34.9% to 47.5% after 10-years. The majority of participants had unchanged or reduced number of nodules. About one quarter had at least one nodule size ≥ 1cm. The proportion of participants with goitre increased from 35% to 37% after 10 years. Nevertheless, individual thyroid volume increased by < 1 ml (95% CI 0.38-3.66) after adjusting for age and BMI irrespective of thyroid medication. Thyroid nodules and goitre are common. After 10 years, the number of nodules did not increase in about 70%. This proportion did not differ substantially when excluding persons with thyroid medication. Thyroid volume increased slightly over the follow up period. These changes do not seem clinically relevant. Our results support a more restrictive approach regarding follow-up diagnostics in asymptomatic patients with thyroid nodules or minimally enlarged thyroid.

The state-of-the-field in complex disorder genetics is marked by large-scale, biobank-ascertained data sets that integrate broad demographic, health, survey, and genetic data. Leveraging the richness of this data can advance our understanding of complex disorders by improving predictions, describing etiology, and augmenting gene-mapping. These themes are especially, but not exclusively, relevant to Major Depressive Disorder (MDD), a leading cause of disability for which genetic predictors underperform, clinical heterogeneity is etiologically enigmatic, and missing heritability persists. The iPSYCH 2015 case-cohort presents a unique opportunity to integrate genotypes, register-based clinical outcomes, and extended genealogies of 30,949 MDD cases and 39,655 random population controls. To make full use of our data, we introduce the Pearson-Aitken framework for Family Genetic Risk Scores (PA-FGRS) to estimate individuals' liabilities for major depressive disorder from extended genealogies of partially observed relatives. PA-FGRS extends previous methods by accounting for censoring, leveraging distant relatives, and utilizing a flexible, model-based approach amenable to analysis and extension, advantages we highlight in simulations. Combining PA-FGRS with genotype data improves classification, replicates and extends known genetic contributions to clinical heterogeneity, and increases power for genome-wide association studies. This study combines novel analyses of unique data and can serve as a model for studies of other outcomes.

Genetics as a science has roots in studying phenotypes of relatives, but molecular approaches facilitate direct measurements of genomic variation within individuals. Agricultural and human biomedical research are both emphasizing genotype-based instruments, like polygenic scores, as the future of breeding programs or precision medicine and genetic epidemiology. However, unlike in agriculture, there is an emerging consensus that family variables act near independent of genotypes in models of human disease. To advance our understanding of this phenomenon, we use 2,066,057 family records of 99,645 genotyped probands from the iPSYCH2015 case-cohort study to show that state-of-the-field genotype- and phenotype-based genetic instruments explain essentially independent components of liability to psychiatric disorders. We support these empirical results with novel theoretical analysis and simulations to describe, in a human biomedical context, parameters affecting current and future performance of the two approaches, their expected interrelationships, relative sample size efficiencies, and the striking consistency of observed results with expectations under simple additive, polygenic liability models of disease. We conclude, at least for psychiatric disorders, that phenotype- and genotype-based genetic instruments are likely to be very noisy measures of the same underlying additive genetic liability, should be seen, in the near future, as complementary, and integrated to a greater extent going forward.

No abstract