Content uploaded by Emilio Rolán-Alvarez
Author content
All content in this area was uploaded by Emilio Rolán-Alvarez on Jan 11, 2014
Content may be subject to copyright.
Comparison of two methods for analysing
the biological factors contributing
to assortative mating or
sexual isolation
Andrés Pérez-Figueroa1, Jacobo de Uña-Alvarez2, Paula Conde-Padín1
and Emilio Rolán-Alvarez1
1Departamento de Bioquímica, Genética e Immunologia, Facultad de Biología
and 2Departamento do Estadística e Investigación Operativa,
Facultad de Ciencias Económicas y Empresariales,
Universidad de Vigo, Vigo, Spain
ABSTRACT
Question: How can we establish the biological factors that contribute to variation in
assortative mating (based on a quantitative or qualitative trait)?
Key assumptions: Assortative (or disassortative) mating for a particular trait can produce
sexual isolation between ecotypes or incipient species. The individual contribution to
population assortative mating for a quantitative trait can be estimated by means of the ri
statistic, which is an additive decomposition of the Pearson correlation coefficient. The mating
pair contribution to population sexual isolation can be estimated by the PSI coefficient.
These statistics can be used to quantify the variability in assortative mating/sexual isolation
in a particular population.
Search method: It was recently proposed that both the ri statistic and the PSI coefficient could
be used as dependent variables in a multiple regression approach to determine which of a set of
independent variables explains the greatest variation in the dependent variable. We describe
both statistics and undertake simulations to compare the efficiency of each statistic to infer
assortative mating when it is caused a priori by a mate choice decision based on a quantitative or
a qualitative trait.
Conclusions: The ri statistic outperforms the PSI coefficient when trying to infer the causes of
both assortative mating and sexual isolation. The applicability of both methods to other cases is
discussed.
Keywords: estimation properties, incipient speciation, mate choice, mate discrimination,
mating behaviour, regression, speciation.
Correspondence: E. Rolán-Alvarez, Departamento de Bioquímica, Genética e Immunologia, Facultad de
Biología, Universidad de Vigo, 36200 Vigo, Spain. e-mail: rolan@uvigo.es
Consult the copyright statement on the inside front cover for non-commercial copying policies.
Evolutionary Ecology Research, 2008, 10: 1201–1216
© 2008 Emilio Rolán-Alvarez
INTRODUCTION
Mating is a fundamental characteristic of sexual organisms, and partners need to be
carefully chosen to avoid wasted time and effort due to unsuccessful mating within
or between species (Andersson, 1994). Within a particular species, the occurrence of similar
or dissimilar mating pair types that exceeds random mating is called assortative or
disassortative mating respectively (Lewontin et al., 1968). One of the most frequent cases of
assortative mating occurs for body size (when mating preferentially occurs between similar-
sized individuals), and this has been documented extensively in insects, molluscs, reptiles,
birds, and humans (Crespi, 1989; Staub and Ribi, 1995; Jonson, 1999; Delestrade, 2000; Forero et al., 2001; Masello and
Quillfeldt, 2003; Silventoinen et al., 2003). It has been demonstrated that size assortative mating is
involved in the reproductive isolation of two morphs or ecotypes that differ in size (Nagel
and Schluter, 1998; Cruz et al., 2004a; McKinnon et al., 2004; Rolán-Alvarez, 2007). In fact, mating behaviour is one
of the main mechanisms able to produce isolation barriers between incipient species [sexual
isolation, sensu Coyne and Orr (2004)]. A full understanding of the biological mechanisms
responsible for sexual isolation is required to explain how reproductive isolation evolves in situ
as well as how it can be reinforced after secondary contact (Coyne and Orr, 2004). In addition, the
behavioural mechanisms responsible for sexual isolation have been key parameters in some
theoretical models of speciation (Turelli et al., 2001; Kirkpatrick and Ravigné, 2002; Gavrilets, 2004).
The degree of assortative mating for a particular quantitative trait is quantified by
the Pearson (or related) correlation coefficient (Johannesson et al., 1995; Masumoto 1999; Silventoinen
et al., 2003). This correlation provides an estimate of assortative mating for the whole
population, but not for each mating pair separately. An index for the estimation of the
individual contribution (of each pair) to overall assortative mating in the population would
help us to understand the causes of assortative mating and sexual isolation. This approach
has been satisfactorily adopted by employing multiple regression to investigate the causes
and consequences of sexual selection in the wild (Arnold and Wade, 1984a, 1984b; Cruz et al., 2001;
reviewed in Brodie et al., 1995), especially when sexual selection estimates were available for each
specimen. This methodology could also be applied to the study of assortative mating if a
similar index was available.
Recently, a study of sexual isolation and parallel ecological divergence in the marine
gastropod Littorina saxatilis (Conde-Padín et al., 2008) used an additive decomposition of the
Pearson correlation coefficient (called ri), which can be estimated for each mating pair
individually and, therefore, allows one to investigate the contribution of each mating pair to
assortative mating for any trait. This statistic was used together with multiple regression to
examine the biological variables that contribute the most to the variation in assortative
mating. Among a series of shape and size morphological measurements, the square of male
size was the main factor responsible for the observed variation in size assortative mating
(Conde-Padín et al., 2008). The relationship is quadratic because under size assortative mating,
mating pairs showing the greatest deviations from mean size also have the most pronounced
size assortative mating. In addition, size assortative mating was closely linked to sexual
isolation in L. saxatilis, since this species commonly presents size assortative mating, and
in the population studied there were two ecotypes showing clear size differences between
them (Rolán-Alvarez, 2007). Conde-Padín and colleagues also used an estimate of sexual
isolation per mating type, the PSI coefficient (Rolán-Alvarez and Caballero, 2000), in com-
bination with the same multiple regression approach, to study the causes of sexual isolation.
The trend observed was very similar using both statistics, although the results obtained
Pérez-Figueroa et al.1202
suggested that ri was better than the PSI for predicting variation in assortative mating using
a regression approach. Here, we describe the sampling properties of the ri statistic, use a
simulation study to compare the ability of the two statistics to infer the causes of size
assortative mating and sexual isolation, and present an example.
DESCRIPTION OF THE ri STATISTIC AND ITS SAMPLING PROPERTIES
The proposed estimator is based on the Pearson correlation coefficient (Pearson, 1894) but is
calculated for each mating pair separately. We will refer to it as the individual correlation
coefficient (ri), and it is the product of the standardized values of a mating pair for a given
variable:
ri=Zm×Zf,
where Zm and Zf are the values of the variable standardized within each sex and sample.
Thus, each Z value from any individual requires knowing the mean (µ) and the standard
deviation (σ) for males and females independently for each sample obtained. For example,
for each individual male with value xm, Zm=(xm−µm)/σm, and analogously for females.
The mean ri value in the whole population is algebraically equal to the parametric Pearson
correlation coefficient. The use of standardized variables has the advantage of allowing
comparisons of the same variable in populations differing in their means and variances
(Sokal and Rohlf, 1995).
To assess the properties of the sampling distribution of this statistic, we developed
explicit formulae for the standard deviation, the asymmetry, and the kurtosis of the ri
statistic. This is interesting, since little is known about the distribution of a product of
possibly correlated random variables (Nadarajah, 2006). Assume that (X, Y) is a random vector
that follows a bivariate normal distribution. Let µX, µY, σX, σY, σXY denote the mean of X,
the mean of Y, the standard deviation of X, the standard deviation of Y, and the covariance
between X and Y respectively. Furthermore, let ρXY denote the Pearson correlation
coefficient between X and Y, and let r=Z(X)Z(Y)=(X−µX)(Y−µY)/σXσY, where Z(X)
stands for a standardized X variable. Asymptotically, since the sample means and standard
deviations converge to their population counterparts, we can approximate the distribution
of ri by that of r. We use the properties of the conditional expectation to write:
E(r)=E[Z(X)E(Z(Y)/X)].
Since (X, Y) is normally distributed, the conditioned variable Y/X is again normally
distributed with mean µY+ρXY(σY/σX)(X−µX) and variance σ2
Y(1 −ρ2
XY). From this it
follows that E(Z(Y)/X)=ρXY(X−µX)/σX and then E(r)=ρXY is obtained. Now write:
E(r2)=E[Z(X)2E(Z(Y)2/X)].
Taking into account the above mentioned properties of Y/X, it is easily shown that
E(Z(Y)2/X)=1−ρ2
XY +ρ2
XY(X−µX)2/σ2
X.
This gives
E(r2)=1−ρ2
XY +ρ2
XYE(Z(X)4)=1+ρ2
XY ,
where we have used the zero kurtosis property of the normal distribution E(Z(X)4)−3=0.
Biological factors contributing to assortative mating 1203
The variance of r can now be derived as Var(r)=E(r2)−(E(r))2=1+ρ2
XY, which gives
our first formula for the standard deviation:
s.d.(ri)≈√1+ρ2
XY . (1)
For the asymmetry formula we need to derive first the moment of order 3 of r, E(r3).
Now,
E(r3)=E[Z(X)3E(Z(Y)3/X)].
Again, assuming Y/X is normally distributed with mean µY+ρXY(σY/σX)(X−µX) and
variance σ2
Y(1 −ρ2
XY), we easily obtain
E(Z(Y)3/X)=(3 σXY(X−µX) σ2
Y(1 −ρ2
XY)/σ2
X+(σXY /σ2
X)3(X−µX)3)/σ3
Y.
This equality and the properties of the standardized normal distribution,
E[Z(X)3]=3E[Z(X)6]=15
give
E(r3)=3 ρXY(3 +2ρ2
XY).
Since the asymmetry of r is given by
a(r)=E[((r−E(r))/Var(r)1/2)3]=[E(r3)−3E(r)E(r2)+2(E(r))3]/Var(r)3/2,
we obtain (after simple algebra)
a(ri)≈2ρXY(3 +ρ2
XY)
(1 +ρ2
XY)3
2
. (2)
Finally, note that the kurtosis coefficient is written as:
k(r)=E[((r−E(r))/Var(r)1/2)4]−3=[E(r4)−4E(r)E(r3)+6E(r2)(E(r))2−3(E(r))4]−3.
Thus, to obtain an explication of the kurtosis we need to investigate the fourth-order
moment of r. Using the normal distribution of Y/X, we obtain (after some lengthy although
straightforward calculations)
E(Z(Y)4/X)=3(1 −ρ2
XY)2+6 ρ2
XY (1 −ρ2
XY)Z(X)2+ρ4
XY Z(X)4
and hence
E(r4)=E[Z(X)4E(Z(Y)4/X)] =9+72ρ2
XY +24ρ4
XY,
and thus obtain
k(ri)≈6+24ρ2
XY
(1 +ρ2
XY)2. (3)
The three parameters (equations 1–3) are increasing functions of the correlation.
For independent variables (ρXY =0), we have s.d.(ri)≈1, a(ri)≈0, and k(ri)≈6
(a symmetric, but not Gaussian, distribution). Under a perfect correlation (ρXY =1),
the parameter values are those corresponding to a chi-squared distribution with one degree
of freedom.
Pérez-Figueroa et al.1204
SAMPLING PROPERTIES AT LOW SAMPLE SIZE
We simulated the sampling of mating pairs from a finite population to determine whether
the above parameters are maintained at biological sample sizes. We assumed X and Y values
(of the studied trait in each mating pair) obtained from different bivariate distributions,
different sampling sizes, and different levels of a priori correlation between the values. There
were three models involving alternative background distributions:
1. Normal model, using a normal distribution with mean equal to 0 and standard deviation
equal to 1.
2. Uniform model, using a uniform distribution with values between 0 and √12, to obtain a
standard deviation of 1, as in the previous case.
3. Chi-squared model, using a chi-squared distribution with 10 degrees of freedom.
To obtain correlations between X and Y values, they were calculated as follows. First, a
value of X was obtained from the corresponding distribution. After that, a Y value was
sampled following the linear model Y=X+E, where E is an error variable independent of
X, with null mean and a standard deviation (σE) determined by the correlation level (ρXY )
desired between X and Y. Thus, for the normal and uniform models, σE=√1−ρ2
XY, and
for the chi-squared model, σE=√20 ×√1−ρ2
XY. We simulated different sampling sizes
(20, 50, 100, 500, and 1000 mating pairs). After the sampling of all pairs of X and Y values,
these were standardized for their sample (e.g. (X−µx)/σx, with µx and σx the sample
mean and standard deviation of variable X). Finally, we calculated the ri statistic for every
pair and its distribution, defined by the mean, standard deviation, asymmetry, and
kurtosis.
The distribution of the ri statistic under different levels of a priori correlation is presented
in Fig. 1. The sample size is very large (1000 pairs) and the scenario is the normal model.
With no correlation (ρXY =0), the distribution is almost symmetric. In that case, averaging
100 replicates, we obtain µ(ri)=0.00 ±0.00, σ(ri)=1.00 ±0.00, a(ri)=−0.04 ±0.05,
and k(ri)=5.97 ±0.32. As the correlation increases, the distribution becomes more
asymmetrical and leptokurtic, up to the point with perfect correlation (ρXY =1), where the
distribution fits to a chi-squared one with one degree of freedom (averaging 100 replicates:
µ(ri)=1.00 ±0.00, σ(ri)=1.41 ±0.01, a(ri)=2.77 ±0.04, k(ri)=11.05 ±0.41).
Table 1 shows the values of parameters (averaging 100 replicates) defining the ri
distribution with different sample sizes under the normal model. The shape of the
distribution remains mostly constant up to sample sizes of 20 mating pairs. Kurtosis is
the only parameter affected by the small sample size. Thus, as sample size decreases, the
distribution becomes less leptokurtic, especially at greater degrees of correlation. Despite
this, the ri statistic seems appropriate for use with biological sample sizes (at least 20 mating
pairs) because it is not biased by the sampling and maintains its distribution as a function
of the degree of correlation.
A comparison of three models of initial distribution for the mating pairs values is
presented in Fig. 2. There is little difference across the normal, uniform, and chi-squared
models, except for the kurtosis in the latter. Thus, the ri statistic could be used under
conditions where the normality of the trait in the population cannot be assumed.
Biological factors contributing to assortative mating 1205
BRIEF DESCRIPTION OF THE PSI STATISTIC
The only alternative approach for estimation of the individual contribution (of each pair) to
overall assortative mating in the population is the PSI (pair sexual isolation) coefficient
(Rolán-Alvarez and Caballero, 2000). The PSI coefficient is defined for every pair combin-
ation (for a categorical trait, i.e. ecotype) as the number of observed pair types divided by
Fig. 1. Distribution of ri under the normal model with different degrees of correlation (ρXY ) between
mating pairs. Large sample size (1000 pairs).
Table 1. Mean ( µ), standard deviation (σ), asymmetry coefficient (a), and kurtosis (k) of the
distribution of ri values, averaged for 100 replicates, with different sample sizes, under the normal
model and correlation (ρXY) between pairs (numbers in parentheses represent standard error)
ρXY =0ρXY =1
Sample size µ(ri)σ(ri)a(ri)k(ri)µ(ri)σ(ri)a(ri)k(ri)
∞0.00 1.00 0.00 6.00 1.00 1.41 2.82 12.00
500 0.01 1.00 0.00 5.58 1.00 1.41 2.75 10.93
(0.00) (0.00) (0.06) (0.28) (0.00) (0.01) (0.05) (0.65)
100 −0.02 0.99 −0.11 5.06 0.99 1.36 2.39 7.31
(0.01) (0.01) (0.11) (0.38) (0.00) (0.01) (0.06) (0.53)
50 0.00 0.96 0.10 3.77 0.98 1.38 2.43 7.41
(0.02) (0.01) (0.11) (0.34) (0.00) (0.02) (0.08) (0.59)
20 0.01 0.94 0.10 3.05 0.95 1.31 2.04 4.90
(0.02) (0.02) (0.13) (0.31) (0.00) (0.03) (0.08) (0.44)
Pérez-Figueroa et al.1206
the expected pair types from mates: let A and B be the two types of males (with A⬘ and B⬘
the corresponding frequencies of females) in a particular natural or laboratory population.
After a particular period of time, the observed number of mates for each male and female
type is aa, ab, ba, and bb, respectively, for a total number t. Then, the PSI for mates of male
A and female B is PSIab =ab ×t
(aa +ab)×(ab +bb). In the same way, we can compute the PSI
coefficient for the other mate types.
The PSI coefficient has been shown to be useful for estimating sexual isolation by
applying the joint isolation index on the PSI:
IPSI =(PSIaa +PSIbb)−(PSIab +PSIba )
(PSIaa +PSIab +PSIba +PSIbb)
Fig. 2. Distribution of ri under the different models studied with different degrees of correlation (ρXY)
between mating pairs. Large sample size (1000 pairs).
Biological factors contributing to assortative mating 1207
(Rolán-Alvarez and Caballero, 2000). This estimator, the IPSI, is presently considered one of
the most accurate and unbiased sexual isolation estimators, especially at biological sample
sizes (Pérez-Figueroa et al., 2005).
Although the PSI is defined for qualitative traits, it could be used to analyse assortative
mating by a quantitative trait by dividing it into a few discrete classes. In this way, the PSI
could be used as the dependent variable in a multiple regression to investigate the biological
variables that contribute the most to the variation in assortative mating (Conde-Padín et al.,
2008).
STATISTICAL COMPARISON BETWEEN ri AND PSI REGRESSION APPROACHES
Computer simulations were developed to compare the ability of the different statistics to
estimate different levels of assortative mating caused by different traits. We simulated the
formation of mating pairs in a population assuming different a priori levels of assortative
mating for a quantitative or a qualitative trait. Thus, we obtained a random sample of mates
from such a population and, a posteriori, we could estimate the assortative mating (ri and
IPSI ). One of the objectives is to check a posteriori which estimator is the best
predictor of the simulated (a priori) assortative mating. In addition, we could also check
which of the estimators (ri or PSI) is the most useful under the exploratory tool developed
to infer the causes of assortative mating following Conde-Padín et al. (2008). This can be
done by linear regression, using the ri and PSI statistics as dependent variables and the
square of the trait (qualitative or quantitative) as an independent variable. Note that the
square of the trait is used because the relationship between any trait and the assortative
mating should be quadratic. For example, in the case of size assortative mating, both low
and high values of size will contribute to a pattern of size assortative mating. The coefficient
of determination (r2) can be used as an estimate of the efficiency of a particular regression
equation for predicting the variation in the dependent variable (Sokal and Rohlf, 1995). This
coefficient can be interpreted (when multiplied by 100) as the percentage of variation
explained by the regression model.
The simulation procedure was as follows. First, we obtained a base population of 100,000
individuals. Each individual consisted of a value for a quantitative trait, taken from a
normal distribution with mean equal to 0 and standard deviation equal to 1, and a state
for a related discriminant trait, 0 if the value for the quantitative trait was negative and
1 otherwise. Note that these traits are causally correlated (empirical correlation in the
simulated population was 0.88). There were two possible scenarios depending on the
a priori mechanism of assortative mating. First, if assortative mating was caused by
the quantitative trait, then two randomly chosen individuals mate with a probability equal
to 1 −(|x−y|×ρ), where x and y are the values for the quantitative trait in each pair of
mates, and ρ is the desired a priori level of assortative mating (range between 0 and 1). This
produced a better chance of mating for those pairs with a small trait difference or a small
assortative mating trend (or both). For example, under random mating the mating
probability is 1. Alternatively, under high a priori assortative mating, only rather small
differences in the trait between mates will result in a reasonable probability of mating.
When assortative mating was caused by differences in a qualitative trait (scaled with
values 0 and 1), we used a different approach. In this case, four different mating pair
combinations are possible: matings between 1/1 (male 1 with female 1), 1/0 (male 1 with
female 0), 0/1 (male 0 with female 1), and 0/0 (male 0 with female 0). Under an assortative
Pérez-Figueroa et al.1208
mating situation, we would expect the following mating preferences for the above
corresponding mating pair combinations: 1, 1 −c, 1 −c, and 1. Then, by using the IPSI
algorithm, it is possible to obtain the best estimate of the isolation (ρ):
ρ=(1 +1) −((1 −c)+(1 −c))
2+2c.
From the above equation it is possible to estimate c:
c=2ρ−2
−2−2ρ.
The simplest way to simulate assortative mating is to allow matings if they have the same
state (1/1 or 0/0), but allowing them to mate only with a probability equal to c if they have
different states (1/0 or 0/1).
We obtained 100 mating pairs by each of the two a priori mechanisms. Four values
(ρ=0, 0.3, 0.6, and 0.9) of a priori levels of assortative mating were run. In those simulated
pairs, we calculated a posteriori the ri and PSI coefficients for every mating pair (note that
the PSI coefficient for a mating pair stands for the one corresponding to its mate pair
combination: 1/1, 1/0, 0/1 or 0/0). We expect that ri and IPSI will be close to ρ if they work
properly. In addition, given that assortative mating is only caused by one trait in this model,
the percentage of explained variation in assortative mating (ri or PSI) by the squared
trait (quantitative or qualitative) gives us the possibility of quantifying their contribution
a posteriori, and so indirectly to evaluate which of the variables causes it. If this approach
is useful we would expect a lower percentage of explained variation under low levels of
a priori assortative mating and a higher percentage of explained variation in the opposite
situation. Each sample of mating pairs was re-sampled from the simulated populations
10,000 times.
Table 2 shows the estimated a posteriori overall assortative mating level (using IPSI and ri
coefficients) and the variance explained (r2 under a regression approach) by the squared
quantitative/qualitative variables on the estimated assortative mating level (ri or PSI). When
the a priori assortative mating (given by ρ) is caused by a preference on a quantitative trait, ri
is a better predictor of ρ than IPSI, and vice versa when it is based on a qualitative difference.
The latter is expected from the estimation properties of IPSI (described in Pérez-Figueroa et al.,
2005) given that this index is unbiased in a wide range of situations. However, when these
estimators were used under the regression approach to infer the causes of assortative
mating, the ri statistic outperformed the PSI coefficient in all circumstances – that is, the
former typically showed a better fit regression with the putatively causal variables of the
assortative mating (Table 2). This happened even when the assortative mating was causally
produced by the qualitative trait and so the IPSI is the best predictor of the assortative
mating. The reason for such an apparent contradiction may be that the PSI coefficient is not
in fact a random variable in the sample, but rather a qualitative index for each mate type
combination and so it may produce very poor fit with regression of individual pairs. In
addition, the degree of variance explained by the regression model increases linearly with
the level of a priori assortative mating simulated; the larger the ρ values, the larger the
percentage of variance explained by the association. For example, the correlation between ρ
and r2
qt was 0.98 when assortative mating was caused by the quantitative trait and 0.93 when
Biological factors contributing to assortative mating 1209
Table 2. Results of comparative analyses for detecting traits contributing to the assortative mating
A priori quantitative A priori qualitative
riPSI riPSI
ρrir2
qt r2
ql IPSI r2
qt r2
ql rir2
qt r2
ql IPSI r2
qt r2
ql
0.0 0.00 0.037 0.004 0.00 0.010 0.00 0.00 0.037 0.004 −0.01 0.010 0.001
(0.00) (0.05) (0.01) (0.00) (0.01) (0.00) (0.00) (0.05) (0.01) (0.00) (0.01) (0.01)
0.3 0.32 0.194 0.008 0.21 0.029 0.003 0.19 0.056 0.005 0.30 0.011 0.005
(0.00) (0.12) (0.01) (0.00) (0.03) (0.01) (0.00) (0.07) (0.01) (0.00) (0.01) (0.01)
0.6 0.68 0.657 0.036 0.46 0.100 0.019 0.38 0.122 0.011 0.60 0.011 0.025
(0.00) (0.10) (0.03) (0.00) (0.04) (0.03) (0.00) (0.01) (0.02) (0.00) (0.01) (0.04)
0.9 0.82 0.826 0.060 0.59 0.128 0.052 0.57 0.274 0.025 0.90 0.015 0.156
(0.00) (0.08) (0.04) (0.00) (0.05) (0.06) (0.00) (0.13) (0.03) (0.00) (0.02) (0.18)
Note: The regression approach is applied on two different estimators of assortative mating (ri and PSI), when it is a priori caused by quantitative or qualitative traits
across different degrees of a priori assortative mating (ρ). For each scenario, the table shows the assortative mating estimate in the population (mean ri and IPSI), and the
variance explained by the regression of traits on assortative mating. In the latter case, the ability to infer the causes of assortative mating is estimated by the squared
correlation of ri (or PSI) with the squared trait (quantitative =r2
qt; qualitative =r2
ql). Such an a posteriori squared correlation (r2) should be positively related with the true
a priori degree of assortative mating (ρ). Numbers in parentheses are standard deviations.
it was caused by the qualitative trait. Somewhat weaker correlations were also observed
between ρ and r2
ql (0.97 and 0.80, respectively). This indicates that ri (and to a lesser extend
also the PSI) can be useful to indirectly infer the factors causing assortative mating,
although the former will always show a higher prediction capability (higher r2).
EXAMPLE OF APPLICATION
In addition, we used published data from Conde-Padín et al. (2008) to apply the ri and PSI
statistics as exploratory tools to detect which variables contribute the most to the individual
variability in assortative mating using multiple regression analysis. We will briefly introduce
a real example to illustrate a practical application of the statistic.
In Galician exposed rocky shores, a striking polymorphism of the marine snail Littorina
saxatilis is found associated with different shore levels and habitats (Johannesson et al., 1993;
Quesada et al., 2007; reviewed by Rolán-Alvarez, 2007). On the upper shore, the RB (ridged and banded)
ecotype is associated with the barnacle belt, while the SU (smooth and unbanded) ecotype
is associated with the mussel belt on the lower shore. Mussels and barnacles overlap at the
mid-shore, creating a patchy micro-habitat, where these two pure ecotypes meet and
occasionally mate, and a variable percentage of intermediate fertile forms (called hybrids;
HY) are observed. These two ecotypes differ for many morphological, behavioural, and
even life-history characteristics, mostly due to the existence of disruptive selection acting
across the vertical environmental gradient (Rolán-Alvarez et al., 1999; Cruz et al., 2001, 2004b, 2004c;
Conde-Padín et al., 2007). In spite of the hybridization at the mid-shore, some partial sexual isolation
(70% of the maximum possible on average) contributes to the maintenance of the ecotype
differences across the environmental gradient (Johannesson et al., 1995; Rolán-Alvarez et al., 1999, 2004;
Cruz et al., 2004a; Quesada et al., 2007). This ecotype assortative mating was indirectly caused by the
existing size assortative mating and the mean size difference between ecotypes (see Cruz et al.,
2004a; Rolán-Alvarez et al., 2004; Rolán-Alvarez, 2007; Conde-Padín et al., 2007). Additionally, it has been proposed
that approximately half of the ecotype assortative mating can be achieved by snail micro-
aggregation, perhaps caused by active search for refuges (Kostylev et al., 1997; Erlandsson et al., 1999),
or by the existence of different preferences in RB and SU ecotypes for mussel and barnacle
micro-patches at the mid-shore (Otero-Schmitt et al., 1997; Carballo et al., 2005). In summary, it could
be advanced that size assortative mating in this population should be influenced to some
extent by the particular micro-habitat conditions.
The mating pairs were captured during May and June 2006 in Silleiro and Centinela in
the L. saxatilis hybrid zone of Galicia (NW of Spain), during low tide directly on the rocky
shore. The specimens from mating pairs were classified as belonging to a particular ecotype
(RB, SU or HY) and sex. Shell size was estimated as the distance between the shell apex
and the shell base (for further details, see Conde-Padín et al., 2008). With these data we could obtain the
individual contribution of each pair to the assortative mating by the ri and the PSI statistics
(see above) to be used as dependent variables in the exploratory multiple regression analysis.
A plastic circle (20 cm in diameter) placed over each mating pair was used to obtain some
environmental/demographic variables inside those micro-areas. We took a digital photo-
graph of the micro-area and in the laboratory we divided the photograph into 16 large
quadrates, each one divided into 16 small quadrates (256 small squares in total). Thus we
obtained from such images the relative abundance of mussels (number of small quadrates
covered by mussels), and the relative abundance of barnacles within the micro-area. The
aggregation of mussels and the aggregation of barnacles were obtained by dividing the mean
Biological factors contributing to assortative mating 1211
number of small quadrates by their variance across large quadrates (following Taylor, 1984; Margalef,
1991). We also obtained two linear profiles (horizontal and vertical profiles in relation to the
shoreline) of the surface of the micro-area following Kostylev et al. (1997). We used the length
of these profiles, and the mean, as estimates of the surface rugosity (Conde-Padín et al., 2008).
A more detailed analysis of morphological variables is presented elsewhere (Conde-Padín et al.,
2008). Some of those environmental variables might partially contribute to the existing size
assortative mating in these populations, and so they were used as independent variables using
a multiple regression approach.
EXPLORATORY REGRESSION ANALYSIS IN L. SAXATILIS
Mating pairs and corresponding environmental variables were used to explore the environ-
mental factors contributing to or affecting assortative mating in the wild. The ri and PSI
values of pairs were used as dependent variables in a multiple regression analysis, using
the square of the associated environmental variables (relative abundance of mussels and
barnacles, aggregation of mussels and barnacles, horizontal, vertical, and mean profiles) as
independent ones. Thus we used seven environmental variables under a step-wise multiple
regression approach, with forward (P<0.05) and backward (P<0.1) criteria to determine
the significant variables contributing to the size assortative mating. All these calculations
were done with the SPSS/PC software version 14.0.
When using ri under multiple regression (Table 3), one environmental variable explained
part of the variation in assortative mating in Silleiro (horizontal profile2) and Centinela
(mean profile2). Interestingly, the regression was not significant when using the PSI
coefficient (Table 3). The relationship was positive in Centinela (suggesting that both low
and high rugosity contribute to assortative mating) and negative in Silleiro (suggesting
that intermediate rugosity contributes most to assortative mating). In summary, one
environmental factor (substrate rugosity) could also contribute to the pattern of assortative
mating. However, the fact that this relationship is different in sign between localities perhaps
suggests that it has no causative role in explaining the pattern of assortative mating in
the wild.
Table 3. Results of step-wise linear regression of the environmental variables explaining the
individual contribution to the assortative mating (ri) in Silleiro and Centinela
Dependent
variable
Locality Independent environmental
variables
r2Coefficient of partial
regression
riSilleiro horizontal profile20.124 −0.353*
Centinela mean profile20.196 0.442***
PSI Silleiro ———
Centinela ———
Note: The variance explained by the only variable introduced (r2) in the model and its partial regression coefficients
are shown. The regression of environmental variables on PSI coefficients was not significant.
*P<0.05, **P<0.01, ***P<0.001.
Pérez-Figueroa et al.1212
DISCUSSION
A modified version of the Pearson correlation coefficient is described here to study the
causes of assortative (or disassortative) mating in any quantitative trait. The statistic shows
relatively robust sampling properties under low sample size or for non-normal variables,
and the development of its sampling standard deviation will allow it to be used under
parametric hypothesis testing or for generating sampling confidence intervals efficiently.
In addition, this index allows exploration of the causes of assortative mating directly
in the wild for a particular species using a regression approach. The ri statistic was
compared with another statistic [PSI (from Rolán-Alvarez and Caballero, 2000)] to allow this
kind of exploration.
The ri method outperformed the PSI in all situations when used to infer the causes
of assortative mating under a regression approach, which suggests that the PSI statistic
should be avoided for such exploratory analysis unless there is no alternative available.
The ri showed a high predictive ability to detect the causal relationship of variables with
assortative mating. Although the simulation conditions were simple, they show that the
percentage of variance explained by the regression is directly related to the a priori degree of
assortative mating, allowing this methodology to be used in more complex situations. For
example, this method has also been employed to determine the morphological traits (from
male or females) that are most relevant to mate choice in the marine gastropod Littorina
saxatilis (Conde-Padín et al., 2008), which represents a case of sympatric ecological speciation
(Rolán-Alvarez, 2007; Quesada et al., 2007). In this case, the dependent variable was a canonical
discriminant score used to distinguish between these ecotypes, while the independent
variables were shell size and shape traits. The results suggest that male (not female) size is
the main determinant of the degree of ecotype assortative mating in this model system.
These results were independently corroborated by a laboratory experiment in which males
could choose among different types of females [with different ecotypes and sizes (Conde-
Padín et al., 2008; see also Cruz et al., 2004a; Rolán-Alvarez et al., 2004)]. The results from our simulations
corroborate the previous use of this approach.
In addition, in a re-analysis of the above data we confirmed that one environmental
variable could explain part of the variation in assortative mating (only when using ri).
This corroborates the results of simulations: the ri statistic was able to detect a significant
relationship in the example, suggesting again that it has greater statistical power than
the PSI statistic under the regression approach. There are many cases in which the use
of this methodology could facilitate the understanding of mating behaviour in the wild
(e.g. Johannesson et al., 1995; Erlandsson and Rolán-Alvarez, 1998; Masumoto, 1999; Shine et al., 2001; Silventoinen et al., 2003;
Hollander et al., 2005).
The applicability of the method rests on the assumption that among a group of biological
traits, the one that shows the strongest correlation with the studied variable (sexual isolation
or size assortative mating) is most likely causally related to the studied variable. It has been
argued, however, that any correlation between two variables does not guarantee any causal
relationship between them (Sokal and Rohlf, 1995), as a third variable, not included in the study,
could be responsible. Such problems with correlation or regression analysis are well-known,
although they have not limited the applicability of the regression model to biology, ecology
or evolution (Brodie et al., 1995). Some caution, however, is needed. For example, the particular
variables included in the model with minor contributions (r2×100 <5–10%) are not
necessarily safely selected by the regression algorithm.
Biological factors contributing to assortative mating 1213
General progress has been made in the last few years studying the genetic causes of
post-zygotic isolation (reviewed in Coyne and Orr, 2004). However, little progress has been made in
understanding the genetic effects of pre-zygotic isolation, as with sexual isolation (Coyne and
Orr, 2004; but see Coyne, 1996; Pugh and Ritchie, 1996; Carracedo et al., 1998). This could be caused in part by
the difficulty of understanding the biological mechanisms contributing to sexual isolation.
The use of the ri statistic could facilitate future research on this mechanism of sexual
isolation, and it has the advantage of being applicable in any case in which two or
more species are being studied (see, for example, Coyne et al., 2005; Giokas et al., 2006). In addition, this
statistic can be also used with many biological variables – anatomical, physiological or even
behavioural ones – used to explore linear and non-linear relationships, the interactions
between independent variables, or measurements previously corrected for spurious effects
caused by environmental variables (see Lande and Arnold, 1983; Rausher, 1992; van Tienderen and De Jong,
1994; Brodie et al., 1995).
ACKNOWLEDGEMENTS
We thank A. Caballero, J.J. Pasantes, and M. Santos for useful suggestions on preliminary versions of
this manuscript, as well as the following institutions for general funding: European Union (code
EVK3-CT-2001-00048), Ministerio de Educación y Ciencia (code CGL2008-00135/BOS), Xunta de
Galicia (code PGIDT02PXIC30101PM; PGIDT06PXIB310247PR), and University of Vigo. P. C.-P.
thanks the Ministerio de Educación y Ciencia from Spain for a research grant. A.P-F. is currently
funded by an Ángeles Alvariño research fellowship from Xunta de Galicia (Spain).
REFERENCES
Andersson, M. 1994. Sexual Selection. Princeton, NJ: Princeton University Press.
Arnold, S.J. and Wade, M.J. 1984a. On the measurement of natural and sexual selection: theory.
Evolution, 38: 709–719.
Arnold, S.J. and Wade, M.J. 1984b. On the measurement of natural and sexual selection:
applications. Evolution, 38: 709–719.
Brodie, E.D., Moore, A.J. and Janzen, F.J. 1995. Visualizing and quantifying natural selection.
Trends Ecol. Evol., 10: 313–318.
Carballo, M., Caballero, A. and Rolán-Alvarez, E. 2005. Habitat-dependent ecotype micro-
distribution at the mid shore in natural populations of Littorina saxatilis. Hydrobiologia, 548:
307–311.
Carracedo, M.C., Suarez, B., Asenjo, A. and Casares, P. 1998. Genetics of hybridization between
Drosophila simulans females and D. melanogaster males. Heredity, 80: 17–24.
Conde-Padín, P., Carvajal-Rodríguez, A., Carballo, M., Caballero, A. and Rolán-Alvarez, E. 2007.
Genetic variation for shell traits in a direct-developing marine snail involved in a putative
sympatric ecological speciation process. Evol. Ecol., 21: 635–650.
Conde-Padín, P., Cruz, R., Hollander, J. and Rolán-Alvarez, E. 2008. Revealing the mechanisms
of sexual isolation in a case of sympatric and parallel ecological divergence. Biol. J. Linn. Soc.,
94: 513–526.
Coyne, J.A. 1996. Genetics of sexual isolation in male hybrids of Drosophila simulans and
D. mauritiana. Genet. Res., 68: 211–220.
Coyne, J.A. and Orr, H.A. 2004. Speciation. Sunderland, MA: Sinauer Associates.
Coyne, J.A., Elwyn, S. and Rolán-Alvarez, E. 2005. Impact of experimental design on Drosophila
sexual isolation studies: direct effects and comparison to field hybridization data. Evolution, 59:
2588–2601.
Pérez-Figueroa et al.1214
Crespi, B.J. 1989. Causes of assortative mating in arthropods. Anim. Behav., 38: 980–1000.
Cruz, R., Rolán-Alvarez, E. and García, C. 2001. Sexual selection on phenotypic traits in a hybrid
zone of Littorina saxatilis (Olivi). J. Evol. Biol., 14: 773–785.
Cruz, R., Carballo, M., Conde-Padín, P. and Rolán-Alvarez, E. 2004a. Testing alternative models
for sexual isolation in natural populations of Littorina saxatilis: indirect support for by-product
ecological speciation? J. Evol. Biol., 17: 288–293.
Cruz, R., Vilas, C., Mosquera, J. and García, C. 2004b. Relative contribution of dispersal and
natural selection to the maintenance of a hybrid zone in Littorina. Evolution, 58: 2734–2746.
Cruz, R., Vilas, C., Mosquera, J. and García, C. 2004c. The close relationship between estimated
divergent selection and observed differentiation supports the selective origin of a marine snail
hybrid zone. J. Evol. Biol., 17: 1221–1229.
Delestrade, A. 2000. Sexual size dimorphism and positive assortative mating in Alpine choughs
(Pyrrhocorax graculus). Auk, 118: 553–556.
Erlandsson, J. and Rolán-Alvarez, E. 1998. Sexual selection and assortative mating by size and
their roles in the maintenance of a polymorphism in Swedish Littorina saxatilis populations.
Hydrobiologia, 378: 59–69.
Erlandsson, J., Kostylev, V. and Rolán-Alvarez, E. 1999. Mate search and aggregation behaviour in
the Galician hybrid zone of Littorina saxatilis. J. Evol. Biol., 12: 891–896.
Forero, M.G., Tella, J.L., Donazar, J.A., Blanco, G., Bertellotti, M. and Ceballos, O. 2001.
Phenotypic assortative mating and within-pair sexual dimorphism and its influence on breeding
success and offspring quality in Magellanic penguins. Can. J. Zool., 79: 1414–1422.
Gavrilets, S. 2004. Fitness Landscapes and the Origin of Species. Princeton, NJ: Princeton University
Press.
Giokas, S., Mylonas, M. and Rolán-Alvarez, E. 2006. Disassociation between weak sexual isolation
and genetic divergence in a hermaphroditic land snail and implications about chirality. J. Evol.
Biol., 19: 1631–1640.
Hollander J., Lindegarth M. and Johannesson, K. 2005. Local adaptation but not geographic
separation promotes assortative mating in a snail – support for ecological speciation. Anim.
Behav., 5: 1209–1219.
Johannesson, K., Johannesson, B. and Rolán-Alvarez, E. 1993. Morphological differentiation and
genetic cohesiveness over a microenvironmental gradient in the marine snail Littorina saxatilis.
Evolution, 47: 1770–1787.
Johannesson, K., Rolán-Alvarez, E. and Ekendahl, A. 1995. Incipient reproductive isolation
between two sympatric morphs of the intertidal snail Littorina saxatilis. Evolution, 49:
1180–1190.
Jonson, L.J. 1999. Size assortative mating in the marine snail Littorina neglecta. J. Mar. Biol. Assoc.
UK, 79: 1131–1132.
Kirkpatrick, M. and Ravigné, V. 2002. Speciation by natural selection and sexual selection: models
and experiments. Am. Nat., 139: 22–35.
Kostylev, V., Erlandsson, J. and Johannesson, K. 1997. Microdistribution of the polymorphic snail
Littorina saxatilis (Olivi) in a patchy rocky shore habitat. Ophelia, 47: 1–12.
Lande, R. and Arnold, S.J. 1983. The measurement of selection on correlated characters. Evolution,
37: 1210–1226.
Lewontin, R., Kirk, D. and Crow, J. 1968. Selective mating, assortative mating, and inbreeding:
definitions and implications. Eugenics Quart., 15: 141–143.
Margalef, R. 1991. Ecología. Barcelona: Ediciones Omega.
Masello, J.F. and Quillfeldt, P. 2003. Body size, body condition and ornamental feathers of
Burrowing Parrots: variation between years and sexes, assortative mating and influences
on breeding success. Emu, 103: 149–161.
Masumoto, T. 1999. Size assortative mating and reproductive success of the funnel-web spider,
Agelena limbata (Aracneae; Agelenidae). J. Insect Behav., 12: 353–361.
Biological factors contributing to assortative mating 1215
McKinnon, J.S., Mori, S., Blackman, B.K., David, L., Kingsley, D.M., Jamieson, L. et al. 2004.
Evidence for ecology’s role in speciation. Nature, 429: 294–298.
Nadarajah, S. 2006. Exact and approximate distributions for the product of inverted Dirichlet
components. Statistical Papers, 47: 551–568.
Nagel, L. and Schluter, D. 1998. Body size, natural selection, and speciation in sticklebacks.
Evolution, 52: 209–218.
Otero-Schmitt, J., Cruz, R., García, C. and Rolán-Alvarez, E. 1997. Feeding strategy and habitat
choice in Littorina saxatilis (Gastropoda: Prosobranchia) and their role in the origin and
maintenance of a sympatric polymorphism. Ophelia, 46: 205–216.
Pearson, K. 1894. Contributions to the mathematical theory of evolution. Phil. Trans. R. Soc. Lond.
A, 185: 71–110.
Pérez-Figueroa, A., Caballero, A. and Rolán-Alvarez, E. 2005. Comparing the estimation properties
of different statistics for measuring sexual isolation from mating frequencies. Biol. J. Linn. Soc.,
85: 307–318.
Pugh, A.R.G. and Ritchie, M.G. 1996. Polygenic control of a mating signal in Drosophila. Heredity,
77: 378–382.
Quesada, H., Posada, D., Morán, P., Caballero, A. and Rolán-Alvarez, E. 2007. Phylogenetic
evidence for multiple sympatric ecological diversification in a marine snail. Evolution, 61:
1600–1612.
Rausher, M.D. 1992. The measurement of selection on quantitative traits: biases due to environ-
mental covariances between traits and fitness. Evolution, 46: 616–626.
Rolan-Alvarez, E. 2007. Sympatric speciation as a by-product of ecological adaptation in the
Galician Littorina saxatilis hybrid zone. J. Mollusc. Stud., 73: 1–10.
Rolan-Alvarez, E. and Caballero, A. 2000. Estimating sexual selection and sexual isolation effects
from mating frequencies. Evolution, 54: 30–36.
Rolán-Alvarez, E., Erlandsson, J., Johannesson, K. and Cruz, R. 1999. Mechanisms of incomplete
prezygotic reproductive isolation in an intertidal snail: testing behavioural models in wild
populations. J. Evol. Biol., 12: 879–890.
Rolán-Alvarez, E., Carballo, M., Galindo, J., Morán, P., Fernández, B., Caballero, A. et al.
2004. Non-allopatric origin of local reproductive barriers between two snail ecotypes. Molec.
Ecol., 13: 3415–3424.
Shine, R., O’Connor, D., Lemaster, M.P. and Mason, R.T. 2001. Pick on someone your size:
ontogenetic shifts in mate choice by male garter snakes result in size assortative mating. Anim.
Behav., 61: 1133–1141.
Silventoinen, K., Kaprio, J., Lahelma, E., Viken, R.J. and Rose, R.J. 2003. Assortative mating by
body height and BMI: Finnish twins and their spouses. Am. J. Human Biol., 15: 620–627.
Sokal, R.R. and Rohlf, F.J. 1995. Biometry, 3rd edn. New York: Freeman.
Staub, R. and Ribi, G. 1995. Size-assortative mating in a natural population of Viviparus ater
(Gastropoda: Prosobranchia) in Lake Zürich, Switzerland. J. Mollusc. Stud., 61: 237–247.
Taylor, L.R. 1984. Assessing and interpreting the spatial distributions of insect populations.
Annu. Rev. Entomol., 29: 321–357.
Turelli, M., Barton, N.H. and Coyne, J.A. 2001. Theory and speciation. Trends Ecol. Evol., 15:
330–343.
Van Tienderen, P.H. and de Jong, G. 1994. A general model of the relation between phenotypic
selection and genetic response. J. Evol. Biol., 7: 1–12.
Pérez-Figueroa et al.1216
