Arizona-Nevada Academy of Sciences

Assortative Mating in the Jewel Wasp. 2: Sequential Canonical Analysis as an Exploratory

Form of Path Analysis
Author(s): Aurelio José Figueredo and Richard L. Gorsuch
Source: Journal of the Arizona-Nevada Academy of Science, Vol. 39, No. 2, Psychology:
Analysis and Measurement (2007), pp. 59-64
Published by: Arizona-Nevada Academy of Sciences
Stable URL: http://www.jstor.org/stable/40022352
Accessed: 21-07-2017 16:36 UTC

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

Arizona-Nevada Academy of Sciences is collaborating with JSTOR to digitize, preserve and extend access
to Journal of the Arizona-Nevada Academy of Science

This content downloaded from 148.209.209.83 on Fri, 21 Jul 2017 16:36:36 UTC
All use subject to http://about.jstor.org/terms
 assortative mating in the jewel wasp! 2. sequential
Canonical Analysis as an
Exploratory Form of Path Analysis

Aurelio Jose Figueredo, Department of Psychology, University of Arizona, Tucson, AZ 85721; and
Richard L. Gorsuch, Graduate School of Psychology, Fuller Theological Seminary, Pasadena, CA
91182

Abstract
This paper discusses the conceptual and mathematical relationships between three statistical models: 1)
traditional, or "Simultaneous", Canonical Analysis, 2) hierarchical, or "Sequential", Canonical Analysis, and 3)
Structural Equations Modeling, or confirmatory path analysis. The advantages of Sequential over Simultaneous
Canonical Analysis are reviewed. The relationships between Sequential Canonical Analysis and Path Analysis are
explored. The need for a legitimate exploratory form of path analysis, analogous to existing exploratory forms of both
multiple regression and factor analysis, is discussed. A logical extension of Sequential Canonical Analysis is proposed
as adequately serving the function of an exploratory path analysis. Empirical data from psychological research is used
to illustrate and qualitatively compare and contrast the results of these three approaches.

Introduction A third alternative is sequential canonical analysis

(SEQCA), which combines some of the advantages of
It is often the case that the multiple outcomes of an
both (Gorsuch 1991, Gorsuch and Figueredo 1991).
experimental treatment or manipulation need to be
Rather than combining all dependent variables into
assessed. Moreover, these dependent variables are of-
uninterpretable linear composites, as a simultaneous
ten selected to measure the impact of the treatment on
regression does for independent variables, it partitions
several conceptually distinct outcomes, rather than con-
their covariance sequentially, as a hierarchical regres-
verge upon a single construct. Thus, multivariate data
sion does, while maintaining their separate identity.
reduction methods, such as common factor modeling,
This method isolates the direct effects of the indepen-
may not be appropriate. On the other hand, because the
dent variables, or interventions, sequentially on each of
intervention may exert a common causal influence on
the dependent variables, or outcomes, controlling for
these outcomes, the dependent variables are likely to be
all indirect effects through the prior dependent vari-
at least spuriously correlated with each other. In addi-
ables, or outcomes. Because the only theoretical guid-
tion, the multiple outcomes may also subsequently
ance required is a tentative specification of a meaningful
exert various causal influences on each other. Thus,
order among the dependent variables, we propose that
separate causal analyses for each of these dependent
this sequential method can be developed into an explor-
variables may also not be appropriate.
atory form of path analysis.
Fortunately, there exist several statistical proce-
The purpose of this paper is to explore the proper-
dures for the analysis of multiple correlated dependent
ties of sequential canonical analysis both in relation to
variables. One of these is structural equations modeling
those of confirmatory path analysis and simultaneous
(SEM), or confirmatory path analysis, in which the
canonical analysis, and as a potential model for explor-
hypothesized causal network between outcomes can be
atory path analysis. To do this, we will use an example
fully specified, estimated, and tested (Bentler 1989).
from insect psychobiology which has the advantage of
This method, however, requires the guidance of a
clearly illustrating these principles with very concrete,
strong causal theory which is often not available. An-"brass tacks", numbers.
other multivariate method is simultaneous canonical
analysis (SIMCA), which requires little theory. This
method, however, produces empirically-derived linear assortative mating in the
composites of dependent variables, or "canonical vari-
Jewel Wasp
ates", which are often difficult to interpret pragmati-
cally in terms of the concrete outcomes that we are The natural history of the Jewel wasp and the
interested in. The linear composites are statistically psychobiological rationale for this study is described in
defined, but are dependent upon the exact data set and Figueredo and Sage (2007). Basically, this experiment
so may not be replicated if the next study uses a slightlywas designed to test the effects of juvenile feeding
different set of variables. experience on adult mate choice in the Jewel wasp. To

FIGUEREDO, A. J., AND R. GORSUCH. 2007. ASSORTATIVE MATING IN THE JEWEL WASP: 2.
SEQUENTIAL CANONICAL ANALYSIS AS AN EXPLORATORY FORM OF PATH ANALYSIS.
JOURNAL OF THE ARIZONA-NEVADA ACADEMY OF SCIENCE 39(2):59-64.

This content downloaded from 148.209.209.83 on Fri, 21 Jul 2017 16:36:36 UTC
All use subject to http://about.jstor.org/terms
ASSORTATIVE MATING In THE JEWEL WASP 2 f FlGUEREDO AND GORSUCH 60

accurately determine the female choice of male wasps as sired by red-eyed, rather than white-eyed, males
offspring
sires for their offspring, genetically marked wasps were
(TRFO), and the total number of female offspring sired
used. Two recessive mutant genotypes were used, one
by blowfly-reared, rather than housefly-reared, males
for white eyes and one for red eyes, which, when crossed,
(TBFO). In the interests of brevity, these female off-
produce hybrids with brown eyes. Each female
springwas
were, respectively, referred to as "red-sired female
given a choice of two males, one of each eye-color
offspring", "white-sired female offspring", "blowfly-
genotype (which were counterbalanced for the particular
sired female offspring", and "housefly-sired female off-
prey species, either housefly or blowfly, that spring".
they hadOf course, the specified eye colors and juvenile
been fed on as larvae). Females of independently crossed
hosts referred to are those of the fathers, rather than those
eye-color genotype and juvenile feeding experience were
of the daughters, because all genetic hybrids had brown
also used. Because these different genotypes might them-
eyes, and all posttest offspring were reared on the larger
selves influence mate choice, the experimentblowfly
was de- hosts for convenience. Because paternal geno-
signed to test the relative influence of two factors, 1) juvenile hosts were also slightly correlated, by
types and
inherited eye-color genotype, and 2) conditioned forag-
the postoperative mortality noted above, eye color was
ing phenotype, on assortative mating in the Jewel wasp. as causally prior to juvenile host in these two
modeled
We wound up with a slight, though nonsignificant,
dependent variables, for the same reasons that they were
negative correlation between the independent in
variables,
the two corresponding independent variables.
female genotype (FG) and female juvenile host (FJ), due
to postoperative differential subject mortality. Such cor-
Alternative
related independent variables are easily modeled by Methods of Analysis
hierarchical partitioning of variance. To be conservative
The question then becomes, precisely how do we
in our testing of the effects of female host, weaccomplish
assigned the requisite feats of multivariate statistical
causal priority to female genotype. Moreover, we did not
control? To expedite our discussion, we report the results
hypothesize an interaction between these two of
factors.
the most uninformative statistical model first, namely,
The dependent variables, however, presented prob-
simultaneous canonical analysis (SIMCA). This analysis
lems that could not be handled by unaidedwas
multiple
performed using PROC CANCORR in SAS (SAS
regression. First, female Jewel wasps that feed Institute
on house- 1989). SIMCA constructed two pairs of canoni-
fly hosts are significantly less fertile than thosecal
that feed V 1 and W 1 , and V2 and W2, to represent our
variates,
on blowfly hosts (Figueredo 1987, 1989). Thismanifest
meansvariables. Pillai's Trace for the whole model
that we could expect significantly different total num-
was .21, F(10,158)= 1.856,/?= .055. Furthermore, VI
bers of offspring (TOT), regardless of paternity, from
was correlated to Wl by a canonical correlation of .42,
subjects reared on different juvenile hosts. Second,
which wasa statistically significant (F(10,156)= 1.89,/?=
variable number of Jewel wasp offspring remain
.05);imma-
V2, however, was correlated to W2 by a canonical
ture in a dormant condition, called diapause, forcorrelation
up to sixof . 17, which was statistically nonsignificant
months past the normal fourteen days (and, thus, the end
(F(4,79)= 0.61,/?= .66). Presumably, we could therefore
of our experiment). The total number of suchproceed
dormant
to interpret V 1 and W 1 , and forget about V2 and
immature offspring (TIO) is known to vary with
W2.certain
Fortunately, V 1 and W 1 were defined by the follow-
environmental factors, such as ambient temperatures andequations, using the standardized canonical
ing linear
photoperiods, but might also be affected in some un-
coefficients, as seen in Table 1 above.
known way by the local mating conditions experienced
Using SIMCA, we were also helpfully provided
by the mother. Because these dormant larvae had not
with yet
the resulting indirect correlations between W 1 and
developed eyes of either color, it was not possible to
the independent manifest variables (FG and FJ) and
determine their paternity, perhaps producingbetween nonran- VI and the dependent manifest variables (TOT
dom missing data. Third, male Jewel wasps are alwaysTBFO), as well as all the corresponding esti-
through
produced asexually by the mother and, therefore, mates have
for the nonsignificant pair of canonical variates,
no biological father at all. Thus, a female Jewel V2
waspandthat
W2, in both their raw and standardized forms.
does not mate with either of the males provided might
Unfortunately, in spite of this apparent wealth of infor-
produce a clutch of all male offspring. The totalmation,
number we still had some trouble understanding what any
of male offspring (TMO) could not be simply ignored in
of this meant for the results of our experiment. How did
this study because it might have indicated female rejec-
some linear combination of female genotype and female
tion of both males, and perhaps constituted more nonran-
juvenile host somehow produce some other linear com-
dom missing data. bination of the total numbers of various types of off-
Finally, the critical information in this study was
spring?
provided by the paternity of the female offspring, but had
A more informative approach was structural equa-
to be controlled for all of the above causally prior
tions modeling (SEM), or confirmatory path analysis.
offspring outcomes. Thus, the two dependent variables
This analysis was done using EQS (Bentler 1989). Be-
of principal interest were the total number of female

This content downloaded from 148.209.209.83 on Fri, 21 Jul 2017 16:36:36 UTC
All use subject to http://about.jstor.org/terms
61 ASSORTATIVE MATING In THE JEWEL WASP 2 4 FlGUEREDO AND GORSUCH

Table L Linear equations for dependent canonical variates VI and Wl.

VI = 0.864*FG - 0.413*FJ

Wl = - 1.015*TOT - 0.082*TIO + .046*TMO + .919*TRFO + 0.274*TBFO

Table 2. Standardized path coefficients for saturated structural model

TOT- - .105*FG +.218*FJ

TIO= .579*TOT - .117*FG +.004*FJ

TMO= .760*TOT - .022* TIC - .054*FG +.012*FJ

TRFO= .856*TOT - .255*TTO -.241*TMO + .263*FG - .090*FJ

TBFO= .753*TOT - .103*TlO - .307*TMO - .083TRFO 4 - .123*FG 4-.004*FJ

Table 3. Standardized path coefficients for restricted model with non-signi

TOT = 4-.230*FJ

TIO - .595*TOT

TMO - .756*TOT

TRFO= .646*TOT -.247*110 +.281*FG

TBFO - .630*TOT -.296*TMO

cause we had no strong causal theory

Such brazenly to
exploratory specify
use of SEM, however, the
is
technically
structural relations between theinappropriate.
multiple The probability
dependent of the model
vari-
chi-squared reported,
ables, a fully saturated structural modelfor example,
was no longer
run, repre-
freely
estimating every causal pathway possible
sented a prior probability between
under the every
null hypothesis, due
to the manyonly
dependent variable, specifying empirical the
model respecifications.
presumed Another
causal
priority. This method has also been
disadvantage referred
was that SEM to
did not provide for as the
perform-
cascade model in cognitive psychology
ing hierarchical (e.g.,
tests of significance, Mouyi
except through
2006, Demetriou et al. 2002). This
such nested model specification
comparisons, which required running pro-
multiple
duced a saturated structural causal models.
model Of course, suchreproducing
perfectly results could be
the data, with a residual obtained by running a series
chi-squared of of separate
zero, hierarchical
based on
zero degrees of freedom. regressions, as specified by the structural
The standardized path equations
coeffi-
cients are summarized in Table 2 above. written above. This piecemeal method had the disadvan-
tage of requiring a separate regression model for each of
Using the statistical tests reported in the EQS output,
the dependent variables and, thus, still providing no valid
we tested each of these path coefficients against zero. We
overall protective test of significance. This alternative
then eliminated nonsignificant causal pathways to de-
procedure was performed first using PROC GLM in SAS
velop a more restricted model in which all the remaining
(SAS Institute 1989), and then using UNIMULT(Gorsuch
pathways were statistically significant. Not surprisingly,
this model was statistically acceptable: the chi-squared1991, Gorsuch and Figueredo 1991), and, in spite of the
algorithmic differences in parameter estimation, both
was 1 1 .06 based on 14 degrees of freedom (p = .68). The
standardized path coefficients for this model are de- produced nearly identical results to each other. Further-
scribed in Table 3. more, this alternative procedure has the practical advan-

This content downloaded from 148.209.209.83 on Fri, 21 Jul 2017 16:36:36 UTC
All use subject to http://about.jstor.org/terms
ASSORTATIVE MATING In THE JEWEL WASP 2 4 FlGUEREDO AND GORSUCH 62

tage of not requiring any specialized software to per-

Table 5. Semipartial Correlation
form. Any standard software package (such as SAS or
Tests of Significance for Each Ind
SPSS) that can estimate hierarchical multiple Variable
regres-in the Multivariate Model
sions (using Type I sums of squares) can be used.
Variable
In spite of all these limitations, at least SEM pro- Eflect Size PBV F( 1,158) p
vided us with some clue as to what is going on. Control-
Dependent variable: TOT
ling for all significant prior dependent variables, it
appeared that there were only two significantFG causal
r=-.13 0.02 1.43 0.2
effects of our two independent variables. As indicated by
FJ r=.22 0.05 3.93 0.05
the positive effect of FJ on TOT, blowfly-reared females
produced more total offspring than housefly-reared fe-
Dependent variable: TIO
males, indicating their generally higher fertility, as was
FG r=-.14 0.03 2.62 0.1
expected. As indicated by the positive effect of FG on
TRFO, red-eyed females produced more red-sired FJ fe- r - .00 0 0 as.
male offspring than white-eyed females, indicating as-
Dependent
sortative mating by eye color. Contrary to the principal variable: TM
experimental hypothesis, however, blowfly-reared FG fe-r=-.08 0.02 1.25 0.3
males apparently produced no more blowfly-sired fe-
FJ r = .02 0 0.05 as.
male offspring than housefly-reared females, indicating
no assortative mating by juvenile host. This was indi-
Dependent variable: TRFO
cated by the lack of a significant causal effect of FJ on
TBFO. FG r=.31 0.14 13 <.001

The third multivariate method used on this data was FJ r=-.O9 0.01 1.12 0.3
sequential canonical analysis (SEQCA). This analysis
was performed using UNIMULT (Gorsuch 1991). All Dependent variable: TBFO
that was required for running this model was the speci- FG r=.12 0.02 1.44 0.2
fication of two hierarchies of causal priority, one for the
FJ r=.01 0 0.02 as.
two independent variables and another for the five de-
pendent variables. The Pillai-Bartlett V for the whole
model was .21, F(10,158)= 1.86,/?= .05, which was the
same, within rounding error, as the corresponding SIMCA
results. In addition, we also obtained overall tests of the
Note that the two statis
proportions of the variance of each of the five dependent
the independent variabl
variables accounted for by the linear combinations of the
were the same as those i
two independent variables, which are shown in Table 4.
for all the prior depend
SEQCA also provided the following hierarchical
regression coefficients
significance tests for the separate effects of each of the
very similar to those pr
two independent variables on each of the five dependent
variables and these are shown in Table 5. become somewhat highe
down the causal hierarch
slightly higher paramet
Table 4. Multiple Correlations and Associated Tests property of the curren
of Significance for Each Dependent Variable in the SEQCA, not an artifact o
Multivariate Model.
algorithms. Recall that w
gressions for the struct
triable Efflect Size PBV F(2,158) p
formed by UNIMULT,
TOT R = .25 0.06 2.68 0.07 estimates were obtaine
systematically higher be
TIO R=.14 0.03 1.31 0.3 what Cohen and Cohen ( 1
rather than the "semipa
TMO R=.08 0.02 0.65 us.
by SEM. In the current
residualized the denomin
TRFO R=.32 0.15 7.06 < 0.005
explained (including the e
TBFO R=.12 0.02 0.73 us. or portion of variance ac
sequential dependent var
affected significance tes
mation, by increasing th

This content downloaded from 148.209.209.83 on Fri, 21 Jul 2017 16:36:36 UTC
All use subject to http://about.jstor.org/terms
63 ASSORTATIVE MATING In THE JEWEL WASP 2 4 FlGUEREDO AND GORSUCH

criterion
Table 6. Standardized Regression Coefficients for variables are analyzed sequ
Each Independent Variable in the Multivariate
to a hypothesized causal order. These
Model rion variables can be entered sequenti
of multiple regression equations with e
TOT - -.105*FG +.217*FJ
prior criterion variable entered as the
TIO - -.144*FG + .004*FJ the next, as we did in the SEM cascad
above. Each successive dependent var
TMO = -.080*FG + .016*FJ
dicted from an initial set of ordered p
TRFO = .300*FG - .092*FJ each time entering the immediately pr
variable hierarchically as the first pred
TBFO = J21*FG + .012*FJ ing all the ordered predictors from th
sion equation. Thus, each successive
all of the preceding dependent variable
order to statistically control for any in

parison with those obtained might be transmitted

by SEM, the SEQCA through
stan- them. W
cal scheme, as with SEQCA, the estima
dardized estimates are reported in Table 6.
The UNIMULT implementationpredictor
ofisSEQCA
limiteddoes
to its direct effe
successive
not explicitly provide estimates dependent
of the sequential variables. The g
effects
this system
of the dependent variables upon each other. of hierarchical
Doing so multiple
shown in Table 7 below.
would provide an alternative implementation of SEQCA
that would be fully equivalent to an exploratory path
Tableestimates
analysis by explicitly including 7. General format offorsequential
multiple dependent
criterion variables analyzed
effects between dependent variables. Such a model would sequentially according
to a hypothesized causal order
restore the status of hierarchically residualized correla-
tions as the "semipartials" (Cohen and Cohen 1983)
Y4= flX, +j32X2 +J3X3
typically estimated in SEM. This new model would have
Y= /?4Y4 4-flX,
the added advantages of both hierarchical +J3.1X2 +£X3of
partitioning
variance and protective overall tests of significance,
such as the Pillai-Bartlett V Y= j35Y5 +yg4Y4
statistic, -lyg.X, +j32X2
so important to+J3ji}
exploratory data analysis. It would also help legitimize
an alternative path analytic model for empirically-as-
sisted theory development, instead of perpetuating the
Where X,, X1 and
widespread abuse of available structural X3 are thethat
models orderedare
predictor
clearly intended exclusively for theory confirmation. As criterion
variables and Y4, Y5, and Y6 are the ordered
variables,
it stands, as illustrated by our numberedwasp
Jewel consecutively after the predictors to
example,
avoid confusion among
SEQCA represents a useful diagnostic tool for isolating the subscripts. What is lost by
using this method, as
and identifying the direct effects of independent vari-opposed to SEQCA, is the protec-
ables on a multiplicity of correlated dependent variables, by this
tive overall test of significance. What is gained
method, as of
if not for estimating the magnitudes opposed to SEQCA,
these is obtaining
effects in the estimates of
the sequential effects
conventional way. In either case, parameter estimation among dependent variables. Thus,
should never be based on the this is superior to saturated
initially merely includingmodel
all the prior depen-
dent variables
used for data exploration because the as "covariates".inclusion
probable What is gained as op-
of nonsignificant variables posed SEM is the ability
substantially to perform
reduces thehierarchical
efficiency of estimation. partitioning of variance and hypothesis testing, as in
Superior estimates can be SEQCA.
obtainedIn addition,
byone avoids compromising
running the the purely
confirmatory nature of the SEM tests of whole-model
various final structural equations, as respecified by pro-
goodness-of-fit.
tected and hierarchical significance testing, as separate
multiple regressions after the initial exploration of the
data, as was briefly describedSubstantive
above. and Methodological
One potential disadvantage of SEQCA in relation to
Conclusions
the alternative methods is that it requires the specialized
UNIMULT software (Gorsuch 1991) To finishto perform.
our story, Armed
our study revealed no evidence
with the basic theory behindof SEQCA, however,
assortative mating one experience
by juvenile feeding may in
use the more commonly available SAS software (SAS
the Jewel wasp. It is therefore unlikely that differential
Institute 1989), or even SPSS, tofeeding
juvenile performexperience acanseries of
lead to sympatric spe-
hierarchical regressions in which multiple dependent
ciation in this particular species by either of the mecha-

This content downloaded from 148.209.209.83 on Fri, 21 Jul 2017 16:36:36 UTC
All use subject to http://about.jstor.org/terms
ASSORTATIVE MATING In THE JEWEL WASP 2 4 FlGUEREDO AND GORSUCH 64

nisms that have been proposed. The requisite SAS reproduc-

Institute, Inc. 1989. SAS Language and Procedures:
tive isolation can apparently not be produced by differ-
Usage, Version 6, First Edition. SAS Institute, Cary, NC.
ential juvenile feeding experience either indirectly by
permanently differentiated foraging behavior or directly
by differentially imprinted sexual preferences. Instead,
assortative mating was shown to be greatly influenced by
genetic markers, such as eye color. Converging evidence
for these conclusions was variously obtained by hierar-
chical multiple regression, conventional path analysis,
and sequential canonical analysis. Finally, the method-
ological implications of these findings for research are as
follows. It was shown how the separate direct effects of
the experimental manipulations on a set of seemingly
hopelessly interdependent outcomes could be readily
discriminated by sequential canonical analysis, produc-
ing results very similar to those obtainable by confirma-
tory path analysis without requiring the stronger
theoretical assumptions of that model. This was done
directly on the variables of practical interest, without
altering the basic nature of the research question by
constructing either inappropriate common factors or
uninterpretable canonical variates. The statistical results
were readily interpretable and directly relevant to the
experimental hypotheses that motivated the study.

Literature Cited
Bentler, P. M. 1989. EQS, Structural Equations, Program
Manual, Program Version 3.0. BMDP Statistical Soft-
ware, Inc., Los Angeles, CA.
Cohen, J., and P. Cohen. 1983. Applied multiple regression/
correlation analysis for the behavioral sciences . Lawrence
Erlbaum, Hillsdale, NJ.
Demetriou, A., C. Christou, G. Spanoudis, and M. Platsidou.
2002. The development of mental processing: Efficiency,
working memory, and thinking. Monographs of the Soci-
ety of Research in Child Development 67(1, Serial No.
268): 1-154.
Figueredo, A. J. 1987. The statistical measurement, develop-
mental mechanisms, and adaptive ecological functions of
conditioned host selection in the par asitoid Jewel wasp.
Dissertation, University of California, Riverside, CA.
Figueredo, A. J. 1989. Host-cue information processing by
foraging Jewel wasps: Response-bias tracking of ex-
pected host-species encounter rates, not modified
profitabilities. Psychobiology 17:435-444.
Figueredo, A. J., and R. S. S. Sage. 2007. Assortative mating
in the Jewel wasp: 1. Female matching of eye-color
genotype, not host-feeding phenotype. Journal of the
Arizona-Nevada Academy of Science, 39:51-57.
Gorsuch, R. L. 199 1 . UniMult: For univariate and multivari-
ate data analysis. UniMult, Altadena, CA.
Gorsuch, R. L., and A. J. Figueredo. 1 99 1 . Sequential canoni-
cal analysis as an exploratory form of path analysis. Paper.
American Evaluation Association Conference, Chicago,
Illinois.

Mouyi, A. 2006. Untangling the cognitive processes web. Paper.

Seventh Annual Conference of the International Society for
Intelligence Research, San Francisco, California.

This content downloaded from 148.209.209.83 on Fri, 21 Jul 2017 16:36:36 UTC
All use subject to http://about.jstor.org/terms