REGULAR ARTI CLE

The Impact of GeneEnvironment Interaction

and Correlation on the Interpretation of Heritability
Omri Tal
Received: 16 November 2010 / Accepted: 10 August 2011
Springer Science+Business Media B.V. 2011
Abstract The presence of geneenvironment statistical interaction (GxE) and
correlation (rGE) in biological development has led both practitioners and philos-
ophers of science to question the legitimacy of heritability estimates. The paper
offers a novel approach to assess the impact of GxE and rGE on the way genetic and
environmental causation can be partitioned. A probabilistic framework is developed,
based on a quantitative genetic model that incorporates GxE and rGE, offering a
rigorous way of interpreting heritability estimates. Specically, given an estimate of
heritability and the variance components associated with estimates of GxE and rGE,
I arrive at a probabilistic account of the relative effect of genes and environment.
Keywords Heritability Quantitative genetics Probability GxE interaction
GE correlation
1 Introduction
We may want to get an intuitive feeling for the importance of genetic
variation in the population and a reasonable measure of this relative
importance is the broad heritability.
Lewontin (1975)
Can the broad heritability provide reliable intuition on the importance of genetic
causes in producing phenotypic variation? Heritability is variably expressed as the
correspondence between a latent genetic variable and a measurable phenotypic one,
the slope of the parent-offspring regression, or the proportion of phenotypic
variance due to genetic differences. This latter formulation is common in the
O. Tal (&)
School of Philosophy and The Cohn Institute for the History and Philosophy of Science and Ideas,
Tel Aviv University, Tel Aviv 69978, Israel
e-mail: [email protected]
123
Acta Biotheor
DOI 10.1007/s10441-011-9139-8
literature, and perhaps the least confusing to conceptualize. However, the variance
is only one measure of population dispersion in a certain quantity. The standard
deviation (SD), specied in the underlying units of the target variables, is arguably a
more natural description, and mathematically, a particular variance ratio V
G
/V
P
necessarily implies a different (higher) ratio of SD. But even the SD is not naturally
congruent with how individual differences are commonly conceptualized, since it is
essentially based on squared distances from the population mean, rather than on
pairwise differences. Moreover, heritability estimates are population measures, and
consequently say little about the relative impact of genetic and environmental causal
factors on individual phenotypic development. When nonlinear and interactive
elements are introduced into the developmental framework, the interpretation and
usage of heritability estimates become even more contentious. In a discussion on the
role of genes in development Rose (1999)criticizes the use of heritability estimates,
arguing that such estimates are meaningful only in the absence of geneenvironment
interaction and under a random distribution of genotypes across environments,
If genotypes are distributed randomly across environments, it is possible to
estimate heritability, which denes the proportion of the variance which is
genetically determined. However, the mathematics only works if all the
relevant simplifying assumptions are made. If there is a great deal of
interaction between genes and environment, that is if genes behave according
to Dobzhanskys (1973) vision of norms of reaction, if genes interact with
each other, and if the relationships are not linear and additive but interactive,
the entire mathematical apparatus of heritability estimates falls apart. Thus the
meaningful application of heritability estimates is only possible in very special
cases, from which the majority of traits of interest outside the special world of
articial selection are likely to escape [added emphasis].
The following observation may shed some light on Roses assertion. A partition of
the phenotypic variance into genetic and nongenetic components is related to the
correspondence of a latent variable to a measured one. The quantication of the
correspondence between phenotypic and genotypic values is central to the analysis
of response to selection, familial resemblance and phenotypic development in
quantitative genetics (Lynch and Walsh 1998). A useful measure of the linear
correspondence of P and G under an additive linear model (P = G ? E) is the
squared correlation coefcient, or the coefcient of determination. Formally, this
measures the proportion of the variance in P that is explained by assuming that the
true regression E[P|G] is linear (ibid., p. 47). From basic principles,
q
2
P; G
COVG E; G
r
P
r
G
_ _
2

V
G
COVG; E
2
V
P
V
G

V
G
V
P
:
Crucially, the coefcient of determination is equivalent to the common formulation
for broad heritabilitythe proportion of the phenotypic variance due to genetic
variationonly when the phenotypic model is additive and COV(G, E) is zero.
Discussions of geneenvironment interaction have been mired with conceptual
confusion. The term GxE interaction is often used loosely to mean that both genes
O. Tal
123
and environment contribute to the response variable, but quantitative geneticists
employ the term in the stricter statistical sense. Simply stated, GxE designates a
contribution that some non-additive function of the hidden variables Gand E makes to
the phenotypic value, independently of the main effects of these variables. It can also
be conceptualized as a relationship between the environment and a phenotype that
depends on the genotype, or alternatively a genotype-phenotype relationship that
depends on the environment (Carey 2002). Figure 1 is a depiction of GxE in terms of
norms of reaction, reecting possible relations between the underlying variables.
The term GE correlation refers to the phenomenon where exposure to
environment may have a genetic basis. Such correlation mainly occurs in
observational studies, whenever the environment cannot be randomly assigned
between genotypes in a controlled setting. More formally, GE correlation is a
feature of the distribution of genotypes within environments and exists whenever a
genetic disposition leads individuals to develop under certain environments. The
present paper proposes a method of incorporating estimates of GE interaction and
covariance within a probabilistic framework of genetic effects on individuals.
2 The Model
Quantitative genetic analysis generally proceeds by using measured phenotypic
variances and covariances to estimate latent variance components within the
framework of a postulated quantitative phenotypic model. The foundational discrete
model of GxE interactions is z
ijk
= l ? G
i
? E
j
? I
ij
? e
ijk
where z
ijk
denotes the
value of the kth replicate of genotype i (G
i
) in environment j (E
j
), I
ij
denotes the
interaction between genotype i and environment j, and e
ijk
the specic environment,
all terms with a mean of zero; To stress, e
ijk
is the residual deviation of an
individuals phenotype from the expectation of G
i
? E
j
? I
ij
, where the residuals
are uncorrelated (Lynch and Walsh 1998, p. 108). A standard abstraction of the
discrete model is P = f(G, E), with a simplied model P = l ? G ? E ? I serving
as a workable approximation. In this model the E variable represents both general
systematic and specic nonsystematic environmental effects. In this respect, the
simplied linear model makes no distinction between epigenetic, systematic,
nonsystematic or stochastic effects.
Fig. 1 The norms-of-reaction for three separate developmental scenarios
Heritability, G 9 E and rGE
123
The partition of phenotypic variance follows directly from the standard
quantitative genetic model (Falconer and Mackay 1996),
P G E I !V
P
V
G
V
E
V
I
2COVG; E 1
where the population mean l is either assumed zero or subsumed in the genetic
component. Modeling the measured and latent variables requires several standard
assumptions. It is standard practice to assume the normality of the phenotypic
distribution for many quantitative traits (some traits, such as litter size or lifespan,
are clearly not normally distributed, but adequate transformations can be invoked).
The marginal normality of the two components I and G ? E then follows from the
decomposition of a normal random variable (P = G ? E ? I) and its stability
features under e-independence and deviations from normality (Tal 2009). Finally,
we adopt the standard assumption of the joint-normality of G and E. This is justied
by observations of the linearity of statistical regressions, such as parent-offspring
regression (Lynch and Walsh 1998, p. 552; Tal 2009; Hill 2010), and a marginal
normality of G, due to high proportion of additive genetic variance for complex
traits (Hill et al. 2008). Formally, the quantitative model requires,
[a] P is a standardized normally distributed trait
[b] The joint distribution of G and E is bivariate normal with a possible covariance
term
[c] I is normally distributed
[d] I is statistically independent of G ? E
[e] Availability of estimates: heritability h
2
= V
G
/V
P
, the fraction of V
P
due to
GxE interaction c
2
= V
I
/V
P
, and a GE correlation coefcient, q.
The goal is to arrive at an expression for prob(|G| [|E| | P). Since we are
modeling only the deviations from the means we can standardize P and assume zero
means for all variables without losing information. Thus V
P
= 1 and the variance of
G becomes h
2
. We then get from Eq. 1 and the above model assumptions the
following marginal distributions and a covariance term,
P$N 0; 1 ; G$N 0; h
2
_ _
; E $N 0; V
E
; I $N 0; V
I
; COV G; E
q h

V
E
p
: 2
Naturally, all the variances and SD should be positive, i.e., h, c and r
E
[0,
where r
E
2
= V
E
. In contrast, the correlation coefcient is allowed full range,
-1 \q\1. We would like to express r
E
in terms of the given estimates, h
2
, c
2
and
q. From Eqs. 1 and 2 we have,
1 h
2
r
2
E
c
2
2qhr
E
: 3
Solving this as a quadratic equation for r
E
we get,
r
E

1 h
2
1 q
2
c
2
_
q h: 4
A corollary of Eq. 4 is that h
2
B (1 - c
2
)/(1 - q
2
), with a stricter upper bound
h
2
\1 - c
2
for q[0, from r
E
[0. Since (from assumptions) the probability
O. Tal
123
density function (pdf) for G and E is bivariate normal with correlation of q, the
general expression is,
f
h
2
;r
E
;q
g; e
1
2phr
E

1 q
2

_ e
1
2 1q
2

g
2
h
2
2q
g
h

e
r
E

e
2
r
2
E
_ _
: 5
To formalize the probability that |G| [|E| we rst need to arrive at the joint
density function of G and E conditional on P, denoted F. From rst principles of
conditional probability we have,
F
p;h
2
;c
2
;q
g; e p
G;EjP
g; ejp
p
G;E;P
g; e; p
p
P
p

p
PjG;E
pjg; e p
G;E
g; e
p
P
p

p
I
p g e p
G;E
g; e
p
P
p
: 6
Note that p
I
p g e p
PjG;E
pjg; e since P = G ? E ? I; hence p
PjG;E
pjg; e
has same variance as I (c
2
) but a mean of g ? e. The pdf of a normal random
variable X with zero mean and non-zero variance v (we assume G, E, and I
GxE
in
Eq. 1 are non-degenerate random variables) is given by
u
v
x
1

2pv
p e
x
2
=2v
: 7
In terms of u
v
x and f we have,
F
p;h
2
;c
2
;q
g; e
u
c
2 p g e f
h
2
;r
E
;q
g; e
u
1
p
8
which results in a bivariate normal pdf after explicit substitution and arranging of
terms (see Appendix),
Fg; e
1
2pr
1
r
2

1 q
2
12
_ _
_ e
1
2 1q
2
12

gl
1
r
1
_ _
2
2q
12
gl
1
r
1
_ _

el
2
r
2
_ _

el
2
r
2
_ _
2
_ _
: 9
Figure 2 depicts F within the domain dened by |G| [|E|.
Finally, we employ F to express the conditional probability, prob(|G| [|E| | P).
This probability is in fact the integral of F in the domain that satises |G| [|E|,
expressed as a function of h
2
, and denoted M
p;q;c
2 h
2
,
M
p;q;c
2 h
2
_ _
prob G j j [ E j jjP p
_
1
1
_
jgj
jgj
F
p;h
2
;c
2
;q
g; ede dg

_
0
1
_
g
g
F
p;h
2
;c
2
;q
g; ededg
_
1
0
_
g
g
F
p;h
2
;c
2
;q
g; ede dg: 10
It can easily be shown that the probability is independent of the sign of P, i.e.,
prob(|G| [|E| | P = p) = prob(|G| [|E| | P = -p) for any p; the probability is the
same, whether the phenotype is below or above the population mean. Figure 3ad
depict several instances of M(h
2
) for combinations of P, GxE and rGE, as a function
of heritability.
Heritability, G 9 E and rGE
123
It is instructive to observe some features that are immediately discernable from
the probability graphs. Figure 3ac show a consistent pattern: as genes and
environment become more positively correlated, or as GxE effects increase, the
probability of |G| [|E| is higher across the whole range of heritability values.
Figure 3a and b depict a divergence of the curves as we traverse the heritability
axis: at low heritability estimates and for individuals close to the population mean,
the probability that |G| [|E| is insensitive to the presence of either interaction or
correlation effects. For instance, Fig. 3a shows that for phenotypic values close to
the mean, at a heritability range up to 0.4 there is less than a 10% change in the
resulting probability across a wide range of GE correlation. A different pattern is
discernable for high phenotypic deviations: the probability curves re-converge at the
high heritability range, as Fig. 3c depicts, for P = 3. This means that for high
heritability values the probability that |G| [|E| is less sensitive to the presence of
interaction for individuals that deviate largely from the mean. Figure 3d shows how
the probability curve depends on the phenotypic value. Crucially, all the probability
curves intersect at a single heritability threshold corresponding to the neutral
probability of 50%, where prob(|G| [|E| | P) = 0.5. It is also discernable that up to
this heritability threshold, for any combination of GxE and rGE, lower values of |P|
correspond to a higher probability that |G| [|E|; above that threshold the pattern is
reversedlower values of |P| correspond to lower probabilities.
The probabilities discussed so far are conditional on a phenotypic value. The
unconditional probability of |G| [|E| is simply the average over the phenotypic
distribution. We denote by M the expected value of M across the population,
M
q;c
2 h
2
_ _
prob G j j [ E j j EMp
_
1
1
u
1
pMpdp
1
1

2p
p
_
1
1
e
p
2
2

_
0
1
_
g
g
Fg; ededg
_
1
0
_
g
g
Fg; ede dg
_
_
_
_
_
_dp:
11
Figure 4 depicts M for a particular instance of GxE and rGE.
1.0
0.5
0.0
0.5
1.0
G
1.0
0.5
0.0
0.5
E
0.0
0.5
1.0
1.5
F
G
,
E
Fig. 2 An instance of the
bivariate normal density F,
given h
2
= 0.6, c
2
= 0.1,
q = 0.3 and p = 0.2, where F is
restricted to the domain
satisfying |G| [|E|
O. Tal
123
We now turn to extracting similar probabilities from a purely additive model
where GxE interaction is absent but GE correlation is potentially present. Note we
cannot merely plug c
2
= 0 in expressions (10) and (11), since the general model in
Eq. 1 assumes that a GxE factor exists and has non-zero variance. However, the
probability M converges to M
0
in Eq. 17 as c
2
approaches zero. This will cover those
study designs that do not model GxE but still factor a variance component from
rGE. We therefore proceed in a similar fashion to Eq. 1, briey,
P G E !V
P
V
G
V
E
2COVG; E: 12
Formally, the quantitative framework is,
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Heritability
P
r
o
b

(
|
G
|
>
|
E
|

|

P
)
P = 3, GE correlation = 0, GxE
0.5
0.1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Heritability
P
r
o
b

(
|
G
|
>
|
E
|

|

P
)
P= 0.5, GE correlation = 0.2, GxE
0.5
0.1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Heritability
P
r
o
b

(
|
G
|
>
|
E
|

|

P
)
P = 0.25, GxE variance component = 0.2,
0.3
0.3
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Heritability
P
r
o
b

(
|
G
|
>
|
E
|

|

P
)
GE correlation = 0.3, GxE variance
p=0
p=2
component = 0.1, P: 0 ... 2
variance component: 0.1 ... 0.5 GE correlation 0.3 ... 0.3
variance component: 0.1 ... 0.5
A B
C D
Fig. 3 The graphs of prob(|G| [|E| | P) as a function of the phenotypic value, GE correlation and the
GxE variance component. a The graphs of M(h
2
) for various values of GE correlation. These are the
probabilities that the genetic deviation of an individual |G| had a greater effect than its environmental
deviation |E| on its phenotypic deviation |P|, for p = 0.25 and GxE variance component of 0.2. At the low
heritability range the probability is insensitive to the presence of GE correlation. b The graphs of M(h
2
)
for various values of the GxE effect for p = 0.5 and a GE correlation of 0.2. At the low heritability
range the probability is insensitive to GxE. c The graphs of M(h
2
) for various values of the GxE effect for
p = 3 and a GE correlation of 0. The probability curves converge at high heritabilitythe effect of
GxE is reduced. d The graphs of M(h
2
) for various values of P for a GE correlation of 0.3 and a
GxE variance component of 0.1. The probability curves intersect at a certain heritability value where the
probability is insensitive to the phenotypic value
Heritability, G 9 E and rGE
123
[a] P = G ? E is a normally distributed trait, standardized to unit variance and
zero mean
[b] G and E are bivariate normal, with possible non-zero covariance
[c] We obtain estimates of h
2
and the rGE correlation coefcient q
This implies,
P$N 0; 1 ; G$N 0; h
2
_ _
; E$N 0; r
2
E
_ _
; COV G; E qhr
E
: 13
From Eqs. 12 and 13 we have 1 = h
2
? r
E
2
? 2qhr
E
and solving this as a
quadratic equation for r
E
we get,
r
E

1 h
2
1 q
2

_
q h: 14
Note that q does not restrict the range of h
2
. The bivariate normal pdf for G and E
with correlation of q, denoted f
0
, is,
f
0
h
2
;q
g; e
1
2phr
E

1 q
2

_ e
1
2 1q
2

g
2
h
2
2q
g
h

e
r
E

e
2
r
2
E
_ _
: 15
To formalize the probability F
0
that |G| [|E| we need to arrive at the joint density
function of G conditional on P. From rst principles of conditional probability and
Eq. 12 we have,
F
0
p;h
2
;q
g p
GjP
gjp
p
P;G
p; g
p
P
p

p
EjG
p gjg p
G
g
p
P
p

p
G;E
g; p g
p
P
p

f
0
h
2
;q
g; p g
u
1
p

1

2p h
2
r
2
E
1 q
2

_ e
gph

1r
2
E
1q
2

p

2
2h
2
r
2
E
1q
2

: 16
0.0 0.2 0.4 0.6 0.8
0.0
0.2
0.4
0.6
0.8
1.0
Heritability
GE Corr =0.3, GxE Var component =0.1, averaged over P
P
r
o
b

(
|
G
|
>
|
E
|
)
Fig. 4 The graph of M for a GE correlation of 0.3 and a GxE variance component of 0.1: the probability
that |G| [|E| averaged over the phenotypic distribution
O. Tal
123
F
0
is expressed in Eq. 16 in a form easily recognizable as a normal pdf, where the
variance is h
2
r
2
E
1 q
2
and the mean is ph

1 r
2
E
1 q
2

_
. The expression for
M
0
prob G j j [ E j jjP is nally,
M
0
p;h
2
prob G j j [ E j jjP
_
1
p=2
F
0
p;h
2
;q
gdg if p !0
_
p=2
1
F
0
p;h
2
;q
gdg if p\0
_

_
: 17
The expected value of M
0
across the population is denoted M
0
,
M
0
h
2
_ _
prob G j j [ E j j E M
0
h
2
p
_ _

_
1
1
u
1
pM
0
h
2
pdp

_
0
1
u
1
p
_
p=2
1
F
0
p;h
2
;q
gdgdp
_
1
0
u
1
p
_
1
p=2
F
0
p;h
2
;q
gdgdp
2
_
1
0
1

2p
p e
p
2
2
_
1
p=2
F
0
p;h
2
;q
gdgdp: 18
Figure 5a and b depict the probabilities generated within this additive framework
for the most basic scenario where GE correlation is zero.
3 Discussion
The framework outlined in this paper allows incorporating estimates of gene
environment interaction and covariance within a probabilistic interpretation of
heritability (Tal 2009; see Tal et al. 2010, for an extension that includes a putative
epigenetic variable). Specically, given estimates of heritability and the variance
components associated with GxEand rGE, a method based on the standard quantitative
model generates the conditional probability that genetic factors had a greater effect than
environmental factors on a deviation fromthe population mean. Previous approaches of
reformulating heritability arise from the application of more complex mixed
quantitative models (see Oakey et al. 2007, for a generalized heritability that
incorporates pedigree information to form an extended pedigree model).
Two underlying assumptions of the present framework should be reemphasized.
First, the model takes as a point of departure the standard linear quantitative model
with interaction, P = G ? E ? I, where the three latent terms are modeled as
continuous variables. Second, the probabilistic framework relies on the availability of
estimates of variance components describing the non-additive effects. In this respect,
it is a methodological issue whether a particular sample size or genetic model has
sufcient power to detect and quantify GxE for a target experimental design or
observational study. Indeed, it is well known that the power to detect interaction is
considerably less than that of the main effects (Wahlsten 1990; Sternberg and
Heritability, G 9 E and rGE
123
Grigorenko 1997; Plomin et al. 2008). There are a few experiential strategies for
increasing the power of GxE detection, primarily, using a greater sample size.
However, in controlled experimental designs the detection of GxE may be an artifact
of the maximization of the phenotypic variance, due to the relatively higher additive
genotypic variance compared to natural randomly breeding populations. Consider-
ations of strain replication in the sample are also pertinent. Using high replication of
fewer genotypes in less environments yields greater sensitivity to GxE detection than
low replication of more genotypes in more environments. This illustrates a basic
tension in dening research targets: whether to increase replication of a few
genotypes in a few environments or to better represent the range of genetic or
environmental variation possible at a cost to the sensitivity to detect small differences
(Hodgins-Davis and Townsend 2009). A related methodological issue with quanti-
tative phenotypes is the mathematical ability to induce or remove interaction effects
simply by transformation of scale (Wahlsten 1990). Ultimately, it is an empirical
issue whether GxE interaction exists with respect to a particular target trait, the given
distribution of genotypes and the range of environments.
An important theoretical issue whether the presence of any amount of
interaction, or even the possibility of undetected interaction, renders the partition
of the phenotypic variance meaningless in terms of its causal explanatory content
(see for e.g., Lewontin 1974; Sarkar 1998, for a critical perspective; Sesardic 2005
for an opposing view; Oftedal 2005 for an attempt at conciliation; Tabery 2009 for
the relation to difference mechanisms). Arguably, there is no clear criterion that
allows distinguishing strong from weak interactions. Is it only when norms of
reaction cross, as in the rightmost graph of Fig. 1? Is it when the variance due to
GxE reaches a certain proportion of the total phenotypic variance? What is clear is
that attempting to separate the effects of genes and environment under substantial
GxE is futile, since the interaction component is ultimately some unknown
combination of G and E. Given the standard model with interaction,
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
Heritability
Zero correlation and interaction,
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Heritability
P
r
o
b
(

|
G
|

>

|
E
|

|

P

)
P
r
o
b

(
|
G
|
>
|
E
|
)
Zero correlation and interaction,
p=0
p=0.25
p=2
for varions values of P: 0...2
Averaged over P
Fig. 5 a The graphs of M
0
from Eq. 17 for various values of P (positive or negative deviations) when the
model used is P = G ? E and V
P
= V
G
? V
E
, i.e., no GxE interaction or GE correlation. The
probability curves intersect at h
2
= 0.5 and the probability for p = 0 is 0.5 irrespective of h
2
(compare to
Fig. 3d). b The graph of M
0
from Eq. 18, the probability that |G| [|E| averaged over the distribution of P
O. Tal
123
P = G ? E ? I, if I is large compared to G and E, our probabilistic framework
cannot capture the full relation between the latent genetic and environmental values.
On the other hand, the presence of strong GE correlation does not pose such
difculty. This follows from the fact that the correlation is not a separate component
of the phenotypic value, P = G ? E ? I, but only a characteristic of the joint
distribution of G and E, which is fully captured by the conditional probability
prob(|G| [|E| | P).
Finally, an application of the probabilistic framework would involve plugging the
various estimates of variance components in the probability function M in Eq. 10, or
M in Eq. 11. For instance, if rGE = 0.3 and the variance portion due to GxE is 0.2,
then for a heritability of 0.7 the probability that |G| [|E| is 0.8 for individuals at
SD from the population mean (see Fig. 3a). It is instructive to compare such results
with the probabilities generated from a model that ignores GxE and rGE, via M
0
in
Eq. 17. Using, for comparative illustration, a phenotypic deviation of SD and
heritability of 0.7, we have M
0
(0.7) = prob(|G| [|E| | P = ) = 0.55, quite in
contrast with the higher probability M(0.7) = 0.8, based on a model with some GxE
and rGE effects. In a similar fashion, one could compare the averaged probabilities
from M in Eq. 11 with M
0
in Eq. 18. Such comparisons are most pertinent in the
context of studies that employ a dual design: a reduced-t additive model that
ignores GxE and rGE effects, and a better-t model that incorporates these effects.
The probabilities generated using variance component estimates from the better-t
model may better describe the relative weight of genes and environment on the
deviation of a target trait from the population mean.
Acknowledgments I would like to thank Jim Tabery, Eva Jablonka, John Loehlin, John C. DeFries,
Neven Sesardic, Samir Okasha, Tamir Tassa and two anonymous reviewers for insightful feedback and
suggestions.
Appendix: Derivation of the Bivariate Normal Form for F
We wish to express F from Eq. 8 such that it complies with a bivariate normal form
Eq. 9, if indeed it can be expressed as such. Towards that end we need to nd
expressions for r
1
, r
2
, l
1
, l
2
and q
12
in terms of the parameters of F. Therefore, we
write F such that terms involving powers of its independent variables g and e are
separately expressed,
Fg; e
1
2p h c r
E

1 q
2
_ e
X
2c
2
1q
2
h
2
r
2
E
19
where,
X g
2
1 q
2
_ _
h
2
hr
2
E
c
2
r
2
E
_ _
e
2
1 q
2
_ _
h
2
r
2
E
c
2
h
2
_ _
g e 2p 1 q
2
_ _
h
2
r
2
E
_ _
2ge 1 q
2
_ _
h
2
r
2
E
c
2
qhr
E
_ _
p
2
1 q
2
_ _
h
2
r
2
E
1 c
2
_ _
20
and where r
E

1 h
2
1 q
2
c
2
_
q h, as derived in Eq. 3.
Heritability, G 9 E and rGE
123
Equating similar terms in Eq. 9 with Eqs. 19 and 20 leads to seven simultaneous
equations for the ve unknowns. For instance, equating the g
2
term, we get,
g
2
1 q
2
h
2
r
2
E
c
2
r
2
E
_ _
2 c
2
1 q
2
h
2
r
2
E

g
2
2 1 q
2
12
_ _
r
2
1
and similarly for the terms related to e
2
, g, e, ge, the constant within the exponent
and the coefcient of the exponent. The mean, variance and covariance parameters
for F resulting are,
q
12
S

1
c
2
1 q
2

1 q
2
r
2
E
c
2
1 q
2
h
2
c
2

where,
S
1 if q[
c
2

c
4
4h
2
r
2
E
p
2hr
E
0 if q
c
2

c
4
4h
2
r
2
E
p
2hr
E
1 if q\
c
2

c
4
4h
2
r
2
E
p
2hr
E
_

_
and,
r
1
h

_
1 q
2
_
r
2
E
c
2
_
; r
2
r
E

_
1 q
2
_
h
2
c
2
_
;
l
1

p
c
2
_
__
1 q
2
_
r
2
E
c
2
_
h
2

__
1 q
2
_
r
2
E
c
2
___
1 q
2
_
h
2
c
2
_
c
2
_
1 q
2
_
_
r
E
h
_
;
l
2

p
c
2
_
__
1 q
2
_
h
2
c
2
_
r
2
E

__
1 q
2
_
r
2
E
c
2
___
1 q
2
_
h
2
c
2
_
c
2
_
1 q
2
_
_
r
E
h
_
;
noting that r
E
itself is a function of the input parameters, p, h
2
, c
2
and q
2
, expressed
in Eq. 3.
References
Carey G (2002) Human genetics for the social sciences, 1st edn. Sage, Thousand Oaks
Dobzhansky T (1973) Nothing makes sense except in the light of evolution. Am Biol Teach 35:12529
Falconer DS, Mackay TFC (1996) Introduction to quantitative genetics, 4th edn. Longmans Green,
Harlow
Hill WG (2010) Understanding and using quantitative genetic variation. Philos Trans R Soc B 365:7385
Hill WG, Goddard ME, Visscher PM (2008) Data and theory point to mainly additive genetic variance for
333 complex traits. PLoS Genet 4(2):e1000008. doi:10.1371/journal.pgen.1000008
O. Tal
123
Hodgins-Davis A, Townsend JP (2009) Evolving gene expression: from G to E to GxE. Trends Ecol Evol
24(12):649658
Lewontin RC (1974) The analysis of variance and the analysis of causes. Am J Hum Genet 26:400411
Lewontin RC (1975) Genetic aspects of intelligence. Annu Rev Genet 9:382405
Lynch M, Walsh B (1998) Genetics and analysis of quantitative traits, 1st edn. Sinauer Associates,
Sunderland
Oakey H, Verbyla AP, Cullis BR, Wei X, Pitchford WS (2007) Joint modeling of additive and non-
additive (genetic line) effects in multi-environment trials. Theor Appl Genet 114:13191332
Oftedal G (2005) Heritability and causation. Philos Sci 72:699709
Plomin R, DeFries JC, McClearn GE, McGufn P (2008) Behavioral genetics, 5th edn. Worth, New York
Rose S (1999) Precis of lifelines: biology, freedom, determinism. Behav Brain Sci 22(5):871885
Sarkar S (1998) Genetics and reductionism. Cambridge University Press, Cambridge
Sesardic N (2005) Making sense of heritability. Cambridge University Press, Cambridge
Sternberg RJ, Grigorenko E (1997) Intelligence, heredity, and environment. Cambridge University Press,
Cambridge
Tabery J (2009) Difference mechanisms: explaining variation with mechanisms. Biol Philos 24:645664
Tal O (2009) From heritability to probability. Biol Philos 24:81105
Tal O, Kisdi E, Jablonka E (2010) Epigenetic contribution to covariance between relatives. Genetics
184:10371050
Wahlsten D (1990) Insensitivity of the analysis of variance to heredity-environment interaction. Behav
Brain Sci 13:109161
Heritability, G 9 E and rGE
123