5 Methods for Association Studies
95
non-Hispanic whites (P value: 1.0 × 10 −211 ) (Hoffmann et al. 2015). More recently,
a PRS for breast cancer based on 313 variants demonstrated strong predictive
performance (AUC = 0.630) and identified 19% of women who could be eligible
for early screening at age 40 (Mavaddat et al. 2019). Still, few individuals will
carry large numbers of risk alleles from GWAS, though essentially all individuals
will carry some risk alleles. Screening for them in the general population is thus
unlikely to be cost-effective, unless individuals receive genome-wide evaluations. In
addition, predictive models may have worse performance in ancestral populations
other than those in which the models were discovered, because effect size estimates
will be diluted when SNPs in populations with one set of LD patterns (e.g.,
Europeans) are applied to populations with a different set of LD patterns (e.g.,
African Americans) (Carlson et al. 2013). Note also that justification for genetic
testing additionally depends on the existence of effective interventions.
5.2.3 Mendelian Randomization
In some instances, genetic variation can be leveraged toward evaluating causal relationships between exposures and outcomes that may be challenging to investigate
in traditional observational studies. By using a genetic predictor of exposure as an
instrumental variable, Mendelian randomization circumvents issues of confounding
and reverse causation that often afflict epidemiological studies. While the method
has been around for several decades (Gray and Wheatley 1991; Katan 1986; Smith
and Ebrahim 2003), its use has exploded with the ever-increasing discovery of traitassociated variants and modern statistical methods for high-dimensional genetic
data. In general, its implementation requires the identification of a set of genetic
variants that is predictive of the exposure of interest followed by the performance of
instrumental variable analyses (Burgess et al. 2013; Pierce and Burgess 2013).
As with all instrumental variable approaches, Mendelian randomization is
premised on three assumptions: (1) the genetic instrument is associated with the
exposure, (2) the genetic instrument shares no common causes with the outcome,
and (3) the genetic instrument only affects the outcome through exposure. The
first assumption is easily satisfied by selecting genetic variants that are strongly
associated with the exposure of interest, such as those reaching genome-wide
significance. The second assumption can be at least partially verified by assessing associations between genetic instruments and known confounders. The third
assumption, however, cannot be substantiated empirically. Nevertheless, sensitivity
analyses can help evaluate the consistency and robustness of observed results
(Bowden et al. 2017; Haycock et al. 2016).
95
non-Hispanic whites (P value: 1.0 × 10 −211 ) (Hoffmann et al. 2015). More recently,
a PRS for breast cancer based on 313 variants demonstrated strong predictive
performance (AUC = 0.630) and identified 19% of women who could be eligible
for early screening at age 40 (Mavaddat et al. 2019). Still, few individuals will
carry large numbers of risk alleles from GWAS, though essentially all individuals
will carry some risk alleles. Screening for them in the general population is thus
unlikely to be cost-effective, unless individuals receive genome-wide evaluations. In
addition, predictive models may have worse performance in ancestral populations
other than those in which the models were discovered, because effect size estimates
will be diluted when SNPs in populations with one set of LD patterns (e.g.,
Europeans) are applied to populations with a different set of LD patterns (e.g.,
African Americans) (Carlson et al. 2013). Note also that justification for genetic
testing additionally depends on the existence of effective interventions.
5.2.3 Mendelian Randomization
In some instances, genetic variation can be leveraged toward evaluating causal relationships between exposures and outcomes that may be challenging to investigate
in traditional observational studies. By using a genetic predictor of exposure as an
instrumental variable, Mendelian randomization circumvents issues of confounding
and reverse causation that often afflict epidemiological studies. While the method
has been around for several decades (Gray and Wheatley 1991; Katan 1986; Smith
and Ebrahim 2003), its use has exploded with the ever-increasing discovery of traitassociated variants and modern statistical methods for high-dimensional genetic
data. In general, its implementation requires the identification of a set of genetic
variants that is predictive of the exposure of interest followed by the performance of
instrumental variable analyses (Burgess et al. 2013; Pierce and Burgess 2013).
As with all instrumental variable approaches, Mendelian randomization is
premised on three assumptions: (1) the genetic instrument is associated with the
exposure, (2) the genetic instrument shares no common causes with the outcome,
and (3) the genetic instrument only affects the outcome through exposure. The
first assumption is easily satisfied by selecting genetic variants that are strongly
associated with the exposure of interest, such as those reaching genome-wide
significance. The second assumption can be at least partially verified by assessing associations between genetic instruments and known confounders. The third
assumption, however, cannot be substantiated empirically. Nevertheless, sensitivity
analyses can help evaluate the consistency and robustness of observed results
(Bowden et al. 2017; Haycock et al. 2016).