5 Methods for Association Studies
97
Studies designed for the purposes of replication should ensure that sample sizes
are sufficiently large to detect associations of the hypothesized magnitudes. In
fact, sample sizes should ideally be larger than those of the initial study so as
to account for overestimation in the original sample (unless one only wishes to
replicate a limited number of variants). The larger the sample, the better success
replication studies will have in reproducing results from and identifying false
positives generated by the initial study. Replication studies should also evaluate
the same ancestral population as the discovery study and, ideally, the same genetic
variant with respect to the same definition of phenotype. Successful replication then
entails finding the same direction of association (for the same effect allele) at a
predetermined threshold for statistical significance. What that threshold should be
is somewhat controversial; some investigators expect that associations be replicated
at a genome-wide significance level, whereas others apply a less conservative
threshold based on evaluating a smaller number of variants in the replication sample.
Still others are not as concerned with the statistical significance of the replication
association as they are with the significance of the joint analysis of discovery and
replication.
In some cases, studies are not designed exclusively for the purposes of replication. Rather, colleagues may help one another replicate their strongest results by
looking them up in independent, existing “discovery” studies. Once results are confirmed in the original target populations, investigators may also choose to evaluate
associations in populations of varying ancestries. Results that replicate from these
studies are often said to generalize, meaning that the effect is relevant to multiple
human populations. In contrast to replication, studies conducted for generalization
should draw from an ancestral population different from the discovery population.
It should be noted that while replication has become standard practice to
corroborate genetic associations, it may not be as necessary as it once was. As
genetic association studies have become increasingly sizeable and larger numbers
of markers have been genotyped in large replication samples, the statistical power
to detect modest effects has substantially increased. As a result, the potential for
winner’s curse has decreased. Still, replication inspires confidence in findings and
remains customary for genetic association studies.
5.2.5.2 Meta-analysis
Results from multiple studies or even multiple stages of the same study can
be combined into a single result via meta-analysis. Meta-analytical methods
synthesize results from analyses that examine the same hypothesis without accessing individual-level data (as mega-analytical studies would). In doing so, they
considerably boost the sample size and power for examining the hypothesis and
thus may achieve a more precise estimate of the association of interest. Several
software packages are available for the implementation of meta-analysis for GWAS,
among which are METASOFT (Han and Eskin 2011), METAL (Willer et al.
2010), GWAMA (Magi and Morris 2010), PLINK (Purcell et al. 2007), and
GenABEL/MetABEL (Aulchenko et al. 2007). Available features in most of these
packages were summarized in a side-by-side comparison (Evangelou and Ioannidis
97
Studies designed for the purposes of replication should ensure that sample sizes
are sufficiently large to detect associations of the hypothesized magnitudes. In
fact, sample sizes should ideally be larger than those of the initial study so as
to account for overestimation in the original sample (unless one only wishes to
replicate a limited number of variants). The larger the sample, the better success
replication studies will have in reproducing results from and identifying false
positives generated by the initial study. Replication studies should also evaluate
the same ancestral population as the discovery study and, ideally, the same genetic
variant with respect to the same definition of phenotype. Successful replication then
entails finding the same direction of association (for the same effect allele) at a
predetermined threshold for statistical significance. What that threshold should be
is somewhat controversial; some investigators expect that associations be replicated
at a genome-wide significance level, whereas others apply a less conservative
threshold based on evaluating a smaller number of variants in the replication sample.
Still others are not as concerned with the statistical significance of the replication
association as they are with the significance of the joint analysis of discovery and
replication.
In some cases, studies are not designed exclusively for the purposes of replication. Rather, colleagues may help one another replicate their strongest results by
looking them up in independent, existing “discovery” studies. Once results are confirmed in the original target populations, investigators may also choose to evaluate
associations in populations of varying ancestries. Results that replicate from these
studies are often said to generalize, meaning that the effect is relevant to multiple
human populations. In contrast to replication, studies conducted for generalization
should draw from an ancestral population different from the discovery population.
It should be noted that while replication has become standard practice to
corroborate genetic associations, it may not be as necessary as it once was. As
genetic association studies have become increasingly sizeable and larger numbers
of markers have been genotyped in large replication samples, the statistical power
to detect modest effects has substantially increased. As a result, the potential for
winner’s curse has decreased. Still, replication inspires confidence in findings and
remains customary for genetic association studies.
5.2.5.2 Meta-analysis
Results from multiple studies or even multiple stages of the same study can
be combined into a single result via meta-analysis. Meta-analytical methods
synthesize results from analyses that examine the same hypothesis without accessing individual-level data (as mega-analytical studies would). In doing so, they
considerably boost the sample size and power for examining the hypothesis and
thus may achieve a more precise estimate of the association of interest. Several
software packages are available for the implementation of meta-analysis for GWAS,
among which are METASOFT (Han and Eskin 2011), METAL (Willer et al.
2010), GWAMA (Magi and Morris 2010), PLINK (Purcell et al. 2007), and
GenABEL/MetABEL (Aulchenko et al. 2007). Available features in most of these
packages were summarized in a side-by-side comparison (Evangelou and Ioannidis