Bivariate association analysis of longitudinal phenotypes in families

Constrained multivariate association with longitudinal phenotypes

BACKGROUND

The incorporation of longitudinal data into genetic epidemiological studies has the potential to provide valuable information regarding the effect of time on complex disease etiology. Yet, the majority of research focuses on variables collected from a single time point. This aim of this study was to test for main effects on a quantitative trait across time points using a constrained maximum-likelihood measured genotype approach. This method simultaneously accounts for all repeat measurements of a phenotype in families. We applied this method to systolic blood pressure (SBP) measurements from three time points using the Genetic Analysis Workshop 19 (GAW19) whole-genome sequence family simulated data set and 200 simulated replicates. Data consisted of 849 individuals from 20 extended Mexican American pedigrees. Comparisons were made among 3 statistical approaches: (a) constrained, where the effect of a variant or gene region on the mean trait value was constrained to be equal across all measurements; (b) unconstrained, where the variant or gene region effect was estimated separately for each time point; and (c) the average SBP measurement from three time points. These approaches were run for nine genetic variants with known effect sizes (>0.001) for SBP variability and a known gene-centric kernel (MAP4)-based test under the GAW19 simulation model across 200 replicates.

RESULTS

When compared to results using two time points, the constrained method utilizing all 3 time points increased power to detect association. Averaging SBP was equally effective when the variant has a large effect on the phenotype, but less powerful for variants with lower effect sizes. However, averaging SBP was far more effective than either the constrained or unconstrained approaches when using a gene-centric kernel-based test.

CONCLUSION

We determined that this constrained multivariate approach improves genetic signal over the bivariate method. However, this method is still only effective in those variants that explain a moderate to large proportion of the phenotypic variance but is not as effective for gene-centric tests.

Rare-Variant Kernel Machine Test for Longitudinal Data from Population and Family Samples

OBJECTIVE

The kernel machine (KM) test reportedly performs well in the set-based association test of rare variants. Many studies have been conducted to measure phenotypes at multiple time points, but the standard KM methodology has only been available for phenotypes at a single time point. In addition, family-based designs have been widely used in genetic association studies; therefore, the data analysis method used must appropriately handle familial relatedness. A rare-variant test does not currently exist for longitudinal data from family samples. Therefore, in this paper, we aim to introduce an association test for rare variants, which includes multiple longitudinal phenotype measurements for either population or family samples.

METHODS

This approach uses KM regression based on the linear mixed model framework and is applicable to longitudinal data from either population (L-KM) or family samples (LF-KM).

RESULTS

In our population-based simulation studies, L-KM has good control of Type I error rate and increased power in all the scenarios we considered compared with other competing methods. Conversely, in the family-based simulation studies, we found an inflated Type I error rate when L-KM was applied directly to the family samples, whereas LF-KM retained the desired Type I error rate and had the best power performance overall. Finally, we illustrate the utility of our proposed LF-KM approach by analyzing data from an association study between rare variants and blood pressure from the Genetic Analysis Workshop 18 (GAW18).

CONCLUSION

We propose a method for rare-variant association testing in population and family samples using phenotypes measured at multiple time points for each subject. The proposed method has the best power performance compared to competing approaches in our simulation study.

Longitudinal data analysis for genetic studies in the whole-genome sequencing era

The analysis of whole-genome sequence (WGS) data using longitudinal phenotypes offers a potentially rich resource for the examination of the genetic variants and their covariates that affect complex phenotypes over time. We summarize eight contributions to the Genetic Analysis Workshop 18, which applied a diverse array of statistical genetic methods to analyze WGS data in combination with data from genome-wide association studies (GWAS) from up to four different time points on blood pressure phenotypes. The common goal of these analyses was to develop and apply appropriate methods that utilize longitudinal repeated measures to potentially increase the analytic efficiency of WGS and GWAS data. These diverse methods can be grouped into two categories, based on the way they model dependence structures: (1) linear mixed-effects (LME) models, where the random effect terms in the linear models are used to capture the dependence structures; and (2) variance-components models, where the dependence structures are constructed directly based on multiple components of variance-covariance matrices for the multivariate Gaussian responses. Despite the heterogeneous nature of these analytical methods, the group came to the following conclusions: (1) the use of repeat measurements can gain power to identify variants associated with the phenotype; (2) the inclusion of family data may correct genotyping errors and allow for more accurate detection of rare variants than using unrelated individuals only; and (3) fitting mixed-effects and variance-components models for longitudinal data presents computational challenges. The challenges and computational burden demanded by WGS data were addressed in the eight contributions.