7 What Have We Learned from GWAS?
163
study, the technicians in the lab generating the data, and the analysts who work
with the output of these labors. A corollary from this principle is that the task
of robustly checking data quality—ensuring that patterns match expectation—is
the first, foremost, and most labor-intensive activity for any genome-wide study,
especially for new data types on cutting-edge technologies. For example, the first
GWAS testing of common copy-number polymorphism data with disease followed
the same trajectory as for single-nucleotide polymorphisms but required additional
checks and filters for quality (Myocardial Infarction Genetics Consortium 2009). As
the field moves forward into new data types and technologies each with their own
biases, the establishment of a procedure of stringent quality control will be essential
to ensure the pace of discovery and the avoidance of false-positive associations (see
retraction from Sebastiani et al. 2011).
7.2.2 Addressing Confounding from Population Stratification
After the technical control and application of high-quality data had been achieved,
an essential biological phenomenon to control in the GWAS design is population
ancestry. It had been known for some time that spurious, false-positive associations
between marker and phenotype could occur when the prevalence of the phenotype
of interest differed across sampled populations and if one does not take this fact into
account when sampling and performing association testing (Knowler et al. 1988;
Campbell et al. 2005). Though well-matched association designs were far more
statistically powerful and cheaper than the previous linkage or family-based studies
(Risch and Merikangas 1996), there was a brief time where it was not entirely
clear how one could achieve ancestry matching in practice. As the need arose,
several statistical developments preceding the GWAS era arose, either to correct the
distribution of generated test statistics (Devlin and Roeder 1999), proposing either
formal tests for cases and controls matching (Pritchard and Rosenberg 1999), using
the genetic data to infer ancestry directly (Pritchard et al. 2000), or summarize the
genetic similarities of samples using principal components analysis and control for
those vectors in association testing (Price et al. 2006). (Readers interested further in
the subject should refer to previous chapters that further detail methods.) Sufficed to
say, the application of these approaches in the context of genome-wide association
studies was highly successful, as overtime, discoveries across a range of phenotypes
continue to be consistently supported by the influx of additional data sets and across
a range of ethnic groups.
7.2.3 Threshold for Declaring Significant Genome-Wide Results
Another point that required resolution was the determination of an appropriate
genome-wide statistical threshold that maintained the appropriate error rates, given
the number of markers tested. If Kruglyak’s estimate of markers was correct
(that ∼500,000 markers would be needed), what would such a threshold actually
163
study, the technicians in the lab generating the data, and the analysts who work
with the output of these labors. A corollary from this principle is that the task
of robustly checking data quality—ensuring that patterns match expectation—is
the first, foremost, and most labor-intensive activity for any genome-wide study,
especially for new data types on cutting-edge technologies. For example, the first
GWAS testing of common copy-number polymorphism data with disease followed
the same trajectory as for single-nucleotide polymorphisms but required additional
checks and filters for quality (Myocardial Infarction Genetics Consortium 2009). As
the field moves forward into new data types and technologies each with their own
biases, the establishment of a procedure of stringent quality control will be essential
to ensure the pace of discovery and the avoidance of false-positive associations (see
retraction from Sebastiani et al. 2011).
7.2.2 Addressing Confounding from Population Stratification
After the technical control and application of high-quality data had been achieved,
an essential biological phenomenon to control in the GWAS design is population
ancestry. It had been known for some time that spurious, false-positive associations
between marker and phenotype could occur when the prevalence of the phenotype
of interest differed across sampled populations and if one does not take this fact into
account when sampling and performing association testing (Knowler et al. 1988;
Campbell et al. 2005). Though well-matched association designs were far more
statistically powerful and cheaper than the previous linkage or family-based studies
(Risch and Merikangas 1996), there was a brief time where it was not entirely
clear how one could achieve ancestry matching in practice. As the need arose,
several statistical developments preceding the GWAS era arose, either to correct the
distribution of generated test statistics (Devlin and Roeder 1999), proposing either
formal tests for cases and controls matching (Pritchard and Rosenberg 1999), using
the genetic data to infer ancestry directly (Pritchard et al. 2000), or summarize the
genetic similarities of samples using principal components analysis and control for
those vectors in association testing (Price et al. 2006). (Readers interested further in
the subject should refer to previous chapters that further detail methods.) Sufficed to
say, the application of these approaches in the context of genome-wide association
studies was highly successful, as overtime, discoveries across a range of phenotypes
continue to be consistently supported by the influx of additional data sets and across
a range of ethnic groups.
7.2.3 Threshold for Declaring Significant Genome-Wide Results
Another point that required resolution was the determination of an appropriate
genome-wide statistical threshold that maintained the appropriate error rates, given
the number of markers tested. If Kruglyak’s estimate of markers was correct
(that ∼500,000 markers would be needed), what would such a threshold actually