100
R. E. Graff et al.
genotyping, investigators sometimes use genotype information from controls who
have been recruited into prior studies and that has been made publicly available to
researchers (e.g., via the database of genotypes and phenotypes (dbGaP)) (Luca et
al. 2008; Burton et al. 2007; Paltoo et al. 2014). The inclusion of controls from
public databases can also increase statistical power without affecting costs (Ho and
Lange 2010). The potential bias arising from the use of such “convenience” or
“public” controls is mitigated by the low measurement error in SNP genotyping,
the absence of recall bias when studying inherited variants, large sample sizes,
stringent criteria for statistical significance, and rigorous replication of findings.
Nevertheless, the use of convenience controls may result in the confounding of
associations due to population stratification (discussed further below). One should
thus address the bias analytically with genetic information (Devlin and Roeder 1999;
Price et al. 2006; Pritchard and Rosenberg 1999; Mitchell et al. 2014). One must
also consider the potential for batch effects due to differences in genotyping quality
control procedures and phenotype misclassification in individuals not thoroughly
screened for common diseases. These issues can be assessed in small subsets of
the sample by comparing genotype concordance in re-genotyped individuals or by
conducting sensitivity analyses of the phenotype (Mitchell et al. 2014). Restricting
the use of controls to those genotyped on the same platform and from the same
genetic ancestry as cases may prevent or reduce these biases (Sinnott and Kraft
2012).
5.3.3 Sample Size
As with any study, it is critical that genetic association studies include a sufficiently
large sample size to ensure good statistical power. The power of a study depends on
the unknown frequency and effect size of the causal genetic variant(s) for which one
is searching. Whenever SNPs in LD with the true causal variant are genotyped rather
than the causal variant itself, power is reduced; the sample size required will be
inflated proportionally to the inverse of the correlation between the genotyped and
causative markers. There undoubtedly exist genetic variants that have a small effect
on disease but that have not been detected due to insufficient sample size. In general,
it is rare for successful GWAS to include fewer than 1000 cases and 1000 controls,
and many include substantially larger numbers of individuals. Among the largest
GWAS conducted to date have investigated smoking initiation (n = 1,232,091)
(Liu et al. 2019), educational attainment (n = 1,131,881), blood pressure traits
(n = 1,006,863) (Evangelou et al. 2018), and risk tolerance (n = 975,353) (Karlsson
Linner et al. 2019).
5.4
Measurement of Genetic Information
Accurate measurement of genetic information is yet another crucial component
of a reliable genetic association study. The study design is likely to inform the
appropriate category of measurement broadly, but each method requires nuanced
R. E. Graff et al.
genotyping, investigators sometimes use genotype information from controls who
have been recruited into prior studies and that has been made publicly available to
researchers (e.g., via the database of genotypes and phenotypes (dbGaP)) (Luca et
al. 2008; Burton et al. 2007; Paltoo et al. 2014). The inclusion of controls from
public databases can also increase statistical power without affecting costs (Ho and
Lange 2010). The potential bias arising from the use of such “convenience” or
“public” controls is mitigated by the low measurement error in SNP genotyping,
the absence of recall bias when studying inherited variants, large sample sizes,
stringent criteria for statistical significance, and rigorous replication of findings.
Nevertheless, the use of convenience controls may result in the confounding of
associations due to population stratification (discussed further below). One should
thus address the bias analytically with genetic information (Devlin and Roeder 1999;
Price et al. 2006; Pritchard and Rosenberg 1999; Mitchell et al. 2014). One must
also consider the potential for batch effects due to differences in genotyping quality
control procedures and phenotype misclassification in individuals not thoroughly
screened for common diseases. These issues can be assessed in small subsets of
the sample by comparing genotype concordance in re-genotyped individuals or by
conducting sensitivity analyses of the phenotype (Mitchell et al. 2014). Restricting
the use of controls to those genotyped on the same platform and from the same
genetic ancestry as cases may prevent or reduce these biases (Sinnott and Kraft
2012).
5.3.3 Sample Size
As with any study, it is critical that genetic association studies include a sufficiently
large sample size to ensure good statistical power. The power of a study depends on
the unknown frequency and effect size of the causal genetic variant(s) for which one
is searching. Whenever SNPs in LD with the true causal variant are genotyped rather
than the causal variant itself, power is reduced; the sample size required will be
inflated proportionally to the inverse of the correlation between the genotyped and
causative markers. There undoubtedly exist genetic variants that have a small effect
on disease but that have not been detected due to insufficient sample size. In general,
it is rare for successful GWAS to include fewer than 1000 cases and 1000 controls,
and many include substantially larger numbers of individuals. Among the largest
GWAS conducted to date have investigated smoking initiation (n = 1,232,091)
(Liu et al. 2019), educational attainment (n = 1,131,881), blood pressure traits
(n = 1,006,863) (Evangelou et al. 2018), and risk tolerance (n = 975,353) (Karlsson
Linner et al. 2019).
5.4
Measurement of Genetic Information
Accurate measurement of genetic information is yet another crucial component
of a reliable genetic association study. The study design is likely to inform the
appropriate category of measurement broadly, but each method requires nuanced