Comparison of two methods for analysis of gene–environment interactions in longitudinal family data: the Framingham heart study

Abstract

Gene–environment interaction (GEI) analysis can potentially enhance gene discovery for common complex traits. However, genome-wide interaction analysis is computationally intensive. Moreover, analysis of longitudinal data in families is much more challenging due to the two sources of correlations arising from longitudinal measurements and family relationships. GWIS of longitudinal family data can be a computational bottleneck. Therefore, we compared two methods for analysis of longitudinal family data: a methodologically sound but computationally demanding method using the Kronecker model (KRC) and a computationally more forgiving method using the hierarchical linear model (HLM). The KRC model uses a Kronecker product of an unstructured matrix for correlations among repeated measures (longitudinal) and a compound symmetry matrix for correlations within families at a given visit. The HLM uses an autoregressive covariance matrix for correlations among repeated measures and a random intercept for familial correlations. We compared the two methods using the longitudinal Framingham heart study (FHS) SHARe data. Specifically, we evaluated SNP–alcohol (amount of alcohol consumption) interaction effects on high density lipoprotein cholesterol (HDLC). Keeping the prohibitive computational burden of KRC in mind, we limited the analysis to chromosome 16, where preliminary cross-sectional analysis yielded some interesting results. Our first important finding was that the HLM provided very comparable results but was remarkably faster than the KRC, making HLM the method of choice. Our second finding was that longitudinal analysis provided smaller P-values, thus leading to more significant results, than cross-sectional analysis. This was particularly pronounced in identifying GEIs. We conclude that longitudinal analysis of GEIs is more powerful and that the HLM method is an optimal method of choice as compared to the computationally (prohibitively) intensive KRC method.

Modeling gene-environment interactions in longitudinal family studies: a comparison of methods and their application to the association between the IGF pathway and childhood obesity

“…One approach involves incorporating random effects with an assumed covariance structure to accommodate both familial and over time correlations. This can be achieved in the linear mixed model framework [31, 32]. Alternatively, simplified correlation structure allows the application of models not designed for longitudinal family data.…”

A Comparison of Statistical Methods for the Discovery of Genetic Risk Factors Using Longitudinal Family Study Designs

“…For methods of type (2), many groups have evaluated either the GEE or GLMM approach to longitudinal analysis when the family structure is simplified. For example, Choi et al ( 23 ), Shi et al ( 24 ), and Sung et al ( 25 ) all include a random family effect in their mixed effect models, implicitly assuming that the correlation between observations from related family members is constant regardless of the relationships. In the GEE approach evaluated by Shi et al ( 24 ) and Sung et al ( 25 ), a compound symmetry correlation structure is assumed for the correlation between family members, which explicitly assumes constant correlation between family members.…”

“…Although many different studies compare the performance of some of the methods outlined above [for example, Ref. ( 23 – 25 )], comparisons are not exhaustive as typically only two methods within each of the three types, for example the marginal to multi-level model, are compared. In addition, we simulated data under multiple different genetic models, and we include the Bayesian approach in our comparison.…”

Assessing the effects of multiple markers in genetic association studies

“…Six manuscripts in this eBook addressed gene-gene interactions, or the related topic of gene-environment interactions. Lee et al ( 2013 ) and Sung et al ( 2014 ) investigated methods for the analysis of interactions in family data. Chen and Guo, and Millstein discussed potential solutions to the challenge of high dimensionality in interaction studies.…”

“…First, computational burden can become prohibitive when interactions are investigated in genome-wide data. Chen and Guo ( 2013 ) explored the possibility of overcoming this constraint through the use of processor graphics cards (GPU), while in the analysis of longitudinal family data, Sung et al ( 2014 ) proposed using a computationally less intensive method based on the hierarchical linear model (HLM). Second, it is well-known that interaction analyses require very large sample sizes, and real data analyses often fail to achieve genome-wide significance (as shown in de las Fuentes et al, 2013 ).…”