# Estimating missing heritability for disease from genome-wide association studies.

**ABSTRACT** Genome-wide association studies are designed to discover SNPs that are associated with a complex trait. Employing strict significance thresholds when testing individual SNPs avoids false positives at the expense of increasing false negatives. Recently, we developed a method for quantitative traits that estimates the variation accounted for when fitting all SNPs simultaneously. Here we develop this method further for case-control studies. We use a linear mixed model for analysis of binary traits and transform the estimates to a liability scale by adjusting both for scale and for ascertainment of the case samples. We show by theory and simulation that the method is unbiased. We apply the method to data from the Wellcome Trust Case Control Consortium and show that a substantial proportion of variation in liability for Crohn disease, bipolar disorder, and type I diabetes is tagged by common SNPs.

**0**Bookmarks

**·**

**141**Views

- Yi Lu, Weronica E Ek, David Whiteman, Thomas L Vaughan, Amanda B Spurdle, Douglas F Easton, Paul D Pharoah, Deborah J Thompson, Alison M Dunning, Nicholas K Hayward, Georgia, Stuart Macgregor[Show abstract] [Hide abstract]

**ABSTRACT:**Common cancers have been demarcated into 'hereditary' or 'sporadic' ('non-hereditary') types historically. Such distinctions initially arose from work identifying rare, highly penetrant germline mutations causing 'heredi-tary' cancer. While rare mutations are important in particular families, most cases in the general population are 'sporadic'. Twin studies have suggested that many 'sporadic' cancers show little or no heritability. To quantify the role of germline mutations in cancer susceptibility, we applied a method for estimating the importance of common genetic variants (array heritability, h 2 g) to twelve cancer types. The following cancers showed a signifi-cant (P < 0.05) array heritability: melanoma USA set h 2 g 5 0.19 (95% CI 5 0.01 –0.37) and Australian set h 2 g 5 0.30 (0.10– 0.50); pancreatic h 2 g 5 0.18 (0.06– 0.30); prostate h 2 g 5 0.81 (0.32–1); kidney h 2 g 5 0.18 (0.04 –0.32); ovarian h 2 g 5 0.30 (0.18 – 0.42); esophageal adenocarcinoma h 2 g 5 0.24 (0.14 – 0.34); esophageal squamous cell carcin-oma h 2 g 5 0.19 (0.07 – 0.31); endometrial UK set h 2 g 5 0.23 (0.01 –0.45) and Australian set h 2 g 5 0.39 (0.02 – 0.76). Three cancers showed a positive but non-significant effect: breast h 2 g 5 0.13 (0 – 0.56); gastric h 2 g 5 0.11 (0 – 0.27); lung h 2 g 5 0.10 (0 – 0.24). One cancer showed a small effect: bladder h 2 g 5 0.01 (0 –0.11). Among these can-cers, previous twin studies were only able to show heritability for prostate and breast cancer, but we can now make much stronger statements for several common cancers which emphasize the important role of genetic var-iants in cancer susceptibility. We have demonstrated that several 'sporadic' cancers have a significant inherited component. Larger genome-wide association studies in these cancers will continue to find more loci, which explain part of the remaining polygenic component. INTRODUCTION06/2014; - SourceAvailable from: Eric R GamazonJason M Torres, Eric R Gamazon, Esteban J Parra, Jennifer E Below, Adan Valladares-Salgado, Niels Wacher, Miguel Cruz, Craig L Hanis, Nancy J Cox[Show abstract] [Hide abstract]

**ABSTRACT:**Top signals from genome-wide association studies (GWASs) of type 2 diabetes (T2D) are enriched with expression quantitative trait loci (eQTLs) identified in skeletal muscle and adipose tissue. We therefore hypothesized that such eQTLs might account for a dispropor-tionate share of the heritability estimated from all SNPs interrogated through GWASs. To test this hypothesis, we applied linear mixed models to the Wellcome Trust Case Control Consortium (WTCCC) T2D data set and to data sets representing Mexican Americans from Starr County, TX, and Mexicans from Mexico City. We estimated the proportion of phenotypic variance attributable to the additive effect of all variants interrogated in these GWASs, as well as a much smaller set of variants identified as eQTLs in human adipose tissue, skeletal muscle, and lymphoblastoid cell lines. The narrow-sense heritability explained by all interrogated SNPs in each of these data sets was substantially greater than the heritability accounted for by genome-wide-significant SNPs (~10%); GWAS SNPs explained over 50% of phenotypic variance in the WTCCC, Starr County, and Mexico City data sets. The estimate of heritability attributable to cross-tissue eQTLs was greater in the WTCCC data set and among lean Hispanics, whereas adipose eQTLs significantly explained heritability among Hispanics with a body mass index R 30. These results support an important role for regulatory variants in the genetic component of T2D susceptibility, particularly for eQTLs that elicit effects across insulin-responsive peripheral tissues.The American Journal of Human Genetics 10/2014; · 11.20 Impact Factor - SourceAvailable from: Weronica Emma EkWeronica E Ek, David M Levine, Maurod ' Amato, Nancy L Pedersen, Patrik K E Magnusson, Lynn E Onstad, Peter T Schmidt, Hanstörnblom, Yvonneromero Helenanordenstedt, Liam J Murray, [......], Nicholas J Risch, Nigel C Shaheen, Brian J Bird, Douglas A Reid, Laura J Corley, Weiminye Hardie, Anna H Wu, Tim D Spector, Thomas L Vaughan, David C Whiteman10/2013;

Page 1

ARTICLE

Estimating Missing Heritability for Disease

from Genome-wide Association Studies

Sang Hong Lee,1Naomi R. Wray,1Michael E. Goddard,2,3and Peter M. Visscher1,*

Genome-wide association studies are designed to discover SNPs that are associated with a complex trait. Employing strict significance

thresholds when testing individual SNPs avoids false positives at the expense of increasing false negatives. Recently, we developed

a method for quantitative traits that estimates the variation accounted for when fitting all SNPs simultaneously. Here we develop

this method further for case-control studies. We use a linear mixed model for analysis of binary traits and transform the estimates to

a liabilityscale by adjustingboth for scale and for ascertainment of the case samples. We show by theory and simulationthat the method

is unbiased. We apply the method to data from the Wellcome Trust Case Control Consortium and show that a substantial proportion of

variation in liability for Crohn disease, bipolar disorder, and type I diabetes is tagged by common SNPs.

Introduction

Heritability is a general and key population parameter that

can help understand the genetic architecture of complex

traits. It is usually defined as the proportion of total pheno-

typic variation that is due to additive genetic factors.1

Methods of obtaining unbiased estimates of heritability

from pedigree data are well established for continuous

phenotypes, for example (restricted) maximum likelihood

for linear mixed models (LMM).2–5For binary traits, such

as disease, familial resemblance is usually parameterized

on an unobserved continuous liability scale so that the

heritability is independent of disease prevalence.6With

genome-wide genotype data, we can derive estimates of

genetic variance tagged by the SNPs from samples of indi-

vidualswho are unrelatedin theconventional sense.7Heri-

tability estimated from pedigree data is not the same as the

proportion of phenotypic variation explained by all SNPs

because the former includes the contribution of all causal

variants, whereas the latter only includes the contribution

of causal variants that are in linkage disequilibrium (LD)

with the genotyped SNPs.8

Genome-wide association studies (GWAS) have reported

hundreds of SNPs that are robustly associated with one or

more complex traits, including quantitative traits and

common disease.9Typically, the associated SNPs in total

only explain a small proportion of the genetic variation

in the population, and this observation has led to the

perceived problem of ‘‘missing heritability.’’10,11We have

argued previously that the two most plausible explana-

tions for these observations are that either the effect sizes

at individual SNPs are so small that they do not reach

genome-wide significance in GWAS or that causal variants

are not in sufficient LD with SNPs on the commercial

arrays to be detected by association.7,12For example, insuf-

ficient LD could arise if causal variants have lower minor

allele frequency (MAF) than genotyped SNPs. To test these

hypotheses, we recently developed a method to estimate

the proportion of variance explained by all SNPs in

GWAS for a quantitative trait.7We showed that a substan-

tial proportion of genetic variation for human height was

associated with common SNPs. For complex diseases it

would be very useful to apply the same estimation proce-

dure to case-control GWAS data. However, there are three

issues that need to be overcome to be able to estimate

genetic variance for disease without bias and with compu-

tationally fast algorithms:

(1) Scale. For quantitative traits the scale of measure-

ment is the same as the scale on which heritability

is expressed. For disease traits, the phenotypes

(case-control status) are measured on the 0–1 scale,

but heritability is most interpretable on a scale of

liability.

(2) Ascertainment. In case-control studies the propor-

tion of cases is usually (much) larger than the prev-

alence in the population yet estimates of genetic

variation are most interpretable if they are not

biased by this ascertainment.

(3) Quality control (QC) of SNPs. QC is more of a

concern for case-control than quantitative GWAS.

For quantitative traits, experimental or genotyping

artifacts are unlikely to be correlated with the trait

value. However, case and control sets are often

collected independently so that experimental arti-

facts could make cases more similar to other cases

and controls more similar to other controls. These

artificial case-control differences could be parti-

tioned as ‘‘heritability’’ in methods that utilize

genome-wide similarity within and differences

between cases and controls.

In the present study, we overcome all three problems

and by using theory, simulations, and analysis of real

1Queensland Institute of Medical Research, 300 Herston Rd, Herston, Queensland 4006, Australia;2Biosciences Research Division, Department of Primary

Industries,Melbourne,Victoria3086,Australia;3DepartmentofAgricultureandFoodSystems,UniversityofMelbourne,Melbourne,Victoria3010,Australia

*Correspondence: peter.visscher@qimr.edu.au

DOI 10.1016/j.ajhg.2011.02.002. ?2011 by The American Society of Human Genetics. All rights reserved.

294

The American Journal of Human Genetics 88, 294–305, March 11, 2011

Page 2

data show that genetic variation in liability to disease that

is in LD with common SNPs can be estimated from GWAS

data. The purpose of this paper is to present these methods

in detail. We demonstrate their application by using three

of the data sets from the Wellcome Trust Case Control

Consortium (WTCCC),13Crohn disease, bipolar disorder,

and type I diabetes. We show that a substantial proportion

of variation in liability to these diseases is captured by

common SNPs.

Material and Methods

We first revisit the concept of using marker data to estimate real-

ized genetic relationships. Next we present the linear model

between binary phenotype and genetic effects (this is the same

model as used for continuous phenotypes) and use this model to

estimate genetic variance. We then demonstrate the derivation

of the classic relationship between additive genetic variance on

the disease and liability scales that allows interpretation of genetic

variance on the liability scale. For case-control studies, we adapt

this relationship to account for ascertainment that generates

a much higher proportion of cases in our analyzed sample

than in the population. The new theory is general and applicable

to any proportion of cases and controls in a case-control study. We

apply these methods first to simulated data in which we can vary

thedisease prevalence and geneticvariance explainedby theSNPs.

We then apply these methods to real GWAS data by focusing on

stringent QC steps required to get meaningful results.

Theory on Random Variables

Throughout subsequent derivations, we repeatedly make use of

a number of known results from statistical theory. For random

variables x and y, their variances and covariance are defined as

varðxÞ ¼ Eðx2Þ ? EðxÞ2;

varðyÞ ¼ Eðy2Þ ? EðyÞ2

(Equation 1)

and

covðx;yÞ ¼ EðxyÞ ? EðxÞEðyÞ:

(Equation 2)

The bivariate regression of y on x has regression coefficient

b ¼ covðx;yÞ=varðxÞ:

(Equation 3)

If y follows a standard normal distribution with a truncation

point at t, with t > 0, so that the fraction of y that is larger than

t is K, then the mean value of y above the truncation point is

Eðyjy > tÞ ¼ i ¼ z=K;

(Equation 4)

with z the height of the normal curve at point t.2,3The mean for y

below the truncation point is

Eðyjy < tÞ ¼ ?iK=ð1 ? KÞ:

(Equation 5)

The variance of y for values above and below the truncation

point are [1 ? i(i-t)] and [1 ? itK/(1-K)(t þ iK/(1 ? K)], respectively.2

It follows from the definition of the variance given above that

E?y2jy > t?¼ ½1 þ it?

(Equation 6)

and

E?y2jy < t?¼ ½1 ? itK=ð1 ? KÞ?:

(Equation 7)

Realized Relationships between Distant Individuals

We showed previously that it is possible to estimate realized rela-

tionships between unrelated (in a conventional sense) individuals

from dense SNP data.7A simple and logical method of estimating

realized additive genetic relationships (bAij) between individual i

between two individuals scaled by the heterozygosity for all L

genotyped SNPs across the genome,7

and j is to use the products of genotype indicator coefficients

bAij¼1

L

X

L

i¼1

?xil? 2pl

?,?xjl? 2pl

???2plql

?

ðisjÞ;

(Equation 8)

where xil¼ 0, 1, or 2 according to whether individual i has geno-

typebb,Bb, or BBat locusl (allelesarearbitrarilycalledbor B), p (q)

is allele frequency of B (b), and 2p is the mean of xl. As in Yang

et al.,7we use the current population as the base (reference) pop-

ulation when estimating relatedness from SNP data so that

E(x) ¼ 2p in the current population. This implies that the average

pairwise relatedness is zero and that some pairs of individuals

are less related to each other than the average in the population,

leading to negative estimates. Relatedness in this definition is

not a probability (as in the classical definition of identity-by-

descent) but a correlation of additive genetic values.2,14The esti-

mate of relatedness for an individual with him/herself (the

diagonal of the matrix) has a slightly different form to the off-

diagonals to minimize sampling variation,7

bAii¼ 1 þ1

L

X

L

i¼1

?x2

il??1 þ 2pl

?xilþ 2p2

l

???2plql

?:

Linear Mixed Model

In a model for analyzing disease, the observations (unaffected or

affected) can be expressed as a linear function of the sum of the

additive effects due to SNPs associated with causal variants and

residual effects. The linear model can be written as

y ¼ m1Nþ u þ e

(Equation 9)

where y is a vector of 0, 1 observations of disease status for N indi-

viduals, m is the overall mean, 1Nis a vector of N ones, u is a vector

of random additive genetic effects from aggregate SNP informa-

tion,and e is avector ofresiduals.The variancestructureof pheno-

typic observations is written as V ¼ As2

realized relationship matrix estimated from SNP data, I is an iden-

tity matrix, s2

uis polygenic additive genetic variance explained by

the SNPs, and s2

eis error variance; these variances are on the

observed 0–1 scale. Therefore, the heritability on the observed

scale is h2

eÞ, the ratio of total phenotypic variance

on that scale that is due to additive genetic effects. The variance

components are estimated via residual maximum likelihood

(REML) analysis.4,15,16

uþ Is2

e, where A is the

o¼ s2

u=ðs2

uþ s2

Liability Threshold Model

One can model the relationship between observations on the

observed risk scale and liabilities on the unobserved continuous

scale by using a probit transformation to generate the classical

liabilitythreshold model6(Figure 1). Liabilityof disease is assumed

to be the sum of environmental and additive genetic components

from independent normal distributions. The advantages of

working on the scale of liability are that population parameters

such as variance components and heritability are independent

of prevalence and can therefore be compared across traits or

The American Journal of Human Genetics 88, 294–305, March 11, 2011

295

Page 3

populations and that statistical methods developed for quantita-

tive traits can be applied to the trait liability.2,6The model can

be written as

l ¼ m1Nþ g þ e

(Equation 10)

where l is a vector of liability phenotypes that are distributed as

N(0, 1), g is a vector of random additive genetic effects on the

liability scale that are distributed N(0, s2

the same as in the linear model but on the liability scale. We

note that the mean of the distribution of liability is zero (m ¼ 0)

when there is no ascertainment. Therefore, because the total

phenotypic variance on the scale of liability is per definition equal

to 1 and the heritability is defined as the genetic variance as

a proportion of total variance, the heritability on the liability scale

is h2

g. In the liability threshold model, all affected individuals

have liability phenotypes exceeding a certain threshold value t

(Figure 1). The population prevalence is K ¼ E(y). Applying the

properties of truncated normal distributions2,17, Equations 4 and

5 give the mean liability

g), and other terms are

l¼ s2

i ¼ Eðljy ¼ 1Þ ¼ z=K for cases and(Equation 11)

i2¼ Eðljy ¼ 0Þ ¼ ?z=ð1 ? KÞ ¼ ?iK=ð1 ? KÞ for controls;

(Equation 12)

and, Equations 6 and 7 then give the squared mean liability as

E?l2jy ¼ 1?¼ 1 þ it for cases and

E?l2jy ¼ 0?¼ 1 þ i2t for controls:

By using Equations 2, 11, and 12, we can derive the covariance

between y (unaffected/affected status) and l (liability) as

(Equation 13)

(Equation 14)

covðy;lÞ ¼ Eðy,lÞ ? EðyÞEðlÞ ¼ K1i þ ð1 ? KÞ0i2¼ Ki ¼ z;

where z is the height of the standard normal probability density

function at the truncation threshold t. The above derivations

describe the relationship between the phenotypes on the two

scales, but what we are interested in is the relationship between

genetic values on those scales. Following Dempster and Lerner,18

we determine the genetic value on the observed 0–1 risk scale

for an individual (u), defined in Equation 9, as

u ¼ c þ bg ¼ c þ zg;

(Equation 15)

where c is a constant.

The linear regression coefficient that links the two scales is

derived from the regression of the phenotype on the observed

scale (y) on the additive genetic effect on the scale of liability (g),

and equals the covariance of y and g divided by the variance of g

(Equation 3),

b ¼ covðy;gÞ=s2

g¼ ½Eðy,gÞ ? EðyÞEðgÞ?=h2

l¼ Kih2

l=h2

(Equation 16)

l¼ z:

Finally, the heritability on the observed scale is the genetic vari-

ance on the observed scale, s2

as a proportion of the total variance of 0–1 observations, which

is the Bernoulli distribution variance K(1 ? K) and can be

written as

u¼ varðzgÞ ¼ z2s2

gfrom Equation 15,

h2

o¼ s2

¼ s2

u=½Kð1 ? KÞ? ¼ s2

gb2=½Kð1 ? KÞ? ¼ h2

g

h

covðy;gÞ=s2

lz2=½Kð1 ? KÞ?:

g

i2=½Kð1 ? KÞ?

This can be rearranged to transform the heritability on the

observed scale to that on the liability scale as

h2

l¼ h2

oKð1 ? KÞ=z2:

(Equation 17)

This linear transformation was derived by Alan Robertson in the

Appendix of Dempster and Lerner.18When applied to estimates of

genetic variation on the observed scale derived from family data,

this transformation can give biased estimates on the liability scale

because the genetic variation estimable from close relatives

contains both additive and nonadditive variance.18,19However,

when the genetic variance is estimated from distant relatives,

the nonadditive genetic component of the variance is small

relative to the additive component, and so the Robertson transfor-

mation provides a good approximation. Because we are using

genetic relationships between ‘‘unrelated’’ individuals, the Robert-

son approximation is valid in samples without ascertainment.

However, in order to obtain a relationship between the estimates

of heritability on the two scales, we need to account for the

inflated proportion of cases in case-control designs.

Ascertainment-Corrected Transformation to the

Estimated Variance in a Case-Control Study

to Estimate h2

l

We consider the same liability model when the proportions of

cases and controls are not a random sample from the population

(Figure 2). The mean and variance for case and control disease

status (ycc), disease liability (lcc), and genetic liability (gcc) following

quantitative genetic theory2are

E?ycc

ðusually 1=2Þ;

?¼ P; which is the proportion of cases in the sample

Figure 1.

lence of K

An underlying continuous random variable determines disease

status. If liability exceeds the threshold t, then individuals are

affected.

The Liability Threshold Model for a Disease Preva-

296

The American Journal of Human Genetics 88, 294–305, March 11, 2011

Page 4

var?ycc

?¼ Pð1 ? PÞ; which is the phenotypic variance on the

observed scale in the case-control sample; and

EðlccÞ ¼ Pi þ ð1 ? PÞi2¼ il; where we define l ¼ ðP ? KÞ=ð1 ? KÞ:

Using Equations 1, 13, and 14 then gives,

varðlccÞ ¼ s2

lcc¼ E?l2

cc

?? EðlccÞ2¼ Pð1 þ itÞ þ ð1 ? PÞð1 þ i2tÞ ? i2l2

¼ 1 þ Pit ? ð1 ? PÞtiK=ð1 ? KÞ ? i2l2¼ 1 þ ilðt ? ilÞ ¼ 1 þ q;

(Equation 18)

where q ¼ ilðt ? ilÞ, that is, the variance of liability is greater than

1 in a case-control study because individuals from the tails of the

distribution of liability have been selected. Because cases (and

controls) are ascertained on the observed phenotypic scale, the

mean of genetic liability depends on the mean liability phenotype

of the cases and the heritability of liability,

EðgccÞ ¼ h2

lEðlccÞ ¼ h2

l½Pi þ ð1 ? PÞi2? ¼ h2

lil:

By using Equations 1, 13, 14, and 18 with the heritability of

liability, we can derive the variance for genetic liability as

varðgccÞ ¼ s2

gcc¼ E?g2

cc

?? EðgccÞ2¼ h2

lE?l2

li2l2¼ h2

cc

???h2

lEðlccÞ?2

?1 þ h2

¼ h2

l½Pð1 þ itÞ þ ð1 ? PÞð1 þ i2tÞ? ? h4

The expression for s2

of estimation of the accuracy of predicting the genetic risk of

disease from case-control studies.20As for the situation of no

ascertainment, we are interested in the regression of phenotype

on the observed risk scale on genetic liability in the case-control

study,

llq?:

gccwas previously derived in the context

bcc¼ cov?ycc;gcc

¼?h2

??varðgccÞ ¼?E?ycc,gcc

?? E?ycc

?EðgccÞ??varðgccÞ

Kð1 ? KÞ

(Equation 19)

liP ? h2

lilP??s2

gcc¼?Ph2

lið1 ? lÞ??s2

gcc¼ zPð1 ? PÞ

s2

g

s2

gcc

:

The termPð1?PÞ

ficient due to ascertainment in a regression of phenotype on the

observed risk scale onto genetic factors on the scale of liability.

In the absence of ascertainment (P ¼ K), this term is 1.

According to Equation 15, the genetic value on the observed

scale (ucc) for an individual in a case-control study is

Kð1?KÞ

s2

s2

g

gccquantifies the change of the regression coef-

ucc¼ c þ bccgcc¼ c þ zPð1 ? PÞ

Kð1 ? KÞ

s2

g

s2

gcc

gcc;

and

s2

ucc¼ b2

ccs2

gcc¼

"

zPð1 ? PÞ

Kð1 ? KÞ

s2

s2

gcc

g

#2

s2

gcc¼

?

zPð1 ? PÞ

Kð1 ? KÞ

?2s2

(Equation 20)

g

s2

gcc

s2

g:

We note that uccis a least-squares estimate of the genetic value

on the observed scale. When residuals are normally distributed,

the least-square estimate is the same as the (residual) maximum-

likelihood estimate. However, normality of liability is violated in

a case-control study. The previous section describes the theoretical

relationships between parameters on different scales in the pres-

ence of ascertainment. In practice, we do not observe parameters

directly but estimate them. We now consider the relationship

between the parameters and their estimates when maximum like-

lihood is used to estimate the variance components.

The estimated genetic variance on the observed scale from

REML analysis (Equation 9) is based on 0–1 observations and the

covariance structure among samples. Without ascertainment,

the mean of estimated genetic values on the observed scale can

be derived from Equations 11, 12, 15, and 16 as Eðbujy ¼ 1Þ ¼

are ascertained, the mean of the estimated genetic values is

zis2

gfor cases and Eðbujy ¼ 0Þ ¼ zi2s2

E?buccjycc¼ 1?

gfor controls. When samples

¼P

K

ð1 ? PÞ

ð1 ? KÞ

ð1 ? PÞ

ð1 ? KÞE?bu jy ¼ 1?

for cases

(Equation 21)

and

E?buccjycc¼ 0?

where the termP

and the decreased proportion of controls, that is, in an ascertained

case-controlstudy, themeangeneticliabilityforcasescanbetrans-

formed to that on the observed scale by bcci?s2

¼P

K

ð1 ? PÞ

ð1 ? KÞ

P

KE?bu jy ¼ 0?

for controls;

K

ð1?PÞ

ð1?KÞis due to the increased proportion of cases

gcc¼P

K

ð1?PÞ

ð1?KÞzi?s2

gand

bcci?

2s2

gcc¼P

ð1?KÞand i?

Equation 17, h2

estimated heritability on the liability scale in an ascertained case-

2

ccis a squared regression coefficient that trans-

forms the estimate of genetic factors on the observed risk scale to

that on the liability scale. This expression can be written as

K

ð1?PÞ

ð1?KÞzi?

2s2

gfrom Equations 11, 12, 15, and 19 and

i?¼ ið1?PÞ

2¼ i2

l¼ h2

P

K, which is derived as follows. According to

2

occPð1 ? PÞ=bz

oKð1 ? KÞ=z2¼bh

2

cc, wherebh

2

occis

control study andbz

s2

u

Kð1?KÞ

1

z

Kð1?KÞ

z

¼

b s

2

ucc

Pð1?PÞ

1

b zcc

Pð1?PÞ

b zcc

. But from Equations 17, 19, and 20,

the terms

s2

u

Kð1?KÞ

1

zon the left-hand side and

b s

2

ucc

Pð1?PÞ

1

b zcc

on the

Figure 2.

sampled as in a Case-Control Study

The Distribution of Liability When Cases Are Over-

The American Journal of Human Genetics 88, 294–305, March 11, 2011

297

Page 5

right-handsidearethesame.Therefore,Kð1?KÞ

expressions i?¼b zcc

above. The REML estimate of the genetic variance (bs

of the genetic values (bu orbucc) with a normality assumption, that

covðy;buÞ=bs

bs

and

z

¼Pð1?PÞ

b zcc

resultsinthe

P¼ ið1?PÞ

ð1?KÞand i?

2¼ ?b zcc

ð1?PÞ¼ i2

P

Kthat were given

2

uorbs

2

ucc) is

the covariance between observations and the unbiased estimate

is, the regression of phenotype on predictor has a slope of 1,

2

u¼ 1 or covðycc;buccÞ=bs

2

2

ucc¼ 1. Therefore,

u¼ cov?y;bu?¼ E?y,bu?? EðyÞE?bu?¼ E?y,bu?¼ E?bu jy ¼ 1?,K

(Equation 22)

bs

2

ucc¼ cov?ycc;bucc

?¼ E?ycc,bucc

?? E?ycc

?E?bucc

2

gfrom Equation 15

?¼ E?ycc,bucc

?

¼ E?buccjycc¼ 1?,P:

from Equations 21 and 22 andbs

2

u¼ z2bs

?

give us

bs

2

ucc¼

?P

K

ð1 ? PÞ

ð1 ? KÞ

?2

bs

2

u¼

zP

K

ð1 ? PÞ

ð1 ? KÞ

?2

bs

2

g:

We note that in an ascertainedcase-controlstudy, the REML esti-

mate is larger than the least-square estimator (Equation 20) by

a factor s2

the normality assumption in REML. Therefore,

gcc=s2

g, i.e.,bs

2

ucc¼ ðs2

gcc=s2

gÞs2

ucc. This difference is due to

h2

l¼ s2

g¼bs

2

ucc

?1

z

Kð1 ? KÞ

Pð1 ? PÞ

?2

¼bh

2

occ

Kð1 ? KÞ

z2

Kð1 ? KÞ

Pð1 ? PÞ:

(Equation 23)

In the absence of ascertainment (P ¼ K), this equation reduces

to Equation 17, hence adjusting for ascertainment when variance

componentsareestimated by

to a generalization of the classical Robertson transformation.

Equation 23 shows the transformation that needs to be applied

to the SNP-attributable variance estimated from Equation 9

to provide an estimate of the liability variance in the total

population explained by the SNPs. The sampling variance of

estimated heritability on the liability scale transformed from

that on the observed scale can be derived with a Taylor series

expansion,

maximumlikelihood leads

var?h2

l

??

"

d?h2

dh2

l

?

?

occ

?

#2

var

?

h2

occ

?

¼

?Kð1 ? KÞ

z2

Kð1 ? KÞ

Pð1 ? PÞ

?2

var

?

h2

occ

?

:

(Equation 24)

With extreme ascertainment, for example, K < 0.01 and P ¼ 0.5,

a high heritability on the liability scale transforms to a greater

genetic than phenotypic variance on the observed scale according

to Equation 23. This is not a problem for the estimation of the

genetic variance. When using REML or ML for estimation, it is

possible to maximize the likelihood for the genetic variance on

the observed scale even if it is larger than the observed phenotypic

variance on that scale. Therefore, we can correctly estimate the

heritability on the liability scale even with extreme K and high

heritability on the liability scale.

In summary, in practice we can estimate the variance ex-

plained on the 0–1 risk scale in a case-control design by using

a LMM and transform both the estimate and its standard error

to the scale of liability while adjusting simultaneously for ascer-

tainment.

Simulated Data

To test the estimation of genetic variance on the liability scale

from ascertained case-control data, we performed a simulation

study. For each simulation replicate we generated 5000 cases and

5000 controls. To achieve a low level of relatedness, we simulated

individuals in independent batches of 100 with genetic values (g)

drawn from a multivariate normal distribution given a 100 3

100 covariance matrix (the mvrnorm function in the R package

was used21). Elements of the covariance matrix were 0.05 s2

off-diagonals and s2

gfor diagonals. Environmental effects were

sampledfromanormaldistributionwithameanofzeroandavari-

ance of s2

echosen such that the desired heritability of

liability was obtained. As in equation (10), liability, l, for each indi-

vidual consists of genetic effects, g, and residuals, e, on the liability

scale; that is, l ¼ g þ e. Disease status for each individual was deter-

mined by comparing l with the threshold of liability determined

by the population prevalence. For example, for K ¼ 0.1, individ-

uals were assigned to be a case if l > 1.282sl. From each batch,

all cases and an equal number of randomly selected controls

contributed to the case-control sample. We continued simulating

batches of individuals until the desired sample size of 5000 cases

and 5000 controls was achieved. The pairwise relationships

between individuals in the case-control sample were 0.05 if they

both came from the same set and zero otherwise.

For the analysis of the simulated data, we used the LMM Equa-

tion 9 with the transformation that was derived in Equation 23.

Simulations were performed for K¼0.001, 0.01, 0.10, 0.20, and

0.5, and the case-control samples were generated either without

ascertainment (P ¼ K) or with ascertainment (P > K), where each

set of 100 individuals contributed approximately the same

number of cases and controls. A range of values of heritability

on the liability scale, h2

l, were tested. One hundred simulation

replicates were conducted for each scenario.

gfor

e, with s2

Real Data

We appliedourestimationmethodtoWTCCC GWASdata,13geno-

typed on the Affymetrix 5.0 platform. QC is important with real

data because artificial allele frequency differences between cases

and controls will generate a spurious ‘‘heritability.’’ As a test of the

robustness of the method, we first used the two independent

control samples and pretended they formed a case-control study.

One of the two control groups was treated as a case group, and the

other was treated as a control group in the analysis. Subsequently,

we estimated genetic variance explained by all SNPs by analyzing

casesamplesforCrohndisease,bipolardisorder,andtypeIdiabetes

alongwiththecombineddatasetofthetwocontrolsamples.Wefit

the first 20 principal components as covariates in the LMM (Equa-

tion 9) to correct for possible population structure.22

For each data set, a standard QC procedure was performed. SNPs

with MAFs <0.01 and missing rates >0.05 were excluded as were

individuals with missing rates >0.01. Because small errors for

each SNP can be accumulated to give incorrect estimates for

genetic variance, additional QC steps were extremely stringent.

We excluded SNPs whose p values were <0.05 for the Hardy-Wein-

berg (H-W) equilibrium test and for missingness-difference

between cases and controls. We also applied a two-locus test based

on the difference in the test statistic of association between single

SNPs and pairs of adjacent SNPs.23Sex chromosomes were

excluded from the analysis. To keep individuals who were only

distantly related, both individuals from a pair with an estimated

relationship >0.05 were excluded; to benchmark this threshold,

relationships approximately closer than second cousins were

298

The American Journal of Human Genetics 88, 294–305, March 11, 2011

Page 6

removed. After the stringent QC process, the number of samples

and SNPs used for estimating genetic variance were 2599 individ-

uals (1395 cases and 1204 controls) and 309,040 SNPs for the

control-control contrast study, 3833 individuals (1504 cases and

2329 controls) and 322,142 SNPs for Crohn disease, 3880 individ-

uals (1433 cases and 2447 controls) and 321,605 SNPs for bipolar

disorder, and 4063 individuals (1640 cases and 2423 controls) and

318,044 SNPs for type I diabetes.

To investigate the robustness of our variance estimates, we also

considered more stringent threshold values for SNP missing rates

(fewer than 20, 7, or 4 genotypes per SNP) and MAF (>0.05). To

benchmark these thresholds, we note that a SNP missing rate of

0.05 is approximately equal to a maximum of 200 missing geno-

types per SNP.

Each QC step was designed to remove potential artifacts from

contributing to the estimate of genetic variance. However, each

step also reduces the number of SNPs used for estimation of the

genetic variance. As the number of SNPs decreases, the LD

between genotyped and causal variants also decreases, and so esti-

mates of genetic variance are expected to decrease. To determine

whether any observed reduction in estimated genetic variance

was a consequence of the reduced number of SNPs rather than

the QC criteria per se, we adjusted the estimate of variance to

take account of imperfect LD between the genotyped SNPs. That

is, we adjusted the estimate because pairwise relatedness is esti-

mated with error. We described previously how this adjustment

was made.7Subsequent to the adjustment for using a finite

number of SNPs, we used the transformation in Equation 23 to

obtainan estimate on the liability scale that takes account of ascer-

tainment of cases in a case-control sample. In the transformation,

we assumed a population prevalence of 0.1%, 0.5%, and 0.5% for

Crohn disease, bipolar disorder, and type I diabetes, respec-

tively.13,24–26Hence, our procedure was as follows: (1) REML anal-

ysis of 0–1 data using relatedness estimated from SNP data,16(2)

adjustment for the number of SNP used to construct relationships,

and (3) adjustment for scale and ascertainment.

Results

Simulations

WeusedREMLtoestimateheritabilityontheobservedscale,

and we transformed estimates to the liability scale with the

ascertainment correction from Equation 23. When simu-

lated data with K ¼ P ¼ 0.5 (no ascertainment) were used,

the estimated heritabilities on the liability scale were unbi-

ased and close to the true values, as expected (Table 1).

When ascertained case-control studies were used, the esti-

mates were largely unbiased although a slight overestima-

tionwasobservedforanextremevalueofK¼0.001(Table1).

Estimated Genetic Variance from the WTCCC Data

after Stringent QC

In preliminary analyses, we recognized the importance of

imposing stringent QC on H-W equilibrium, on differen-

tial missingness between cases and controls, and on

a two-locus QC test (see Figures S1–S3, available online).

Genotyping conducted on other platforms might not

require this level of stringency. We conducted a range of

additional tests and checks to ensure the validity of our

results (see Discussion and Supplemental Data).

Control-Control Contrast Study

Estimates for genetic variance between the two control

groups were not significantly different from zero, as ex-

pected (Table 2). The estimate and its likelihood ratio grad-

ually decreased when the threshold for the SNP missing

rate decreased. When SNPs with an MAF > 0.01 and

missing <4 genotypes/SNP were used, the estimate was

0.06 (SE ¼ 0.11). When SNPs with an MAF > 0.05 were

used, the decreasing patterns of the estimates and their

likelihood ratios were very similar to those of SNPs with

an MAF > 0.01. The estimated values observed when

SNPs with an MAF > 0.05 were used were slightly higher

than those with an MAF > 0.01 although the difference

was small. These results suggest that our QC procedure

was stringent enough to allow robust estimates of genetic

variation, that is, the likelihood ratio was already not

significant for a SNP genotype missingness of up to 200.

Crohn Disease

We investigated the impact of SNP missingness on the esti-

mates of variance explained by SNPs for Crohn disease.

While the threshold for missingness becomes more strin-

gent, the number of SNPs reduces from ~322,000 to

~196,000 when MAF > 0.01 (Table 3). While this happens,

the raw proportion of variance estimate drops from 0.56 to

0.50. Part of this decline is due to the reduced number of

SNPs used rather than artifacts of genotype missingness.

After we adjust for the number of SNPs, the proportion of

Table 1. Simulation Results: Estimated Heritability on the Liability Scale and Empirical Standard Error over Replicates

Prevalence of Disease

in the Population (K)

Heritability of Liability

0.1 0.30.50.7 0.9

K ¼ 0.5

K ¼ 0.2

K ¼ 0.1

K ¼ 0.01

K ¼ 0.001

0.09 (0.006)0.28 (0.010)0.51 (0.013)0.70 (0.016)0.90 (0.016)

0.10 (0.007) 0.31 (0.009)0.49 (0.011)0.71 (0.012) 0.91 (0.013)

0.11 (0.007)0.30 (0.009)0.49 (0.009) 0.71 (0.012)0.89 (0.012)

0.11 (0.009)0.30 (0.011)0.49 (0.012) 0.70 (0.013) 0.90 (0.012)

0.17 (0.020)0.31 (0.020) 0.56 (0.021)0.75 (0.021) 0.94 (0.022)

In all examples the proportion of cases in the case-control sample was p ¼ 0.5. Sample size was 10,000 (5000 cases, 5000 controls) for all situations. The number of

replicates was 100.

The American Journal of Human Genetics 88, 294–305, March 11, 2011

299

Page 7

variance estimate drops from 0.64 to 0.61 but reaches

a plateau at that value (Table3, Adjusted column). Therefore

we conclude that there is no need to make the missing

threshold more stringent than 20. On the liability scale, the

heritability estimate (i.e., the variance in liability explained

by the SNPs) is 0.22 (SE ¼ 0.04), which is much higher than

that explained by genome-wide significant SNPs.27Similar

results are obtained if the SNPs with MAF >0.05 are used

(Table 3). This indicates that common SNPs (MAF > 0.05)

are in substantial LD with casual variants for Crohn disease.

Bipolar Disorder

For bipolar disorder, we found that we needed a slightly

more stringent threshold for SNP missingness than for

Crohn disease. When we excluded SNPs with an MAF <

0.01 and missing rate >200, the heritability estimate on

the liability scale was 0.4 (Table 4). The estimates gradually

decreased and became stable after we excluded SNPs

with more than seven missing genotypes. The estimated

value was 0.38 (SE ¼ 0.04). When we used SNPs with an

MAF > 0.05, the decreasing pattern of the estimate and

its likelihood was similar to that with an MAF > 0.01,

and the values were slightly lower than those with an

MAF > 0.01. When we used SNPs with missingness of <7

or <4 genotypes, we obtained a stable estimate of ~0.37

(SE~0.04) (Table 4).

Type I Diabetes

When we used SNPs with an MAF > 0.01 and missingness

of <200 genotypes, the estimates on the liability adjusted

for reduced number of SNPs was 0.32 (SE ¼ 0.04). After

excluding SNPs with a missingness of >7 genotypes, esti-

mates and likelihood ratio showed little change (Table 5).

The estimate was 0.30 (SE ¼ 0.04) for a missingness <7

genotypes and 0.31 (SE ¼ 0.04) for a missingness of <4

genotypes. When SNPs with an MAF > 0.05 were used,

estimated values were slightly lower compared to those

with an MAF > 0.01. The estimate was 0.28 (SE ¼ 0.04)

for a missingness of <7 genotypes, and 0.29 (SE ¼ 0.04)

for <4 missing genotypes (Table 5). For type I diabetes,

some SNPs on chromosome 6 had extremely significant

associations, for example, WTCCC13reported a p value

of 5.47e-134 for rs9272346 in the region of the major

histocompatibility complex (MHC). We performed an

analysis without chromosome 6 or with chromosome 6

only when we used SNPs with an MAF > 0.01 (Table 6).

We observed that the estimates substantially decreased

Table 2.

Explained by All SNPs for Two Control Samples in the WTCCC Data

Estimated Genetic Variance in the Observed Scale

Thresholda

No. SNPb

Estimatec(SE)LRd

p valuee

MAF > 0.01

200 309,0400.17 (0.11)2.29 0.07

20 297,1980.13 (0.11)1.31 0.13

7 266,5340.08 (0.11)0.59 0.22

4226,165 0.06 (0.11) 0.290.30

MAF > 0.05

200278,564 0.19 (0.11)3.420.03

20267,043 0.16 (0.10)2.330.06

7 239,6140.12 (0.10)1.33 0.12

4203,6980.09 (0.10)0.92 0.17

aExcluding SNPs with more than the listed number of missing genotypes. Two

hundred missing genotypes are approximately equal to a missingness rate of

5% (depending on sample size).

bAfter filtering on the basis of SNP missing rate.

cEstimate of genetic variance proportional to the total phenotypic variance on

the observed scale.

dLikelihood-ratio test statistic.

ep values were calculated assuming that the LR is distributed as a 50:50

mixture of zero and c2

1under the null hypothesis.

Table 3.Estimated Genetic Variance on the Observed and Liability Scale Explained by All SNPs for Crohn Disease in WTCCC Data

Thresholda

No. SNPb

Estimatec(SE)LR Adjustedd(SE) Transformede(SE)

MAF > 0.01

200 322,1420.56 (0.07)63.160.64 (0.08)0.24 (0.03)

20294,850 0.53 (0.07)57.480.61 (0.08)0.22 (0.03)

7 248,7910.52 (0.07)57.300.61 (0.08) 0.22 (0.03)

4 195,9770.50 (0.07) 54.940.60 (0.08)0.22 (0.03)

MAF > 0.05

200293,2690.56 (0.07)69.00 0.63 (0.08)0.23 (0.03)

20 266,8430.53 (0.07)63.27 0.60 (0.08)0.22 (0.03)

7225,043 0.52 (0.07)63.940.60 (0.08)0.22 (0.03)

4 177,6150.50 (0.07) 62.140.60 (0.08)0.22 (0.03)

aExcluding SNPs with more than the listed number of missing genotypes.

bAfter filtering on the basis of SNP missing rate.

cEstimate of genetic variance proportional to the total phenotypic variance on the observed scale.

dEstimate adjusted for reduced number of SNPs.

eTransformed genetic variance proportional to the total phenotypic variance on the liability scale under the assumption that the population prevalence is 0.1%,

the heritability on the liability scale explained by the SNPs.

300

The American Journal of Human Genetics 88, 294–305, March 11, 2011

Page 8

when we excluded chromosome 6 from the analysis; that

is, it decreased to 0.13 (SE ¼ 0.04). On the other hand,

the estimate based on SNPs on chromosome 6 was rela-

tively high, that is, 0.19 (SE ¼ 0.01) (Table 6). Hence,

although the known risk locus on chromosome 6 explains

a substantial proportion of variation in liability to type I

diabetes, common SNPs on other chromosomes explain

a substantial additional proportion of variation.

Discussion

In this study, we have provided a computationally fast

methodofestimatingtheproportionofvariationindisease

liability that is captured in GWAS by considering all SNPs

simultaneously. Compared to previous analyses on quanti-

tative traits, we needed three improvements: (1) a suitable

transformation from the 0–1 risk scale to an underlying

scale of liability, (2) a proper adjustment to take account

of the fact that case-control proportions are not the

same as the proportion of cases and controls in the popula-

tion, and (3) a calibration of SNP and sample QC to

avoid spurious case-control differences in relatedness. We

demonstrated by simulation that the LMM implementa-

tion gives unbiased estimates and applied the method to

WTCCC data. We showed that a substantial proportion of

disease liability is tagged by common SNPs for Crohn

disease,bipolardisorder,andtype1diabetes.Toimplement

these methods, we have created a user-friendly software

Table 4.Estimated Genetic Variance on the Observed and Liability Scale Explained by All SNPs for Bipolar Disorder in WTCCC Data

Thresholda

No. SNPb

Estimatec(SE)LRAdjustedd(SE)Transformede(SE)

MAF > 0.01

200 3216050.71 (0.07)107.760.81 (0.08) 0.41 (0.04)

20291724 0.68 (0.07)100.480.78 (0.08) 0.40 (0.04)

7245127 0.65 (0.07) 94.69 0.76 (0.08)0.38 (0.04)

4 187597 0.62 (0.07)92.210.76 (0.08) 0.38 (0.04)

MAF > 0.05

200292969 0.68 (0.07) 110.450.77 (0.08)0.39 (0.04)

20 264151 0.65 (0.07)103.460.75 (0.08)0.38 (0.04)

7 2219470.62 (0.07)97.640.72 (0.08)0.37 (0.04)

4 1701430.60 (0.06) 95.47 0.73 (0.08)0.37 (0.04)

aExcluding SNPs with more than the listed number of missing genotypes.

bAfter filtering on the basis of SNP missing rate.

cEstimate of genetic variance proportional to the total phenotypic variance on the observed scale.

dEstimate adjusted for reduced number of SNPs.

eTransformed genetic variance proportional to the total phenotypic variance on the liability scale under the assumption that the population prevalence is 0.5%.

Table 5.Estimated Genetic Variance on the Observed and Liability Scale Explained by All SNPs for Type I Diabetes in WTCCC Data

Thresholda

No. SNPb

Estimatec(SE)LRAdjustedd(SE)Transformede(SE)

MAF > 0.01

200318,044 0.57 (0.07)70.36 0.65 (0.08)0.32 (0.04)

20289,463 0.56 (0.07)70.320.65 (0.08) 0.32 (0.04)

7238,8050.52 (0.07)61.510.61 (0.08) 0.30 (0.04)

4 178,8920.51 (0.07)64.740.64 (0.08) 0.31 (0.04)

MAF > 0.05

200 289,693 0.54 (0.07) 70.480.61 (0.08) 0.30 (0.04)

20262,0910.53 (0.07) 70.490.61 (0.08)0.30 (0.04)

7216,136 0.49 (0.06)61.81 0.57 (0.08)0.28 (0.04)

4 162,1620.48 (0.06) 63.540.58 (0.08) 0.29 (0.04)

aExcluding SNPs with more than the listed number of missing genotypes.

bAfter filtering on the basis of SNP missing rate.

cEstimate of genetic variance proportional to the total phenotypic variance on the observed scale.

dEstimate adjusted for reduced number of SNPs.

eTransformed genetic variance proportional to the total phenotypic variance on the liability scale under the assumption that the population prevalence is 0.5%.

The American Journal of Human Genetics 88, 294–305, March 11, 2011

301

Page 9

tool that is called Genome-wide Complex Trait Analysis

(GCTA) and that is available from the our website.28

The estimation of the variance explained by all SNPs is

important because it tells us how much genetic variation

is in linkage disequilibrium with SNPs on commercial

arrays. It has direct impact on further experimental design,

for example on the decision whether to invest in ever

larger GWAS samples or whether to sequence a smaller

number of samples. The difference between the proportion

of disease liability variation accounted for by robustly asso-

ciated SNPs, as in standard GWAS analysis, and our

method is that wefocus on estimationrather than on strin-

gent hypothesis testing. Causal variants that arein LD with

common SNPs but have small effect sizes are not detected

by GWAS but do contribute to genetic variation in our

method.Ourmethodcannot

whether the variance detected represents LD with causal

variants that are common or rare. However, it is very

unlikely that the detected variation represents only rare

variants, and the results provide further evidence that

common variants contribute to the genetic architecture

of complex genetic disease. Our results suggest that ever

larger GWAS samples will continue to identify robustly

associated variants that could reflect both common and

rare causal variants. The robustly associated variants will

continue to provide information about the underlying

biology, and construction of genomic profiles from

genome-wide SNPs can be useful in genetic-risk predic-

tion.29With increasing sample size, the proportion of vari-

ance explained in validation samples by using genomic

profiles developed from discovery samples will approach

the proportion of variance that we have estimated to be

tagged by all of the SNPs.

differentiate between

In order to estimate the proportion ofvariance explained

by SNPs on the liability scale, we needed to derive a trans-

formation from the variance estimated on the observed

scale, accounting for the ascertainment typical of case-

control studies. Without ascertainment, the range of heri-

tability on the observed scale is smaller compared to

that on the liability scale (see Figure 2 of Dempster and

Lerner18). However, because there is a much higher propor-

tion of cases in case-control studies than in the general

population, the range on the observed scale is larger than

that on the liability scale. For example, when K ¼ 0.01

and P ¼ 0.5, proportions of variance estimated on the

observed scale of 0.18, 0.54, and 0.91 correspond to a heri-

tability on the liability scale of 0.1, 0.3, and 0.5. Therefore,

a large change of heritability on the observed scale

becomes relatively small on the liability scale, particularly

for an extreme ascertainment. This is why the standard

errors for values on the liability scale were small relative

to those on the observed scale when the WTCCC data

were used.

We used a linear model for estimation of the variance

attributable to SNPs (Equation 9). Nonlinear models

might be considered a reasonable and appropriate alterna-

tive. However, generalized LMMs (GLMMs) that use

maximum likelihood for estimation and approximations

to avoid numerical integration30and that have been

widely used for binary traits have a problem of serious

bias induced by the approximations.31In addition, these

methods do not take account of ascertainment typical of

case-control studies. We explored GLMM by using Markov

Chain Monte Carlo sampling methods and observed that

although it gives unbiased estimates in the absence of

ascertainment, estimates were biased when samples were

Table 6.

Chromosome 6 or of Chromosome 6 Only

Estimated Genetic Variance on the Observed and Liability Scale Explained by All SNPs for Type I Diabetes from an Analysis without

Thresholda

No. SNPb

Estimatec(SE)LRAdjustedd(SE) Transformede(SE)

Analysis without chromosome 6

200 297,0280.23 (0.07) 11.98 0.26 (0.08)0.13 (0.04)

20270,3320.22 (0.07)10.660.25 (0.08)0.12 (0.04)

7223,0390.20 (0.07) 9.080.23 (0.08)0.12 (0.04)

4 167,099 0.20 (0.06)10.17 0.26 (0.08) 0.13 (0.04)

Analysis of chromosome 6 only

20021,016 0.33 (0.02)268.55 0.37 (0.03)0.18 (0.01)

2019,131 0.33 (0.02)278.090.37 (0.03) 0.18 (0.01)

715,7660.32 (0.02) 255.650.36 (0.03)0.18 (0.01)

411,7930.31 (0.02)264.630.38 (0.03)0.19 (0.01)

aExcluding SNPs with more than the listed number of missing genotypes.

bAfter filtering on the basis of SNP missing rate.

cEstimate of genetic variance proportional to the total phenotypic variance on the observed scale.

dEstimate adjusted for reduced number of SNPs.

eTransformed genetic variance proportional to the total phenotypic variance on the liability scale assuming that the population prevalence is 0.5%. SNPs with an

MAF > 0.01 were used.

302

The American Journal of Human Genetics 88, 294–305, March 11, 2011

Page 10

ascertained (results not shown). In addition to the problem

of bias, GLMM methods are computationally much slower

than LMMs.

Stringent QC is important for the analyses we have per-

formed because artificial allele frequency differences

between cases and controls will result in apparent genetic

variance. We explicitly checked data quality with various

tests, aiming to avoid spurious results. We applied a very

stringent threshold (p value < 0.05) for the H-W equilib-

rium test because more SNPs showed weak departures

from equilibrium than expected by chance (Figure S1).

There were SNPs whose missing rate was significantly

different between cases and controls. These SNPs could

be problematic because of artifact effects and would influ-

ence estimation of genetic variance. We excluded SNPs

with p value < 0.05 for differential missingness. After

excluding these SNPs, we assessed data quality by test

statistics from a comparison of single and pairwise SNP

analyses by using a two-locus QC test23(Figure S2).

Although a large number of problematic SNPs and erro-

neous signals had gone after filtering out SNPs whose

differential missingness was significant, there were still

a number of potentially problematic SNPs (e.g., 963 data

points deviating from expectation in Figure S2). We subse-

quently applied more stringent QC allowing only 20 or

only four missing genotypes across all samples, i.e., a SNP

missingness rate of <20/N (¼ ~0.005) and 4/N (¼ ~0.001),

and we showed how the erroneous signals changed

(Figure S3), where N is the total sample size. Therefore, it

islikelythat onlyhigh-qualitySNPsareretained forpheno-

type-genotype analysis. We visually checked the distribu-

tions of diagonal and off-diagonal elements from the

estimated relationship matrices. We generated histograms

of the distributions of the diagonal and off-diagonal

(Equation 8) elements of the estimated realized additive

genetic relationship matrix (Figures S4–S9). The distribu-

tions of the control-control and case-case off-diagonals are

centered slightly higher than the distribution of the case-

control off-diagonals. The means and standard deviations

are presented in Tables S1 and S2. Additionally, we per-

formed a number of analyses to make sure that there was

no estimation bias due to artifacts. Heterogeneity for two

independent sets of case-control studies was tested (Table

S3). Haseman-Elston regression for the control-control

contrast study was performed (Table S4). The original

WTCCC study13reported that ‘‘selected samples were

normalized to 50 ng ml21 and rearrayed robotically into

96-well plates so that each plate was composed of 94

samples representing at least two different collections at

a ratio of 1:1. For each collection, the selected samples

were balanced first for sex and then geographical region.’’

Given this statement, age (at which the participants

entered a study) might be more vulnerable to be associated

with systematic artifact bias due to batch or plate effects

than sex or geographical region. After removing problem-

atic SNPs, we hypothesized that individual relationships

within an age group should not be more related than those

across the rest of age groups. This was tested by Haseman-

Elston regression in which a case-control study where the

individuals from one age group were treated as cases

and the other individuals were treated as controls. (Tables

S5–S8). A case-case contrast study was carried out to check

whether genotypes from cases with different diseases

were too similar to each other (Table S9). In bivariate

analyses, we showed that the genetic correlations between

the three diseases were not significantly different from

zero (Table S10). From the test results, we concluded

that there were no apparent artifacts that were confounded

with genetic effects. However, ultimate confirmation

will come from replication analyses in other independent

data sets.

Our estimate of the proportion of variation in liability

that is tagged by all SNPs relies on knowledge of the

population prevalence (K), just as it does when one esti-

mates total heritability of liability from pedigree or twin

analyses by using binary traits. What is the effect of mis-

specifying this population parameter? We derived the

ratio of bias in the estimate of the total variance of

liability explained up to a two-fold misspecification of

disease prevalence, that is,bK ¼ 0.5K, 0.75K, 1.5K, or 2K

the ratio of bias was small, at 0.91–1.14, for all values of

K; a value of 1.0 indicates no bias. ForbK ¼ 0.5K or 2K,

K ¼ 0.1 was largest (0.81~1.24). Therefore, misspecifying

the population prevalence by a factor of two results in an

upward ordownward bias of theestimate ofthe proportion

of variance in liability explained by all SNPs of approxi-

mately 20%.

In conclusion, we have developed the methodology

needed to estimate the proportion of variance explained

on the liability scale in the population by sets of SNPs

on the basis of observations in ascertained samples of

cases and controls. We have tested our methodology by

simulation and by application to real GWAS data for

three diseases and have implemented our methodology

into freely available software.28Stringent QC of GWAS

data is necessary to prevent inflated estimates of herita-

bilityattributable toartifactual differences

case and control genotypes. Using genotypes from Affy-

metrix 5.0, we estimate that for Crohn disease, bipolar

disorder, and type I diabetes, genotyped SNPs tag

between a quarter and one half of the heritability esti-

mated from family studies. Our estimates provide an

upper limit on the variance that can be explained in

genomic profiling as sample sizes increase when the

same genotyping platform is used. Genotyping platforms

with more SNPs are expected to tag more of the genetic

variance. We show that a good proportion of the herita-

bility is not missing. The variance explained by the SNPs

is likely to tag both common and rare causal variants. We

anticipate that a proportion of the heritability will always

remain missing, reflecting rare causal variants of small

effect.

(Table S11). For a misspecification ofbK ¼ 0.75K or 1.5K,

the ratio of bias increased, and the range for the value

between

The American Journal of Human Genetics 88, 294–305, March 11, 2011

303

Page 11

Supplemental Data

Supplemental Data include nine figures and 11 tables and can be

found with this article online at http://www.cell.com/AJHG.

Acknowledgments

We acknowledge funding from the Australian National Health

and Medical Research Council (Grants 389892, 442915, 496688,

613672, and 613601) and the Australian Research Council

(Grants DP0770096 and DP1093900 and Future Fellowship to

N.R.W.). This study makes use of data generated by the Wellcome

Trust Case-Control Consortium. A full list of the investigators

who contributed to the generation of the WTCCC data is avail-

able from www.wtccc.org.uk. Funding for the WTCCC project

was provided by the Wellcome Trust under award 076113.

We thank Stuart Macgregor for useful discussions and sugges-

tions. S.H.L. acknowledges the use of the Genetic Cluster

Computer for carrying out simulations. The cluster is financially

supported by the Netherlands Scientific Organization (NWO

480-05-003). We thank the referees for many helpful comments

and suggestions.

Received: September 30, 2010

Revised: December 10, 2010

Accepted: February 1, 2011

Published online: March 3, 2011

Web Resources

The URLs for data presented herein are as follows:

Genome-wide Complex Trait Analysis (GCTA), http://gump.qimr.

edu.au/gcta

National Human Genome Research Institute, www.genome.gov/

26525384

References

1. Visscher, P.M., Hill, W.G., and Wray, N.R. (2008). Heritability

in the genomics era—concepts and misconceptions. Nat.

Rev. Genet. 9, 255–266.

2. Falconer, D.S., and Mackay, T.F.C. (1996). Introduction to

Quantitative Genetics (Harlow, UK: Longman).

3. Lynch, M., and Walsh, B. (1998). Genetics and Analysis of

Quantitative Traits (Sunderland: Sinauer Associates).

4. Patterson, H.D., and Thompson, R. (1971). Recovery of inter-

block information when block sizes are unequal. Biometrika

58, 545–554.

5. Henderson, C.R. (1984). Applications of Linear Models in

Animal Breeding (Guelph, Canada: University of Guelph).

6. Falconer, D.S. (1965). The inheritance of liability to certain

diseases, estimated from the incidence among relatives. Ann.

Hum. Genet. 29, 51–71.

7. Yang, J., Benyamin, B., McEvoy, B.P., Gordon, S., Henders,

A.K., Nyholt, D.R., Madden, P.A., Heath, A.C., Martin, N.G.,

Montgomery, G.W., et al. (2010). Common SNPs explain

a large proportion of the heritability for human height. Nat.

Genet. 42, 565–569.

8. Visscher, P.M., Yang, J., and Goddard, M.E. (2010). A commen-

tary on ‘common SNPs explain a large proportion of the

heritability for human height’ by Yang et al. (2010). Twin

Res. Hum. Genet. 13, 517–524.

9. Manolio,T.A.(2010). Genomewide

and assessment of the risk of disease. N. Engl. J. Med. 363,

166–176.

10. Maher, B. (2008). Personal genomes: The case of the missing

heritability. Nature 456, 18–21.

11. Manolio, T.A., Collins, F.S., Cox, N.J., Goldstein, D.B.,

Hindorff, L.A., Hunter, D.J., McCarthy, M.I., Ramos, E.M.,

Cardon, L.R., Chakravarti, A., et al. (2009). Finding the

missingheritabilityofcomplexdiseases.Nature461, 747–753.

12. Goddard, M.E., Wray, N.R., Verbyla, K., and Visscher, P.M.

(2009). Estimating effects and making predictions from

genome-wide marker data. Stat. Sci. 24, 517–529.

13. Wellcome Trust Case Control Consortium. (2007). Genome-

wide association study of 14,000 cases of seven common

diseases and 3,000 shared controls. Nature 447, 661–678.

14. Powell, J.E., Visscher, P.M., and Goddard, M.E. (2010). Recon-

ciling the analysis of IBD and IBS in complex trait studies. Nat.

Rev. Genet. 11, 800–805.

15. Gilmour, A.R., Thompson, R., and Cullis, B.R. (1995). Average

information REML: An efficient algorithm for variance param-

eter estimation in linear mixed models. Biometrics 51, 1440–

1450.

16. Gilmour, A.R., Gogel, B.J., Cullis, B.R., and Thompson, R.

(2006). ASReml User Guide Release 2.0 (Hemel Hempstead,

UK: VSN International).

17. Barr, D.R., and Sherrill, E.T. (1999). Mean and variance of trun-

cated normal distribution. Am. Stat. 53, 357–361.

18. Dempster, E.R., and Lerner, I.M. (1950). Heritability of

threshold characters. Genetics 35, 212–236.

19. Van Vleck, L.D. (1972). Estimation of heritability of threshold

characters. J. Dairy Sci. 55, 218–225.

20. Daetwyler, H.D., Villanueva, B., and Woolliams, J.A. (2008).

Accuracy of predicting the genetic risk of disease using

a genome-wide approach. PLoS ONE 3, e3395.

21. Venables, W.N., Smith, D.M., and R Development Core Team.

(2010) An Introduction to R. Version 2.11.1, cran.r-project.

org/doc/manuals/R-intro.pdf.

22. Price, A.L., Patterson, N.J., Plenge, R.M., Weinblatt, M.E.,

Shadick, N.A., and Reich, D. (2006). Principal components

analysis corrects for stratification in genome-wide association

studies. Nat. Genet. 38, 904–909.

23. Lee, S.H., Nyholt, D.R., Macgregor, S., Henders, A.K., Zonder-

van, K.T., Montgomery, G.W., and Visscher, P.M. (2010). A

simple and fast two-locus quality control test to detect false

positives due to batch effects in genome-wide association

studies. Genet. Epidemiol. 34, 854–862.

24. Hyttinen, V., Kaprio, J., Kinnunen, L., Koskenvuo, M., and

Tuomilehto, J. (2003). Genetic liability of type 1 diabetes

and the onset age among 22,650 young Finnish twin pairs:

A nationwide follow-up study. Diabetes 52, 1052–1055.

25. Gottesman, I.I., Laursen, T.M., Bertelsen, A., and Mortensen,

P.B. (2010). Severe mental disorders in offspring with 2

psychiatricallyillparents.

252–257.

26. Lichtenstein,P.,Yip,B.H.,Bjo ¨rk,C.,Pawitan,Y.,Cannon,T.D.,

Sullivan, P.F., and Hultman, C.M. (2009). Common genetic

determinantsofschizophreniaandbipolardisorderinSwedish

families: A population-based study. Lancet 373, 234–239.

27. Barrett, J.C., Hansoul, S., Nicolae, D.L., Cho, J.H., Duerr, R.H.,

Rioux,J.D., Brant, S.R., Silverberg, M.S., Taylor, K.D., Barmada,

association studies

Arch. Gen.Psychiatry

67,

304

The American Journal of Human Genetics 88, 294–305, March 11, 2011

Page 12

M.M., et al; NIDDK IBD Genetics Consortium; Belgian-French

IBD Consortium; Wellcome Trust Case Control Consortium.

(2008). Genome-wide association defines more than 30

distinct susceptibility loci for Crohn’s disease. Nat. Genet.

40, 955–962.

28. Yang, J., Lee, S.H., Goddard, M.E., and Visscher, P.M. (2011).

GCTA: A tool for genome-wide complex trait analysis. Am.

J. Hum. Genet. 88, 76–82.

29. Wray, N.R., Yang, J., Goddard, M.E., and Visscher, P.M. (2010).

The genetic interpretation of area under the ROC curve in

genomic profiling. PLoS Genet. 6, e1000864.

30. Schall, R. (1991). Estimation in generalized linear models with

random effects. Biometrika 78, 719–727.

31. Gilmour, A.R., Anderson, R.D., and Rae, A.L. (1985). The anal-

ysis of binomial data by a generalized linear mixed model.

Biometrika 72, 593–599.

The American Journal of Human Genetics 88, 294–305, March 11, 2011

305