Principal components analysis corrects for stratification
in genome-wide association studies
Alkes L Price
1,2
, Nick J Patterson
2
, Robert M Plenge
2,3
, Michael E Weinblatt
3
, Nancy A Shadick
3
&
David Reich
1,2
Population stratification—allele frequency differences between cases and controls due to systematic ancestry differences—can
cause spurious associations in disease studies. We describe a method that enables explicit detection and correction of population
stratification on a genome-wide scale. Our method uses principal components analysis to explicitly model ancestry differences
between cases and controls. The resulting correction is specific to a candidate marker’s variation in frequency across ancestral
populations, minimizing spurious associations while maximizing power to detect true associations. Our simple, efficient approach
can easily be applied to disease studies with hundreds of thousands of markers.
Population stratification—allele frequency differences between cases
and controls due to systematic ancestry differences—can cause spur-
ious associations in disease studies
1–8
. Because the effects of stratifica-
tion vary in proportion to the number of samples
9
, stratification will
be an increasing problem in the large-scale association studies of the
future, which will analyze thousands of samples in an effort to detect
common genetic variants of weak effect.
The two prevailing methods for dealing with stratification are
genomic control and structured association
9–14
. Although genomic
control and structured association have proven useful in a variety of
contexts, they have limitations. Genomic control corrects for stratifi-
cation by adjusting association statistics at each marker by a uniform
overall inflation factor. However, some markers differ in their allele
frequencies across ancestral populations more than others. Thus, the
uniform adjustment applied by genomic control may be insufficient at
markers having unusually strong differentiation across ancestral
populations and may be superfluous at markers devoid of such
differentiation, leading to a loss in power. Structured association
uses a program such as STRUCTURE
15
to assign the samples to
discrete subpopulation clusters and then aggregates evidence of
association within each cluster. If fractional membership in more
than one cluster is allowed, the method cannot currently be applied to
genome-wide association studies because of its intensive computa-
tional cost on large data sets. Furthermore, assignments of individuals
to clusters are highly sensitive to the number of clusters, which is not
well defined
14,16
.
We propose a method to detect and correct for population
stratification that addresses these limitations. Our method, EIGEN-
STRAT, consists of three steps (Fig. 1). First, we apply principal
components analysis
17
to genotype data to infer continuous axes of
genetic variation. Intuitively, the axes of variation reduce the data to a
small number of dimensions, describing as much variability as
possible; they are defined as the top eigenvectors of a covariance
matrix between samples (see Methods). In data sets with ancestry
differences between samples, axes of variation often have a geographic
interpretation: for example, an axis describing a northwest-southeast
cline in Europe would have values that gradually range from positive
for samples from northwest Europe, to near zero in central Europe, to
negative in southeast Europe. Second, we continuously adjust
genotypes and phenotypes by amounts attributable to ancestry
along each axis, via computing residuals of linear regressions;
intuitively, this creates a virtual set of matched cases and controls.
Third, we compute association statistics using ancestry-adjusted
genotypes and phenotypes.
The EIGENSTRAT method has arisen out of our systematic
exploration of the use of principal components analysis in a more
general population genetic context. Principal components analysis was
originally applied to genetic data to infer worldwide axes of human
genetic variation from the allele frequencies of various popula-
tions
18,19
. We have further developed this approach in a parallel
paper (N.J.P., A.L.P. and D.R., unpublished data), focusing instead
on individual genotype data and placing the method on a firm
statistical footing by rigorously assigning statistical significance to
each axis of variation
20–22
. EIGENSTRAT applies this toolkit to analyze
population structure in the context of disease studies.
Correcting for stratification using continuous axes of variation has
several advantages. Continuous axes provide the most useful descrip-
tion of within-continent genetic variation, according to recent stu-
dies
23
. Because our continuous axes are constructed to be orthogonal,
results are insensitive to the number of axes inferred, as we verify
Received 23 March; accepted 21 June; published online 23 July 2006; doi:10.1038/ng1847
1
Department of Genetics, Harvard Medical School, Boston, Massachusetts 02115, USA.
2
Program in Medical and Population Genetics, Broad Institute of MIT and
Harvard, 7 Cambridge Center, Cambridge, Massachusetts 02142, USA.
3
Division of Rheumatology, Immunology and Allergy, Brigham and Women’s Hospital, Boston,
Massachusetts 02115, USA. Correspondence should be addressed to A.L.P. (aprice@broad.mit.edu).
904 VOLUME 38 [ NUMBER 8 [ AUGUST 2006 NATURE GENETICS
ARTICLES
© 2006 Nature Publishing Group http://www.nature.com/naturegenetics