Statistical analysis of genome-wide association

(GWAS) data

Jim Stankovich
Menzies Research Institute
University of Tasmania
J.Stankovich@utas.edu.au
 Outline

• Introduction
• Confounding variables and linkage disequilibrium
• Statistical methods to test for association in case-control GWA
studies
– Allele counting chi-square test
– Logistic regression
• Multiple testing and power
• Example: GWAS for multiple sclerosis (MS)
– Data cleaning / quality control
– Results
 GWA studies have been very successful since 2007

• Prior to the advent of GWA studies, there was very little success in
identifying genetic risk factors for complex multifactorial diseases
• GWA studies have identified over 200 separate associations with
various complex diseases in the past two years
• “Human Genetic Variation” hailed as “Breakthrough of the Year” by
Science magazine in 2007

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009

Human genome project

SNP genotyping arrays

The SNP consortium
GWA studies
The International HapMap Project
 This talk: case-control GWA studies

• Obtain DNA from people with disease of interest (cases) and

unaffected controls
• Run each DNA sample on a SNP chip to measure genotypes at
300,000-1,000,000 SNPs in cases and controls
• Identify SNPs where one allele is significantly more common in
cases than controls
– The SNP is associated with disease

SNP: rs12425791 stroke

 This talk: case-control GWA studies

• Obtain DNA from people with disease of interest (cases) and

unaffected controls
• Run each DNA sample on a SNP chip to measure genotypes at
300,000-1,000,000 SNPs in cases and controls
• Identify SNPs where one allele is significantly more common in
cases than controls
– The SNP is associated with disease
• Alternative strategy (Peter Visscherʼs talk):
test for association between SNPs and a quantitative trait that
underlies the disease (endophenotype)

SNP: rs12425791 blood pressure stroke

 Association does not imply causation

• Suppose that genotypes at a particular SNP are significantly

associated with disease
• This may be because the SNP is associated with some other factor
(a confounder), which is associated with disease but is not in the
same causal pathway

SNP near lactase gene multiple sclerosis (MS)

 Association does not imply causation

• Suppose that genotypes at a particular SNP are significantly

associated with disease
• This may be because the SNP is associated with some other factor
(a confounder), which is associated with disease but is not in the
same causal pathway

Northern European
ancestry

SNP near lactase gene multiple sclerosis (MS)

 Association does not imply causation

• Suppose that genotypes at a particular SNP are significantly

associated with disease
• This may be because the SNP is associated with some other factor (a
confounder), which is associated with disease but is not in the same
causal pathway
• Possible confounders of genetic associations:
– Ethnic ancestry
– Genotyping batch, genotyping centre
– DNA quality
• Environmental exposures in the same causal pathway
– Nicotine receptors --> smoking --> lung cancer
Hung et al, Nature 452: 633 (2008) + other articles in same issue
– Alcohol dehydrogenase genes --> alcohol consumption --> throat cancer
Hashibe et al, Nature Genetics 40: 707 (2008)
 Helpful confounding: linkage disequilibrium

Linkage disequilibrium (LD) is the non-independence of alleles at

nearby markers in a population because of a lack of recombinations
between the markers

50,000 years ago: Today:

~50kb “Haplotype block”

 Direct and indirect association testing
Hirschhorn and Daly: Nature Reviews Genetics 6: 95 (2005)

Functional SNP is genotyped Functional SNP (blue) is not

and an association is found genotyped, but a number of
other SNPs (red), in LD with
the functional SNP, are
genotyped, and an
association is found for these
SNPs
LD is helpful, because not all SNPs have to be genotyped

Peʼer et al: Nature Genetics 38: 663 (2006)

 Allele counting to test for association between SNP
genotype and case / control status

GG GT TT Total
Cases r0 r1 r2 R
Controls s0 s1 s2 S
Total n0 n1 n2 N

Observed allele counts

G T Total
Cases 2r 0+r 1 r 1+2r 2 2R
Controls 2s 0+s 1 s 1+2s 2 2S
Total 2n 0+n 1 n 1+2n 2 2N
 Allele counting to test for association between SNP
genotype and case / control status

GG GT TT Total
Cases r0 r1 r2 R
Controls s0 s1 s2 S
Total n0 n1 n2 N

Observed allele counts Expected allele counts

G T Total G T
Cases 2r 0+r 1 r 1+2r 2 2R 2R(2n 0+n 1)/(2N) 2R(n 1+2n 2)/(2N)
Controls 2s 0+s 1 s 1+2s 2 2S 2S(2n0+n1)/(2N) 2S(n 1+2n 2)/(2N)
Total 2n 0+n 1 n 1+2n 2 2N
 Allele counting to test for association between SNP
genotype and case / control status

GG GT TT Total
Cases r0 r1 r2 R
Controls s0 s1 s2 S
Total n0 n1 n2 N

Observed allele counts Expected allele counts

G T Total G T
Cases 2r 0+r 1 r 1+2r 2 2R 2R(2n 0+n 1)/(2N) 2R(n 1+2n 2)/(2N)
Controls 2s 0+s 1 s 1+2s 2 2S 2S(2n0+n1)/(2N) 2S(n 1+2n 2)/(2N)
Total 2n 0+n 1 n 1+2n 2 2N

Chi-square test for independence of rows and columns (null hypothesis):

(Obs – Exp)2
∑
Exp
~ χ2 with 1 df

PLINK --assoc option Other options (e.g. dominant/recessive models)

--model
 The odds ratio: a measure of effect size

Odds of an event occurring = Pr(event occurs) / Pr(event doesnʼt occur)

= Pr(event occurs) / [1 - Pr(event occurs)]
Allele counts
G T
Cases a b
Controls c d

Consider all the G alleles in the sample, and pick one at random.
The odds that the G allele occurs in a case: a/c

Consider all the T alleles in the sample, and pick one at random.
The odds that a T allele occurs in a case: b/d

odds ratio = odds that G allele occurs in a case = a/c = a d

odds that T allele occurs in a case b/d bc
 Interpretation of the odds ratio
G T
Cases a b
Controls c d

odds ratio (OR) = odds that G allele occurs in a case = a d

odds that T allele occurs in a case bc

OR = increase in odds of being a case for each additional G allele

OR = 1: no association between genotype and disease

OR > 1: G allele increases risk of disease
OR < 1: T allele increases risk of disease

If the disease is rare (e.g. ~0.1% for MS), the odds ratio is roughly equal to
the genotype relative risk (GRR):
the increase in risk of disease conferred by each additional G allele

e.g. if OR = 1.2,
Pr(MS | TT) = 0.1% Pr(MS | GT) = 0.12% Pr(MS | GG) = 0.144%
 Logistic regression: more flexible analysis for GWA studies
• Similar to linear regression, used for binary outcomes instead of
continuous outcomes
• Let Yi be the phenotype for individual i
Yi = 0 for controls
Yi = 1 for cases
• Let Xi be the genotype of individual i at a particular SNP
TT Xi = 0
GT Xi = 1
GG Xi = 2
 Logistic regression: more flexible analysis for GWA studies
• Similar to linear regression, used for binary outcomes instead of
continuous outcomes
• Let Yi be the phenotype for individual i
Yi = 0 for controls
Yi = 1 for cases
• Let Xi be the genotype of individual i at a particular SNP
TT Xi = 0
GT Xi = 1
GG Xi = 2
• Basic logistic regression model
Let pi = E(Yi | Xi), expected value of pheno given geno
Define logit(pi) = loge[pi /(1- pi) ]
 Logistic regression: more flexible analysis for GWA studies
• Similar to linear regression, used for binary outcomes instead of
continuous outcomes
• Let Yi be the phenotype for individual i
Yi = 0 for controls
Yi = 1 for cases
• Let Xi be the genotype of individual i at a particular SNP
TT Xi = 0
GT Xi = 1
GG Xi = 2
• Basic logistic regression model
Let pi = E(Yi | Xi), expected value of pheno given geno
Define logit(pi) = loge[pi /(1- pi) ]

logit(pi) ~ β0 + β1 Xi
 Logistic regression: more flexible analysis for GWA studies
• Similar to linear regression, used for binary outcomes instead of
continuous outcomes
• Let Yi be the phenotype for individual i
Yi = 0 for controls
Yi = 1 for cases
• Let Xi be the genotype of individual i at a particular SNP
TT Xi = 0
GT Xi = 1
GG Xi = 2
• Basic logistic regression model
Let pi = E(Yi | Xi), expected value of pheno given geno
Define logit(pi) = loge[pi /(1- pi) ]

logit(pi) ~ β0 + β1 Xi
Test whether β1 differs significantly from zero:
roughly equivalent to allele counting chi-square test

Estimate of odds ratio: exp(β1)

 Logistic regression: more flexible analysis for GWA studies
• Similar to linear regression, used for binary outcomes instead of
continuous outcomes
• Let Yi be the phenotype for individual i
Yi = 0 for controls
Yi = 1 for cases
• Let Xi be the genotype of individual i at a particular SNP
TT Xi = 0
GT Xi = 1
GG Xi = 2
• Add extra terms to adjust for potential confounders: e.g. ethnicity,
genotyping batch, genotypes at other SNPs
Let pi = E(Yi | Xi,Ci, Di,…)

logit(pi) ~ β0 + β1Xi + β2Ci + β3Di +…

PLINK --logistic
 Multiple testing
• Suppose you test 500,000 SNPs for association with disease
• Expect around 500,000 x 0.05 = 25,000 to have p-value less than 0.05
• More appropriate significance threshold
p = 0.05 / 500,000 = 10-7
genome-wide significance
• In our MS GWAS we considered SNPs for follow-up if they had p-
values less than 0.001
• To detect a smaller p-value need a larger study
 The power to detect an association
• Suppose the G allele of a SNP has frequency 0.2. If each additional
G allele increases odds of disease by 1.2, and 1618 cases and 3413
controls are genotyped, what is the power (chance) of detecting an
association with significance p<0.001?

Null hypothesis: true OR=1

Observed OR from this
distribution

p=0.001

1 1.2 Odds ratio (OR)

 The power to detect an association
• Suppose the G allele of a SNP has frequency 0.2. If each additional
G allele increases odds of disease by 1.2, and 1618 cases and 3413
controls are genotyped, what is the power (chance) of detecting an
association with significance p<0.001?

Null hypothesis: true OR=1 Alternative hypothesis: true OR=1.2

Observed OR from this Observed OR from this
distribution distribution

p=0.001

1 1.2 Odds ratio (OR)

 The power to detect an association
• Suppose the G allele of a SNP has frequency 0.2. If each additional
G allele increases odds of disease by 1.2, and 1618 cases and 3413
controls are genotyped, what is the power (chance) of detecting an
association with significance p<0.001?

Null hypothesis: true OR=1 Alternative hypothesis: true OR=1.2

Observed OR from this Observed OR from this
distribution distribution

Power = blue shaded area

= 59%

p=0.001

1 1.2 Odds ratio (OR)

 Effect of increasing sample size
Observed OR tends to be closer to true OR (narrower distributions)
⇒ Null and alternative distributions become more separate
⇒ Power increases

Null hypothesis: true OR=1 Alternative hypothesis: true OR=1.2

Observed OR from this Observed OR from this
distribution distribution

Power is now greater

p=0.001

1 1.2 Odds ratio (OR)

Multiple sclerosis - degradation of myelin sheath around
nerve fibres

www.msif.org/en/about_ms/ demyelination.html
 Multiple sclerosis
• neurodegenerative disease
• autoimmune attack on myelin sheaths around nerve cells
• more females affected than males (3:1)
• average age-at-onset ~30
• ~16,000 people with MS in Australia ($2 billion p.a.)
• no cure
 Risk factors
• Epstein-Barr virus
• Exposure to infant siblings (Ponsonby et al, JAMA, 2005)
• Latitude gradient, childhood sun exposure
(van der Mei et al, Lancet, 2003)
• Only genetic risk factor known before 2007 (first GWAS):
HLA-DRB1*1501 discovered in 1972
(60% MS and 30% controls) IL7R CD58
IL2RA EVI5/RPL5
CLEC16A CD226
KIF1B
TYK2

2000 2001 2002 2003 2004 2005 2006 2007 2008 2009

Human genome project

SNP genotyping arrays
The SNP consortium
GWA studies
The International HapMap Project
 Australian and New Zealand MS GWAS

• Assemble collection of DNA samples (all states + NZ)

• Genotype 1952 MS cases from around Australia and New
Zealand with Illumina 370CNV BeadChips
(Patrick Danoy, Matt Brown, Diamantina Institute, UQ)
• Analyse GWAS data
– Quality control (Devindri Perera, Menzies)
– Impute genotypes at millions of other SNPs
(Sharon Browning, Univ of Auckland)
– Compare case genotypes with >3500 controls from the UK and US
(publicly available data)
• Replication genotyping
(Justin Rubioʼs lab, Howard Florey Institute, Univ of Melbourne)
Quality control - MS samples (PLINK)
• Start with 1952 samples
• Exclusions
– Samples with >2% of SNPs not called 70 --mind
 Genotype call rate
No call rate

Oragene saliva DNA Samples

Quality control - MS samples (PLINK)
• Start with 1952 samples
• Exclusions
– Samples with >2% of SNPs not called 70 --mind
– Suspect batch of samples 128
– Uncertain phenotype 10
– Duplicates / relatives 88 --genome
– Ancestry outliers 35
 Quality control - ethnicity

N European
 Principal components
African
analysis: EIGENSTRAT
Price et al (2006).
Nat Genet 38: 904

Japanese
Chinese  Use an independent
set of ~77,000 SNPs
--indep-pairwise

 178 outliers removed:

- 35 MS
- 143 controls
Quality control - MS samples
• Start with 1952 samples
• Exclusions
– Samples with >2% of SNPs not called 70 --mind
– Suspect batch of samples 128
– Uncertain phenotype 10
– Duplicates / relatives 88 --genome
– Ancestry outliers 35
– Sex discrepancies 3 --check-sex
• Leaves 1618 samples

Quality control - SNPs

• Start with 310,504 SNPs in both case and control datasets
• Exclude SNPs
– Not called in >5% of samples --geno
– In Hardy-Weinberg disequilibrium --hwe
--maf
– Where one allele has frequency < 1%
• Leaves 302,098 SNPs
 Choice of 5% no-call threshold
• We originally planned to use a 10% threshold, but lots of SNPs with no call
rate 5-10% showed deviations from Hardy-Weinberg equilibrium

• Closer look at SNPs with call rates between 5% and 10% suggested that
they were unreliable
 GWAS - results

Total sample = 1618 MS cases + 3413 controls

HLA
P=7 x 10-84

P=10-7

# 50
# 500
P=0.001
 Extra QC for associated SNPs: cluster plots

UK controls ANZ cases both

 The replication phase
• Selected 100 SNPs for replication genotyping

• 2,256 ANZ MS cases + 2,310 ANZ controls

• Two chromosome regions on chr 12 and chr 20 showed (almost)

genome-wide significant (p<5 x 10-7) association with MS after
combining GWAS and replication data

• SNPs in 13/53 other regions with replication p-values < 0.1:

more than expected by chance (p=0.002)
Chromosome 12 association: the downside of LD

P=4.1 x 10-6

P=0.00001

P=0.0001

rs703842

GWAS
P = 4.1 x 10-6

replication
P = 1.4 x 10-6

GWAS + rep
P = 5.4 x 10-11

Allele
frequency 0.33

Odds ratio 0.81

(protective)
Chromosome 12 association: the downside of LD

P=4.1 x 10-6

P=0.00001

P=0.0001

KIF5A SNP associated

with rheumatoid arthritis
and type 1 diabetes
Chromosome 12 association: the downside of LD

P=4.1 x 10-6

P=0.00001

P=0.0001

Logistic regression with both

SNPs in the same model:
rs10876994 p = 0.004
rs703842 p = 0.08
Chromosome 12 association: the downside of LD

P=4.1 x 10-6

P=0.00001

P=0.0001

CYP27B1: most likely

candidate??
 CYP27B1
• Cytochrome p450 gene family (drug metabolizing)
• Encodes 25-hydroxyvitamin D-1 alpha hydroxylase (1α−OHase)
• Converts 25(OH)D3 to bioactive 1,25(OH)2D3
• 1,25(OH)2D3 regulates calcium metabolism and the immune system
via vitamin D receptor (VDR)

25(OH)D3 UVB
Liver
7 dehydro-
cholesterol

Vit D

Diet

HLA-DRB1*1501

Adorini and Penna (2008)

Nat Clin Prac Rheum 4: 404-12
The chromosome 20 association

P=0.0001
rs6074022

GWAS
P = 2.5 x 10-5

replication
P = 4.6 x 10-4

GWAS + rep
P = 1.3 x 10-7

Allele
frequency 0.25

Odds ratio 1.20

(increased risk)
 CD40
• Member of TNF receptor superfamily: regulates many cell- and
antibody-mediated immune responses

• SNPs in CD40 are associated with risk of rheumatoid arthritis and

Gravesʼ disease
• Functional SNP rs1883832C>T, 1 base pair upstream of the ATG
translation initiation codon
• Allelic heterogeneity

T allele CD40 protein MS

-1 ATG C allele CD40 protein RA GD

www.hapmap.org
 Another use of logistic regression:
test for gene-gene interaction

MS risk alleles
Chr 12 = rs703842A
Chr 20 = rs6074022G

Odds
ratio

Chr20 risk

Chr12 risk

Modest evidence that each risk allele has a bigger effect in the presence of
the other risk allele (p = 0.03)
 Summary

• Case-control GWA studies have been very successful in the past

couple of years
• Linkage disequilibrium means that most, but not all, common human
genetic variation is captured by genotyping a few hundred thousand
SNPs
• Small effect sizes (e.g. OR 1.2) mean that GWA studies need to be
large, with thousands of cases and controls --> big collaborations
• Methods of statistical analysis are fairly straightforward, but care is
required to clean data
• The ultimate test of any association: replication in an independent
population
 Acknowledgments - MS GWAS
Hobart: Devindri Perera
Sydney: Graeme Stewart Adelaide: Mark Slee
Bruce Taylor David Booth
Karen Drysdale Robert Heard
Preethi Guru Jim Wiley
Brendan McMorran New Sharon Browning
Simon Foote Zealand: Brian Browning
Gold Coast: Simon Broadley
Deborah Mason
Lyn Griffiths
Melbourne: Justin Rubio Ernie Willoughby
Lotfi Tajouri
Melanie Bahlo Glynnis Clarke
Michael Pender
Helmut Butzkueven Ruth McCallum
Vicky Perreau Marilyn Merriman
Laura Johnson Brisbane: Matthew Brown Tony Merriman
Judith Field Patrick Danoy
Cathy Jensen Johanna Hadler
Ella Wilkins Karen Pryce
Caron Chapman Peter Csurshes
Mark Marriott Judith Greer
Niall Tubridy
Trevor Kilpatrick
Perth: Bill Carroll
Newcastle: Jeanette Lechner-Scott Alan Kermode
Rodney Scott
Pablo Moscato The Australian and NZ MS Genetics Consortium (2009).
Mathew Cox Nat Genet 41: 824

Funding: MS Research Australia, John T Reid Charitable Trusts,

Trish MS Research Foundation, Australian Research Council