98
R. E. Graff et al.
2013). In addition, there exist tools to meta-analyze results from populations of
varying ethnicities (Hong et al. 2016; Morris 2011).
Meta-analytical methods can also be used to discover novel genetic loci with
pleiotropic effects and to explore associations across phenotypes or disease subtypes. Association analysis based on subsets (ASSET) is a flexible meta-analysis
framework that can evaluate associations for a given SNP across phenotypes and
identify the combination of associated traits that maximizes the overall test statistic
(Bhattacharjee et al. 2012). In addition to boosting power in the presence of
heterogeneity, attractive features of ASSET are its ability to account for sample
overlap across contributing studies and its internal correction for the multiple tests
required by the subset search. ASSET has been applied to a number of traits, among
which are multiple cancers (Fehringer et al. 2016) and immune-related diseases
(Marquez et al. 2018; Zhu et al. 2018).
To conduct a rigorous meta-analysis, all studies should be subject to a standard
quality control procedure that determines which SNPs are included in each study. It
is a fundamental assumption of meta-analysis that the studies provide independent
information, so it is also critical to ensure that there not be any overlap in the
samples included from each study. In addition, the design of each study incorporated
into a meta-analysis should ideally be similar; the measurement of covariates and
phenotypes should be analogous, analytic procedures should be comparable, and
covariate adjustment should be standardized (Zeggini and Ioannidis 2009). It is also
important that all studies report results using the same reference allele and mode of
inheritance. Imputation is often required to ensure that all studies in a meta-analysis
offer data about the same SNPs (discussed further below).
The most common method to estimate an average effect across studies is fixedeffects modeling that weights each study effect based on its inverse variance.
Mixed-effects models may also be used when there is substantial heterogeneity
of effects across studies; their random effect parameters can help account for the
heterogeneity. Regardless of the model selected for meta-analysis, it is important to
quantify the differences across studies, particularly given that it is rare that studies
perfectly fulfill the stringent criteria for meta-analysis. The most commonly used
measures to do so are the Q statistic and I 2 index (Evangelou and Ioannidis 2013;
Huedo-Medina et al. 2006; Panagiotou et al. 2013).
5.3
Design of Association Studies
The first step toward obtaining meaningful results from any genetic association
study is designing it effectively. Investigators must always define appropriate
phenotypes, designate a valid study population, and ensure a sufficient sample size.
In this section, we outline some of these key elements that should be contemplated
in conceiving new studies.
R. E. Graff et al.
2013). In addition, there exist tools to meta-analyze results from populations of
varying ethnicities (Hong et al. 2016; Morris 2011).
Meta-analytical methods can also be used to discover novel genetic loci with
pleiotropic effects and to explore associations across phenotypes or disease subtypes. Association analysis based on subsets (ASSET) is a flexible meta-analysis
framework that can evaluate associations for a given SNP across phenotypes and
identify the combination of associated traits that maximizes the overall test statistic
(Bhattacharjee et al. 2012). In addition to boosting power in the presence of
heterogeneity, attractive features of ASSET are its ability to account for sample
overlap across contributing studies and its internal correction for the multiple tests
required by the subset search. ASSET has been applied to a number of traits, among
which are multiple cancers (Fehringer et al. 2016) and immune-related diseases
(Marquez et al. 2018; Zhu et al. 2018).
To conduct a rigorous meta-analysis, all studies should be subject to a standard
quality control procedure that determines which SNPs are included in each study. It
is a fundamental assumption of meta-analysis that the studies provide independent
information, so it is also critical to ensure that there not be any overlap in the
samples included from each study. In addition, the design of each study incorporated
into a meta-analysis should ideally be similar; the measurement of covariates and
phenotypes should be analogous, analytic procedures should be comparable, and
covariate adjustment should be standardized (Zeggini and Ioannidis 2009). It is also
important that all studies report results using the same reference allele and mode of
inheritance. Imputation is often required to ensure that all studies in a meta-analysis
offer data about the same SNPs (discussed further below).
The most common method to estimate an average effect across studies is fixedeffects modeling that weights each study effect based on its inverse variance.
Mixed-effects models may also be used when there is substantial heterogeneity
of effects across studies; their random effect parameters can help account for the
heterogeneity. Regardless of the model selected for meta-analysis, it is important to
quantify the differences across studies, particularly given that it is rare that studies
perfectly fulfill the stringent criteria for meta-analysis. The most commonly used
measures to do so are the Q statistic and I 2 index (Evangelou and Ioannidis 2013;
Huedo-Medina et al. 2006; Panagiotou et al. 2013).
5.3
Design of Association Studies
The first step toward obtaining meaningful results from any genetic association
study is designing it effectively. Investigators must always define appropriate
phenotypes, designate a valid study population, and ensure a sufficient sample size.
In this section, we outline some of these key elements that should be contemplated
in conceiving new studies.