Content uploaded by Norman B Schmidt
Author content
All content in this area was uploaded by Norman B Schmidt
Content may be subject to copyright.
Taxometric and Factor Analytic Models of Anxiety Sensitivity: Integrating
Approaches to Latent Structural Research
Amit Bernstein and Michael J. Zvolensky
University of Vermont
Peter J. Norton
University of Houston
Norman B. Schmidt
Florida State University
Steven Taylor
University of British Columbia
John P. Forsyth
University at Albany, State University of New York
Sarah F. Lewis
Marshall University and Meridian Behavioral Health Services
Matthew T. Feldner and Ellen W. Leen-Feldner
University of Arkansas, Fayetteville
Sherry H. Stewart
Dalhousie University
Brian Cox
University of Manitoba
This study represents an effort to better understand the latent structure of anxiety sensitivity (AS), as
indexed by the 16-item Anxiety Sensitivity Index (ASI; S. Reiss, R. A. Peterson, M. Gursky, & R. J.
McNally, 1986), by using taxometric and factor-analytic approaches in an integrative manner. Taxo-
metric analyses indicated that AS has a taxonic latent class structure (i.e., a dichotomous latent class
structure) in a large sample of North American adults (N ⫽ 2,515). As predicted, confirmatory factor
analyses indicated that a multidimensional 3-factor model of AS provided a good fit for the AS
complement class (normative or low-risk form) but not the AS taxon class (high-risk form). Exploratory
factor analytic results suggested that the AS taxon may demonstrate a unique, unidimensional factor
solution, though there are alternative indications that it may be characterized by a 2-factor solution.
Findings suggest that the latent structural nature of AS can be conceptualized as a taxonic latent class
structure composed of 2 types or forms of AS, each of these forms characterized by its own unique latent
continuity and dimensional structure.
Keywords: anxiety sensitivity, taxometrics, factor analysis, anxiety disorders
Anxiety sensitivity (AS) is defined as the fear of bodily sensa-
tions related to anxiety, which arises from beliefs that the sensa-
tions have harmful consequences (Reiss & McNally, 1985). With
the recognition of the importance of AS as a cognitive vulnerabil-
ity factor for anxiety problems, a concerted scientific effort has
focused on better understanding the latent structure of the con-
struct. The basic premise of this work is that by explicating the
latent structure of AS, theoretical models can be refined, leading to
advances in research and clinical intervention (Ruscio & Ruscio,
2002).
There have thus far been two principal approaches to under-
standing the latent structure of AS. The first and most common
Amit Bernstein and Michael J. Zvolensky, Department of Psychol-
ogy, University of Vermont; Peter J. Norton, Department of Psychol-
ogy, University of Houston; Norman B. Schmidt, Department of Psy-
chology, Florida State University; Steven Taylor, Department of
Psychiatry, University of British Columbia, British Columbia, Canada;
John P. Forsyth, Department of Psychology, University at Albany, State
University of New York; Sarah F. Lewis, Department of Psychology,
Marshall University, and Meridian Behavioral Health Services, Sylva,
North Carolina; Matthew T. Feldner and Ellen W. Leen-Feldner, De-
partment of Psychology, University of Arkansas, Fayetteville; Sherry
H. Stewart, Department of Psychiatry, Dalhousie University, Halifax,
Nova Scotia, Canada; Brian Cox, Department of Psychiatry, University
of Manitoba, Winnipeg, Manitoba, Canada.
This article was supported by National Institute on Drug Abuse Grants 1
R01 DA018734-01A1, R21 DA016227-01, and R03 DA016566-01A2 to
Michael J. Zvolensky; National Research Service Award Predoctoral Fellow-
ship F31 MH073205-01 to Amit Bernstein; and National Institute of Mental
Health Grant R21 MH62056 to Norman B. Schmidt.
Correspondence concerning this article should be addressed to Michael J.
Zvolensky or Amit Bernstein, Department of Psychology, University of Ver-
mont, 2 Colchester Avenue, John Dewey Hall, Burlington, VT 05405-0134.
E-mail: Michael.Zvolensky@uvm.edu or Amit.Bernstein@uvm.edu
Psychological Assessment Copyright 2007 by the American Psychological Association
2007, Vol. 19, No. 1, 74 – 87 1040-3590/07/$12.00 DOI: 10.1037/1040-3590.19.1.74
74
type of latent structural research has involved exploratory and
confirmatory factor analytic approaches (e.g., Silverman, Goed-
hart, Barrett, & Turner, 2003). Research on adults across diverse
populations suggests that AS, as indexed by both the 16-item
Anxiety Sensitivity Index (ASI; Reiss, Peterson, Gursky, & Mc-
Nally, 1986) and the 36-item Anxiety Sensitivity Index—Revised
(ASI–R; Taylor & Cox, 1998), is multidimensional and hierarchi-
cal in nature. Specifically, AS is composed of a higher order factor
with a number of specific lower order facets (e.g., Zinbarg, Brown,
& Barlow, 1997; Zinbarg, Mohlman, & Hong, 1999). Other studies
have produced conceptually similar results using the 18-item
Childhood Anxiety Sensitivity Index (CASI; Silverman, Fleisig,
Rabian, & Peterson, 1991) among youth (e.g., Silverman et al.,
2003).
Taxometrics is an additional type of latent structural research
approach that has been applied to the study of AS. Taxometrics
entails a set of statistical procedures used to determine whether the
latent structure of a construct is continuous or categorical (i.e.,
taxonic; Schmidt, Kotov, & Joiner, 2004). Though taxometrics
have been less frequently used in psychopathology research com-
pared with other latent structural approaches, such as factor anal-
ysis, they similarly serve to explicate the latent structure of psy-
chological constructs. Although similar in general purpose, there
are many direct and unique clinical implications that arise from the
study of whether a construct is continuous or categorical and,
hence, from the use of taxometric procedures. As one illustrative
example, if AS is dimensional, then it may be best to use scores
from an established scale to index continuous clinical improve-
ment during preventative or treatment programs. However, if AS is
taxonic in nature, then it may prove to be more useful to gauge
change in this construct by assessing whether an individual moves
from a vulnerability-conferring form to a normative form.
It is natural to ask why one would expect AS to vary as a
function of a latent class. First, foremost in the theoretically based
hypothesis of AS taxonicity is the important consideration that AS,
at its core, is an adaptive factor and process (Barlow, 2002). A
sensitivity to anxiety-related cues may serve to alert the organism
to threatening internal or external cues. It is unclear how an
adaptive process, normally distributed in the population, could
therefore confer vulnerability for anxiety psychopathology, such as
panic. It is, however, possible that certain aversive emotional
learning histories or gene– environmental interactions may quali-
tatively change the adaptive nature of AS to become maladaptive
and thereby facilitate the developmental bifurcation of AS into
unique classes and taxonic trajectories. Indeed, basic research on
associative learning of anxiety suggests that once such maladap-
tive emotional associative learning occurs, it is often highly resis-
tant to extinction (Bouton, 2004). This resistance to extinction may
thereby facilitate and maintain the taxonic split between adaptive
and maladaptive latent forms of AS.
Consistent with predictions derived from this type of theoretical
perspective, studies using the ASI with nonclinical adults suggest
that AS is taxonic (i.e., has a dichotomous latent class structure
with a high-risk form and a normative or low-risk form) rather than
dimensional (i.e., single latent continuous structure; Bernstein,
Leen-Feldner, Kotov, Schmidt, & Zvolensky, 2006; Schmidt, Ko-
tov, Lerew, Joiner, & Ialongo, 2005; Zvolensky, Forsyth, Bern-
stein, & Leen-Feldner, in press). Using a sample of individuals
with panic disorder and nonclinical controls matched on demo-
graphic characteristics, Taylor, Rabian, and Fedoroff (1999) found
equivocal evidence regarding the latent continuous or taxonic
structure of AS. Differences between Taylor et al.’s (1999) inves-
tigation and the other AS taxometric research reports among adults
may be due to a host of unique methodological factors specific to
this particular investigation, including sampling for bimodality
(i.e., systematically selecting and then mixing clinical and non-
clinical participants in the same sample), indicator selection crite-
ria, and statistical procedures. Confidence in the taxonic structure
of AS is strengthened by other work using the ASI–R (Taylor &
Cox, 1998) among adults across six different samples from sepa-
rate countries (Bernstein, Zvolensky, Kotov, et al., 2006) and two
large independent samples of youth using the CASI (Bernstein,
Zvolensky, Stewart, Comeau, & Leen-Feldner, 2006; Bernstein,
Zvolensky, Weems, Stickle, & Leen-Feldner, 2005). Collectively,
available data suggest that there is a large degree of consistent
evidence that AS may be taxonic.
Although the replicated taxonic structure of AS is increasingly
compelling, there also is evidence testifying to its construct valid-
ity. The AS taxon, specifically, has demonstrated incremental
validity relative to (a) the well-established continuous index of the
construct and (b) negative affectivity in the prediction of panic
symptoms and processes (Bernstein, Leen-Feldner, et al., 2006;
Schmidt et al., 2005) as well as posttraumatic stress symptoms
(Bernstein, Zvolensky, Feldner, Lewis, Fauber, et al., 2005; Bern-
stein, Zvolensky, Feldner, Lewis, & Leen-Feldner, 2005). For
example, Zvolensky et al. (in press) found that the items measuring
the AS taxon accounted for significant variance above and beyond
that accounted for by the 16-item ASI total score in terms of scores
on bodily vigilance and perceived controllability of anxiety-related
events. Moreover, after accounting for variance explained by the
full-scale ASI total score, the total score for the ASI items not
indexing the AS taxon was associated with significant variance in
these same dependent measures but in the opposite direction
predicted by contemporary anxiety vulnerability theory (Zvolen-
sky et al., in press). These data suggest that it may be practically
and theoretically useful to begin to conceptualize one form of AS
as a normative form of AS and the other as a vulnerability-
conferring form of AS (i.e., conferring risk for specific kinds of
anxiety psychopathology). To be conservative, however, in the
present article we refer to the AS complement class as the “nor-
mative form” or “low-risk form” of AS and the taxon as the
“high-risk form” of AS.
A logical next research step is to use taxometric and factor
analytic approaches in an integrative fashion to further examine
the nature of the AS construct. To understand why an integration
of these approaches is important, it is useful to briefly highlight
core assumptions underlying each statistical approach. A central
assumption of factor analysis and structural equation modeling
more broadly is population homogeneity (Thurstone, 1935). This
assumption necessitates that the studied sample be selected from a
single population such that a single covariance matrix fully cap-
tures the relations among the variables in that sample. In fact, to
account for observed population heterogeneity, Jo¨reskog (1971)
extended structural equation modeling confirmatory factor analy-
sis (CFA) to multiple groups, each with its own mean vector and
covariance matrix (Bauer & Curran, 2004). A problem is encoun-
tered, however, when population heterogeneity is at the latent
taxonic level rather than at the manifest level—an issue we return
75
LATENT STRUCTURE OF ANXIETY SENSITIVITY
to later in this article. A core assumption of the general covariance
mixture theorem underlying coherent cut kinetic taxometric pro-
cedures (Waller & Meehl, 1998) and similar latent class variable
analytic approaches (Bauer & Curran, 2004) is of within-class
independence or zero within-class between-variable covariance
(i.e., variables do not covary within each class but only among the
admixed classes). This issue is particularly noteworthy with re-
spect to AS, as multiple studies of AS have identified within-class
nuisance correlations that, although within tolerable limits for
taxometric analysis (Waller & Meehl, 1998), nevertheless are not
fully within-class independent and so demonstrate some within-
class between-variable covariance.
Thus, it is important to note that taxonicity does not imply
absence of latent within-class quantitative gradations or hierarchi-
cal structure (Pickles & Angold, 2003). When taxonicity is de-
tected, a nonarbitrary qualitative difference may be arbitrarily
superimposed on within-class quantitative variability (Schmidt et
al., 2004). In other words, just because taxonicity is detected and
nonarbitrary qualitative variability is identified, such a finding
does not at all reflect on concurrent existence or systematic mean-
ingfulness of within-class quantitative variability. To help ensure
that further theoretical development of AS validly reflects its latent
structure and so is clinically meaningful, both these latent quali-
tative and quantitative forms of variability need to be explicitly
recognized (Bernstein, Zvolensky, Kotov, et al., 2006; Pickles &
Angold, 2003; Waller & Meehl, 1998).
With this background, it should not be assumed that AS would
have only a single latent factor structure for both the complement
class and the taxon class. Past studies have not distinguished
between latent taxa but rather have used the single manifest
distribution of ASI scores, assuming no latent population hetero-
geneity. Therefore, there has been no explicit test of the factor
structure of AS after the conjectured taxonic (i.e., normative and
high-risk) forms of the construct have been distinguished from one
another. This is important because in the absence of such a test, we
have only limited knowledge about the latent structure of AS and,
specifically, no empirical knowledge regarding the latent nature
within each AS class. Available theory and research suggest that
one should not expect both the AS taxon and complement class to
have an identical factor structure. Indeed, there are three key
reasons to hypothesize that the AS taxon and complement class
would be expected to demonstrate distinct factor structures.
First, differential base rates between the taxonic latent AS
classes may be expected to have systematically affected the ob-
served factor structures among mixed-class samples (i.e., samples
composed of admixed complement and taxon classes) in previous
factor analytic studies of AS. The complement class base rate
(range across studies: p ⫽ .82–.89) is much larger than the taxon
class base rate (range across studies: p ⫽ .11–.18; Bernstein,
Leen-Feldner, et al., 2006; Bernstein, Zvolensky, Feldner, Lewis,
& Leen-Feldner, 2005; Bernstein, Zvolensky, Kotov, et al., 2006;
Bernstein, Zvolensky, Weems, et al., 2005; Schmidt et al., 2005;
Zvolensky et al., in press). As the base rate of the complement
class is significantly larger than the taxon class, the observed factor
structure of AS may, by statistical artifact, reflect the much larger
complement class when the latent classes are mixed. Indeed,
previous factor analytic studies have universally failed to distin-
guish between the taxometrically indicated taxonic classes (i.e.,
violated assumption of population heterogeneity) and have exclu-
sively studied mixed-class samples. After separating the latent AS
classes based on taxometric analyses, the well-replicated multidi-
mensional factor structure would be expected only for the larger
base rate complement class but not the smaller base rate taxon
class. This naturally is important to know because it would help
provide an empirical test of whether the well-established AS factor
structure only validly reflects the dimensional structure of the
complement class and not the taxon class.
Second, on the basis of the documented nomological differences
between the taxon and complement classes, the AS taxon could be
expected to demonstrate a distinct factor structure from the mul-
tidimensional structure expected among the complement class. The
AS taxon is presumably composed of a group of individuals
qualitatively distinct from those individuals who belong to the
complement AS class in terms of their vulnerability for certain
forms of anxiety psychopathology. The AS taxon has demon-
strated differential, unique associations with anxiety-relevant cri-
terion variables relative to the complement AS class (e.g., Bern-
stein, Zvolensky, Feldner, Lewis, & Leen-Feldner, 2005). If the
AS taxon and complement classes are systematically meaningful
in terms of their differential nomological networks, then it is
neither theoretically nor statistically plausible that these discrete
forms of AS will demonstrate the same underlying structural forms
and therefore factor structures.
Finally, the AS taxon and complement class could be expected
to demonstrate distinct factor structures for the same conceptual
reason that taxometric analyses of AS would, in the first place, be
expected to reveal latent taxonic classes rather than a single latent
continuous variable structure. Contemporary conceptual models
posit that adverse learning experiences and biological predisposi-
tions play formative roles in shaping individual susceptibility to
anxiety problems (Barlow, 2002). Here, theory and research sug-
gest that a variety of maladaptive associative learning experiences
may be an important factor in contributing to differing levels of AS
(Stewart et al., 2001) and could be related to the putative bifurca-
tion of qualitatively distinct developmental trajectories with re-
spect to AS and anxiety psychopathology vulnerability (see
Beauchaine, 2003). To the extent that negative life experiences,
and perhaps gene– environmental interactions (Schmidt, Lerew, &
Joiner, 2000), shape clinically significant levels of AS, as sup-
ported by extant empirical work (Stewart et al., 2001), it is possible
that such sensitivity will be a distinguishing cognitive-based fea-
ture associated with the pathogenesis and maintenance of a tax-
onic, anxiety-vulnerability-conferring, developmental trajectory.
That is, if an individual is exposed to any number of aversive
learning experiences, then it is likely that this individual may
develop higher, and perhaps structurally distinct, sensitivity to
anxiety states (a taxonic or high-risk form of AS). By contrast,
those individuals not exposed to such aversive learning experi-
ences may not develop in a similar manner. In this sense, the
processes underlying AS change may be similar but the type of
experiences fostering such change (e.g., exposure to aversive
events) may be distinct among those in an AS taxon class and
normative complement class. Thus, it is possible that for AS, an
adaptive cognitive process at its etiological core, to become mal-
adaptive and thereby confer risk for anxiety psychopathology, it
needs to change qualitatively or categorically over time. For this
reason, we might not expect individuals in the taxon class to be
sensitive in the same way or to the same types of cues as those with
76
BERNSTEIN ET AL.
the normative form of AS. Accordingly, the AS complement and
taxon class members could be expected to show different latent
natures, including factor structures on instruments, such as the
16-item ASI, designed to tap sensitivity to anxiety-related sensa-
tions.
The overall aim of this investigation was to use taxometric and
factor analytic approaches in an integrative fashion to study the
latent structure of AS, as indexed by the ASI. It was predicted that
(a) taxometric analyses would replicate existing findings that AS is
taxonic among a large multisite sample of young adults, (b) CFA
using structural equation modeling (SEM) would demonstrate that
the preexisting hierarchical three-factor model of AS (Zinbarg et
al., 1997, 1999) would fit the ASI data for the complement class,
and (c) this same factor solution would not fit the ASI data among
the taxon class of AS. Rather, we expected (d) that an exploratory
factor analysis would identify a different factor structure among
the taxon class. As the latter hypothesis is novel to the AS
literature, no empirical findings or theory were available to guide
specific hypotheses regarding the precise nature of the AS factor
structure among the taxon class.
Method
Participants
Data for the present investigation were collected across seven
separate universities, reflecting different regions of North America
(Albany, New York; Columbus, Ohio; Houston, Texas; Hunting-
ton, West Virginia; and British Columbia, Manitoba, and Nova
Scotia, Canada). In general, data were collected in a similar man-
ner across each of the sites, with an attempt to target the same type
of population. Specifically, large samples composed of non-
treatment-seeking persons. This tactic was used to protect against
purposively selecting for bimodality. These samples were collec-
tively composed of seven aggregated samples of individuals, with
a total sample size of 2,515 individuals (1,566 female; mean age ⫽
20.0 years, SD ⫽ 4.7); the large majority were university students
(approximately 85%), and a small minority (15%) were
community-based participants. In terms of ethnicity, 1,186 (47%)
participants were Caucasian, 234 (9.3%) were African American,
221 (8.8%) were Asian American, 193 (7.7%) were Hispanic, 3
(0.1%) were Native American–Alaskan, 119 (4.7%) self-identified
as “other,” and the remainder did not specify their ethnicity.
Measures
The ASI (Reiss et al., 1986) is a 16-item measure on which
respondents indicate on a 5-point Likert-type scale (0 ⫽ very little,
4 ⫽ very much) the degree to which they are concerned about
possible negative consequences of anxiety-related sensations. The
ASI has high levels of internal consistency and good test–retest
reliability (Peterson & Plehn, 1999).
Procedure
Recruitment of participants was generally similar across sites.
After providing written informed consent, participants completed
the assessment anonymously in an individual or group format.
Participants were debriefed as to study objectives prior to their
departure. All participants received either monetary compensation
(community members of the sample) or course credit (student
members of the sample) for their efforts.
Analytic Approach: Taxometric Analyses
Indicator selection. Three theoretically and quantitatively
based item domains of the 16-item ASI were identified and used to
build item-parcel manifest indicators of AS: (a) physical concerns,
(b) mental incapacitation concerns (heretofore described as “psy-
chological concerns”), and (c) fears of publicly observable anxiety
reactions (heretofore described as “social concerns;” Bernstein,
Leen-Feldner, et al., 2006). Each of the three indicators bore item
compositions identical to a well-established factor structure of the
ASI (i.e., Physical Concerns ⫽ Items 3, 4, 6, 8, 9, 10, 11, 14;
Psychological Concerns ⫽ Items 2, 12, 15, 16; and Social Con-
cerns ⫽ Items 1, 5, 7, 13; Zinbarg et al., 1997, 1999). The three
indicators are consistent with the dominant theoretical model of
AS (McNally, 2002). Although there are various means of select-
ing manifest indicators for taxometric procedures, factor analyti-
cally derived indicators may produce optimal taxometric results.
Factor analytically derived indicators limit artifactual nuisance
correlations, provide meaningful sample indicators from all facets
of a construct, and optimize internal consistency and distinctive-
ness of the indicators (Schmidt et al., 2004).
MAXCOV-HITMAX. First, the maximum covariance proce-
dure (MAXCOV; Ruscio, 2004) was conducted. On a rotating
basis, each pair of indicators served as output variables and the
third remaining indicator formed the input variable. To exhaust all
possible bivariate combinations of output variables, three MAX-
COV plots were generated. The MAXCOV analysis was con-
ducted in line with recent recommended guidelines (Ruscio &
Ruscio, 2004). Specifically, equal-sized nonoverlapping intervals
were used to divide each input variable. To concurrently limit
sampling error and to ensure a sufficient number of intervals to
detect the expected low base rate taxon (Schmidt et al., 2005), we
chose 125 fixed-size intervals a priori because this value assigns at
least 20 cases to each window (Schmidt et al., 2004). Fifty internal
replications were used to divide cases by repeatedly resorting cases
along the input indicator at random. At each replication, the
correlation of the two output indicators within each interval was
calculated, and these calculations were averaged across all repli-
cations. These replications are particularly useful when MAXCOV
input variables are divided by fixed intervals and placed arbitrarily
between equal-scoring cases. Thus, use of these replications is
intended to minimize sampling error and bolster the reliability of
the results (Ruscio, 2004).
Covariance plot “nose count” (i.e., the number of taxonic,
ambiguous, and nontaxonic covariance plots across indicators) and
the coherency of the standard deviation of the base rate estimates
served as internal consistency tests. Consistency tests are intended
to rule out the detection of a pseudotaxonic or pseudodimensional
conclusion (Haslam & Kim, 2002). The specific criteria used
herein to conduct the covariance plot “nose count” were based on
the guidelines provided by Waller and Meehl (1998) and standards
used to interpret the plots in previous AS research (e.g., Bernstein,
Zvolensky, Weems, et al., 2005). In addition, only peaked MAX-
COV plots were interpreted as evidence of latent taxonicity. Al-
though in some instances right-end cusped MAXCOV plots also
may represent latent taxonicity (Schmidt et al., 2004), such cusped
77
LATENT STRUCTURE OF ANXIETY SENSITIVITY
plots may equally likely represent positively skewed dimensional
data; thus, cusped plots were interpreted conservatively in this
investigation as ambiguous rather than as evidence of taxonicity.
Plots that could not be clearly discerned as taxonic or nontaxonic
were labeled ambiguous, and all other plots were labeled nontax-
onic.
If the MAXCOV analysis indicated taxonic latent structure, then
individual participants were assigned to the taxon and complement
classes. The validity of the manifest indicators to discriminate be-
tween complement and taxon class members was estimated by mea-
suring the effect size of each indicator in discriminating between
complement and taxon class membership. The level of nuisance
correlation was used to gauge the accuracy of the estimated parame-
ters. When data are taxonic, the level of nuisance correlation may be
used as an index of the accuracy of parameter estimates because it can
be used as a marker of the sensitivity and specificity of latent class
membership assignment (Schmidt et al., 2004).
MAXEIG-HITMAX. The maximum eigenvalue (MAXEIG) was
used as an external consistency test of the MAXCOV procedure
(Ruscio, 2004). External consistency tests are premised on the same
notion as internal consistency tests and are specifically intended to
rule out pseudotaxa or pseudodimensions that result from an artifact
of a taxometric procedure under particular methodological or statis-
tical conditions (Schmidt et al., 2004; Waller & Meehl, 1998).
MAXEIG is a multivariate extension of MAXCOV. In contrast to
MAXCOV, which calculates bivariate correlations between pairs of
output indicators at varying levels of an input indicator, MAXEIG
derives the maximum eigenvalues of the covariance matrix of a set of
output indicators at varying levels of an input indicator.
Although MAXCOV and MAXEIG are mathematically similar
insofar as correlations and eigenvalues are alike, they also are
systematically different in a number of important respects (Waller
& Meehl, 1998), which justify their concurrent utility in the
present study. For example, unlike MAXCOV, MAXEIG divides
the input variables with overlapping intervals. Overlapping
windows enable investigators to detect very low base rate taxa.
MAXEIG also affords a powerful and unique internal consistency
test. This “inchworm” test is achieved by repeatedly conducting
the MAXEIG analysis while systematically increasing the number
of overlapping windows, thereby systematically decreasing the
size of the subsamples used to calculate maximum eigenvalues in
each overlapping window (Waller & Meehl, 1998). Plots of tax-
onic data yield an increasingly better defined unimodal peak, as the
number of windows is increased, resembling the head of an inch-
worm. In the instance of dimensional data, in contrast, the sys-
tematic increase of the number of overlapping windows will pro-
duce plots that do not systematically peak (e.g., flat or cusped
plots). Fifty internal replications were used to divide cases into
overlapping windows by repeatedly resorting cases along the input
indicator at random. At each replication, the maximum eigenvalue
of the two output indicators within each window was calculated,
and these calculations were averaged across all replications. In the
present investigation, we conducted the inchworm test by gener-
ating four sets of MAXEIG plots, from 500 windows to 1,250
windows, in 250-window increments. These specific values were
selected on an a priori basis to produce four sets of MAXEIG
analyses that assign at least 20 cases per window in the final
inchworm test, so that a sufficiently wide range of windows is used
to detect what was expected to be a low base rate taxon while
guarding against sampling error. In addition, covariance plot “nose
count” and the coherency of the standard deviation of the base rate
estimates served as additional internal consistency tests. Finally, as
described for the MAXCOV procedure, the MAXEIG taxon base
rate, indicator validity, and nuisance correlations were estimated.
Parameter-matched Monte Carlo simulations. Following re-
cently recommended guidelines for taxometric analyses (Ruscio,
2004) parameter-matched Monte Carlo simulated dimensional and
taxonic data were derived. Simulated dimensional data were re-
produced via bootstrap and thereby matched to the number of
indicators, sample size, observed indicator correlation matrix, and
distributions of all indicators, including their skew, kurtosis, and
number of levels. Simulated taxonic data were similarly matched
to the parameters of the research data. Ten separate sets of simu-
lated dimensional and taxonic data were derived to bolster the
reliability and precision of the simulated data and the taxometric
analyses conducted with these simulated data (Ruscio, 2004).
Monte Carlo simulated data were derived for two primary
purposes. First, for the purpose of a priori suitability testing
(Ruscio & Ruscio, 2004), and as conducted in previous investiga-
tions (Bernstein, Leen-Feldner, et al., 2006), all proposed taxo-
metric analyses were conducted on each of the 10 simulated
dimensional and each of the 10 simulated taxonic data before
taxometric analyses of the research data were interpreted. By
examining the degree to which it is possible to distinguish the
taxometric plots of the simulated dimensional and taxonic data
under the parameter-matched conditions of the research data, the
simulations demonstrate the capacity of the research data to afford
meaningful taxometric analyses. Thus, only if each taxometric
procedure could distinguish between the simulated dimensional
and taxonic data could the data be interpreted meaningfully. If the
simulated data pass the suitability tests, then the simulated data
permit a second primary function. Specifically, the simulations
allow investigators to compare the shape of the research data plots
with the plots of the simulated taxonic and dimensional data. If
simulated dimensional and taxonic data are discernible, then a
nontaxonic pattern of findings in the research data may be reliably
interpreted as a marker of latent continuity or nontaxonicity,
whereas a taxonic pattern of findings in the research data can be
more reliably interpreted as evidentiary of taxonicity that is less
likely artifactual or pseudotaxonic. Thus, in addition to contrasting
the taxometric plots of the research data to predefined criteria
(Waller & Meehl, 1998), these parameter-matched simulations
provide an additional comparative idiographic benchmark for in-
terpreting the data with respect to their parameters.
MAXCOV and MAXEIG plot rating reliability. Three investi-
gators trained in taxometrics independently rated each of the
MAXCOV and MAXEIG plots as nontaxonic, ambiguous, or
taxonic. Ratings were assigned on the basis of rater consensus.
Although this is likely sufficient to guard against biased plot
ratings, we also asked a naive rater, who was blind to the
investigation and not trained in taxometrics but familiar with
statistics, to rate each research and simulated MAXCOV and
MAXEIG plot using the same guidelines as those used by the
investigators. To estimate plot rating reliability, we calculated
the degree of agreement regarding plot ratings between the
investigators and the naive rater.
78
BERNSTEIN ET AL.
Analytic Approach: Confirmatory and Exploratory Factor
Analyses
If taxometric analyses indicated taxonic latent structure and
assigned individual participants to complement and taxon classes,
then a CFA of the ASI items was planned; AMOS SEM (Arbuckle,
1999) was to be used to test whether the well-replicated hierarchi-
cal three-factor model of AS (Zinbarg et al., 1997, 1999) fit the
data for complement and taxon classes separately. We planned to
evaluate fit using chi-square tests (optimally p ⬍ .05; Bollen,
1989), root-mean-square residuals (RMR; optimally ⬍ .05; Hu &
Bentler, 1995), the goodness-of-fit index (GFI; optimally ⬎ .90;
Hu & Bentler, 1995), the parsimony goodness-of-fit index (PGFI;
optimally ⬎ .50; James, Mulaik, & Brett, 1982), and the root-
mean-square error of approximation (RMSEA; optimally 90%
confidence interval ⫽ .02–.07; Steiger, 1990).
If the three-factor model fit the ASI data among the complement
class and among the taxon class, then no more analyses were to be
conducted. If the three-factor model did not fit the ASI data among
the complement class and/or the taxon class, then, as needed,
exploratory factor analyses (EFAs) were planned. EFAs via prin-
cipal axis factoring (PAF) between ASI items would be conducted.
The factor solutions would be subjected to oblique rotations.
Results
Taxometric Analyses
Suitability testing. Suitability of the data for MAXCOV anal-
yses was first tested (see Figure 1, simulated plots). MAXCOV
analyses were conducted with the parameter-matched simulated
dimensional and taxonic data. The covariance plots of simulated
dimensional data were cusped, whereas the covariance plots of the
simulated taxonic data were distinctly peaked. Thus, the research
data likely were suitable for MAXCOV analysis. The suitability of
the data for MAXEIG analyses was then tested (see Figure 2,
simulated plots). MAXEIG analyses were conducted with the
parameter-matched simulated dimensional and taxonic data. The
inchworm consistency test of the simulated dimensional data failed
to yield unimodally peaked plots as the number of overlapping
windows was increased and instead yielded cusped plots, whereas
the taxonic data yielded unimodally peaked plots resembling the
head of an inchworm (Waller & Meehl, 1998). Thus, the research
data likely were suitable for MAXEIG analysis. Because the
parameter-matched data each passed suitability, MAXCOV and
MAXEIG analyses of the research data are capable of distinguish-
ing between dimensional and taxonic latent structures of AS.
1
AS taxonicity and parameters. Table 1 provides a summary of
the taxometric analyses. Consistent with prediction, MAXCOV
analysis (Ruscio, 2004) of the three item-parcel indicators pro-
duced two unimodally peaked taxonic plots and one ambiguous
plot (see Figure 1, research plots). This pattern of covariance plots
provides evidence in support of AS taxonicity (Schmidt et al.,
2004; Waller & Meehl, 1998). The MAXCOV covariance plots of
the research data were not cusped like the parameter-matched
simulated dimensional data and, in support of latent taxonicity,
were more similar to the peaked parameter-matched simulated
taxonic data (see Figure 1). Also consistent with prediction, base
rate estimates from all indicator combinations were similar; the
standard deviation of base rate estimates was .04 (SD ⬍ .10 is
desirable; Schmidt et al., 2004). The mean base rate estimate of the
taxon was .11. It is important to note that the intraclass nuisance
correlation was greater than zero level among the complement
class but within the tolerable range (r ⫽ .34; Schmidt et al., 2004;
Waller & Meehl, 1998) and at zero level among the taxon class
(r ⫽ .00). Thus, the estimated parameters of the latent distributions
likely were reliable approximations. All three indicators discrim-
inated between latent classes (average indicator validity ⫽ 2.3
standard deviations; 2.3 standard deviations for the Physical Con-
cerns indicator, 2.8 standard deviations for the Psychological Con-
cerns indicator, and 1.7 standard deviations for the Social Con-
cerns indicator). An effect size between taxon and complement
class members of 1.2 standard deviations or greater is desirable for
coherent cut kinetic taxometric procedures. The Social Concerns
indicator did not as clearly or strongly discriminate between latent
taxa as the Physical Concerns and Psychological Concerns indi-
cators, as indicated by the ambiguous versus taxonic covariance
plot and lower validity parameter estimate, respectively.
MAXEIG analysis and the inchworm consistency test of the
research data produced further convergent evidence in support of
taxonic latent structure. Consistent with prediction and evidentiary
of latent taxonicity, cusped plots became peaked for all three
indicators as the number of overlapping windows was increased
(see Figure 2, research plots; Waller & Meehl, 1998). The
MAXEIG plots of the research data were dissimilar to the matched
simulated dimensional data and more similar to the parameter-
matched simulated taxonic data in that they produced peaks con-
sistent with the head of an inchworm (see Figure 2). The base rate
estimates were similar across the three MAXEIG plots, yielding an
average base rate estimate of .11 (SD ⫽ .06). The intraclass
nuisance correlation was greater than zero level among the com-
plement class but within the tolerable range (r ⫽ .35) and at zero
level among the taxon class (r ⫽ .00). Thus, the estimated param-
eters of the latent distributions likely were reliable approximations.
All three indicators discriminated between latent classes, as ob-
served in the MAXCOV analysis (average indicator validity was
2.3 standard deviations; 2.3 standard deviations for the Physical
Concerns indicator, 2.9 standard deviations for the Psychological
Concerns indicator, and 1.7 standard deviations for the Social
Concerns indicator). The Social Concerns indicator, again, did not
as clearly or strongly discriminate between latent taxa compared
with the other indicators. Class assignment for the purpose of the
planned factor analyses was based on the MAXEIG analysis.
MAXCOV and MAXEIG plot rating reliability. The three in-
vestigators and the naive rater agreed about whether the plots
reflected latent taxonicity on five out of the total five MAXCOV
plots (100% raw agreement). Raters agreed about whether the plots
reflected latent taxonicity on four out of the total five MAXEIG
plots (80% raw agreement). The naive rater indicated that the
Social Concerns MAXEIG plot was ambiguous. The investigators
rated this plot as taxonic because it resembled the head of an
(text continues on page 82)
1
Additional external consistency tests (e.g., MAMBAC) were tested on
the parameter-matched simulated data but did not pass suitability testing
and therefore were not used to test the latent structure of AS in the research
data.
79
LATENT STRUCTURE OF ANXIETY SENSITIVITY
Figure 1. MAXCOV-HITMAX plots of anxiety sensitivity manifest indicators: matched research, simulated taxonic, and simulated dimensional data. The
x-axis shows 125 intervals across each of the individual manifest indicators (i.e., values of the input indicator); the y-axis shows covariance values between
the three bivariate combinations of manifest indicators not serving as an input indicator (i.e., covariance values between the pairs of output indicators).
Simulation figure x-axis ⫽ 125 intervals, y-axis ⫽ Covariance. Indicator 1 ⫽ Physical Concerns; Indicator 2 ⫽ Psychological Concerns; Indicator 3 ⫽
Social Concerns.
80
BERNSTEIN ET AL.
Figure 2. MAXEIG-HITMAX plots of anxiety sensitivity manifest indicators: matched research, simulated taxonic, and simulated dimensional data.
Although MAXEIG plots were cusped at 500 windows, consistent with latent taxonic structure, peaks resembling the head of an inchworm were observed
as the number of overlapping windows was increased. The x-axis shows 1,250 intervals across each of the individual manifest indicators (i.e., values of
the input indicator); the y-axis shows eigenvalues between the three bivariate combinations of manifest indicators not serving as an input indicator (i.e.,
eigenvalues between the pairs of output indicators). Simulation figure x-axis ⫽ 1250 Windows, y-axis ⫽ Eigenvalue. Vars ⫽ variables; Indicator 1 ⫽
Physical Concerns; Indicator 2 ⫽ Psychological Concerns; Indicator 3 ⫽ Social Concerns.
81
LATENT STRUCTURE OF ANXIETY SENSITIVITY
inchworm as the number of windows was increased across
MAXEIG analyses of the inchworm consistency test. After pro-
viding ratings, the naive rater was asked for the rationale for each
rating. The rater indicated that although the Social Concerns
MAXEIG plot looked like the head of an inchworm, consistent
with a taxonic MAXEIG plot (Waller & Meehl, 1998), this plot
was not as peaked as the other two plots that were rated as taxonic
and so he decided to rate the plot as ambiguous.
Descriptive characteristics of participants by latent class. In
terms of gender, consistent with previous empirical study of AS
taxonicity across gender (Bernstein, Zvolensky, Stewart, et al.,
2006), there were more women in the taxon class (71.4%) relative
to the complement class (61.5%),
2
(1, N ⫽ 2,504) ⫽ 9.5, p ⬍ .01.
In terms of ethnicity, there were no significant differences in the
rates of any one or more ethnic groups in the taxon relative to the
complement class,
2
(5, N ⫽ 1,956) ⫽ 7.5, p ⫽ ns. Regarding
education, there were no significant differences in the levels of
education among participants in the taxon class relative to the
complement class,
2
(8, N ⫽ 1,960) ⫽ 8.0, p ⫽ ns. Finally, the
mean age of taxon and complement class members did not differ,
F(1, 2499) ⫽ 0.57, p ⫽ ns.
Prior to conducting the planned factor analyses, the variability
of AS scores among each class was evaluated. The range of scores
of the Physical Concerns indicator among the complement class
was 0–25 (M ⫽ 7.5, SD ⫽ 5.1) and among the taxon class was
4 –35 (M ⫽ 19.4, SD ⫽ 4.8). The range of scores for the Psycho-
logical Concerns indicator among the complement class was 0 –12
(M ⫽ 2.1, SD ⫽ 2.2) and among the taxon class was 1–16 (M ⫽
8.6, SD ⫽ 3.0). The range of scores for the Social Concerns
indicator among the complement class was 0 –15 (M ⫽ 6.9, SD ⫽
2.6) and among the taxon class was 6 –16 (M ⫽ 11.2, SD ⫽ 2.2).
The range of scores for the ASI full-scale total score among the
complement class was 0 –36 (M ⫽ 16.5, SD ⫽ 7.7) and among the
taxon class was 28 – 62 (M ⫽ 39.2, SD ⫽ 6.2). Consequently,
sufficient variability in AS scores was observed among both taxo-
metrically indicated taxonic classes to permit the planned factor
analyses.
Factor Analyses of Taxometrically Identified Taxonic
Forms of AS
CFA: Complement class. Figure 3 presents the structural
model with standardized path coefficients. Indices of fit indicate
acceptable overall model fit,
2
(101, N ⫽ 2,258) ⫽ 1,707.0, p ⬍
.01, RMR ⫽ .06, GFI ⫽ .91, PGFI ⫽ .68, RMSEA ⫽ .08 (90%
confidence interval ⫽ .080 to .087).
2
Indicative of poorer fit, the
chi-square statistic was significant. Highly sensitive to sample
size, chi square must be interpreted cautiously in light of conver-
gence across multiple other fit indices (Bollen, 1989). It is impor-
tant to note that we did not refine the model, based on preexisting
factor analytic findings that identified poor fitting items (Zinbarg
et al., 1997, 1999), observed poor fitting items (e.g., Items 1 and
5), or modification indices.
All indicator variables, except Item 1 and Item 5, were strongly
associated (i.e., ⬎ .40) with their respective latent first-order
factors, suggesting good estimation of all three variables. Results
also provide support for the hierarchical structure of AS among the
complement class. All first-order factors demonstrated large sig-
nificant loadings on the second-order Global AS factor (.84, .70,
and .85 for the Physical, Psychological, and Social Concerns
factors, respectively).
CFA: Taxon class. Figure 3 presents the structural model with
standardized path coefficients. Indices of fit indicate inadequate
overall model fit,
2
(101, N ⫽ 257) ⫽ 310.0, p ⬍ .01, RMR ⫽ .11,
GFI ⫽ .86, PGFI ⫽ .64, RMSEA ⫽ .09 (90% confidence inter-
val ⫽ 0.08 to 1.01). Eight of the 16 indicator variables demon-
strated relatively weak associations (⬍ .40) with their respective
latent first-order factors, suggesting poor overall estimation. Spe-
cifically, five items (i.e., Items 3, 4, 6, 8, and 14) demonstrated
relatively weak associations with Factor I (Physical Concerns), one
item (i.e., Item 12) demonstrated poor association with Factor II
(Psychological Concerns), and two items (i.e., Items 7 and 13)
demonstrated poor associations with Factor III (Social Concerns).
Factors I and II demonstrated smaller loadings on the second-order
Global AS factor (i.e., .53 and .43, respectively) than Factors I and
II demonstrated among the complement class (i.e., .84 and .70).
Factor III (Social Concerns) demonstrated a large negative loading
(–.73) on the second-order Global AS factor, unlike among the
complement class (.85) and unlike theoretically predicted by AS
theory. These results converge on the conclusion that the tested
three-factor solution does not fit the ASI data for the taxon class.
Given that the taxon class data did not fit the hypothesized model,
in contrast to the complement class that did show acceptable fit,
formal tests of structural invariance across groups were unneces-
sary (Byrne, 2001).
EFA: Taxon class. An exploratory PAF (Floyd & Widaman,
1995) was conducted to identify the possible latent factor structure
of the AS taxon. The number of factors to retain was determined
by applying the Kaiser rule (i.e., eigenvalue ⬎ 1.0; Kaiser, 1961),
two quantitative tests of the scree-plot evaluation (Cattell, 1966), a
well-established visual inspection of the scree-plot (Cattell, 1966),
2
The fit of the three-factor model was tested among the mixed-class full
sample. The three-factor solution demonstrated similar, but not better, fit
among the full sample than among the complement class. Because nearly
90% of the full sample belongs to the complement class, this was expected.
Table 1
Summary of Taxometric Analyses
Variable MAXCOV-HITMAX MAXEIG-HITMAX
Taxonic plots 2 3
Nontaxonic plots 0 0
Ambiguous plots 1
Mean P .11 .11
SD .04 .06
Interindicator r .52
Within-complement r .34 .35
Within-taxon r .00 .00
Mean indicator validity 2.3 2.3
Note. Mean P ⫽ grand mean taxon base rate; SD ⫽ standard deviation of
taxon base rate estimates; interindicator r ⫽ mean correlation between
three indicators in the full (mixed-class) sample; within-complement r ⫽
intracomplement class between-variable nuisance correlation; within-taxon
r ⫽ intrataxon class between-variable nuisance correlation; mean indicator
validity ⫽ estimated effect size of indicators in discriminating between
latent classes.
82
BERNSTEIN ET AL.
and factor interpretability. Two factors were extracted by the
Kaiser rule. Relying on the Kaiser rule alone, however, may lead
to factor overextraction (Gorsuch, 1983). Indeed, tests of scree-
plot evaluation converged on a unidimensional solution (i.e., one
factor) as best representing the data. The difference between the
eigenvalues of Factor I minus that of Factor II (.88) was much
greater than between the eigenvalue of Factor II minus Factor III
(.15) and so on for subsequent eigenvalue difference values. The
ratios between factor eigenvalues mirrored the eigenvalue differ-
ence scores. Consistent with the quantitative tests, the visual
inspection of the scree-plot also strongly supported a one-factor or
unidimensional solution. This unidimensional solution yielded a
clearly interpretable factor structure within existing AS theory
(Reiss & McNally, 1985). Thus, following scree-plot evaluation as
described, the PAF analysis was rerun, constraining it to a one-
factor solution.
The PAF pattern matrices (loadings) and communalities for the
one-factor solution are shown in Table 2. Seven hyperplane items
(i.e., items failing to have a salient loading ⬎ .30) were identified
and removed from the final factor solution. Specifically, all four
Social Concerns items (i.e., Items 1, 5, 7, and 13) were thereby
removed empirically from the model. Notably, Items 1, 5, and 7
demonstrated negative loadings (–.23, –.26, and –.16, respec-
tively). Item 2 (“When I cannot keep my mind on a task, I worry
that I might be going crazy”), Item 8 (“It scares me when I am
nauseous”), and Item 12 (“It scares me when I am unable to keep
my mind on a task”) were removed because of weak loadings (.23,
.16, and .14, respectively). Therefore, the unidimensional factor is
Figure 3. Confirmatory factor analysis model with standardized coefficients: complement class and taxon
class. Complement class parameters are presented in Roman text, whereas taxon class parameters appear in bold
italics. err ⫽ error; ASI ⫽ Anxiety Sensitivity Index; Res ⫽ residual.
83
LATENT STRUCTURE OF ANXIETY SENSITIVITY
composed of the following items: Item 3 (“It scares me when I feel
‘shaky’ [trembling]”), Item 4 (“It scares me when I feel faint”),
Item 6 (“It scares me when my heart beats rapidly”), Item 9
(“When I notice that my heart is beating rapidly, I worry that I
might have a heart attack”), Item 10 (“It scares me when I become
short of breath”), Item 11 (“When my stomach is upset, I worry
that I might be seriously ill”), Item 14 (“Unusual body sensations
scare me”), Item 15 (“When I am nervous, I worry that I might be
mentally ill”), and Item 16 (“It scares me when I am nervous”).
Next, following rigorous guidelines for the conduct of EFA
factor extraction (O’Connor, 2000), a parallel analysis was con-
ducted (Horn, 1965). Parallel analysis was thus used as a consis-
tency test for the purpose of factor extraction to either provide
convergent support for a one-factor solution or divergent support
for an alternative multidimensional solution. Specifically, Monte
Carlo simulations were used to generate 100 random data sets of
raw data permutations matched to the research data (e.g., number
of items, cases), except that the simulated permutations of the raw
data are random and not correlated (O’Connor, 2000). The eigen-
values of the research and the 95th percentile of the eigenvalues
across the 100 simulated data sets were compared. Factors that
demonstrated eigenvalues ⬎ 1.0 and that were greater than those
observed among the random simulated data were extracted and
retained. Research data eigenvalues that are smaller than those
observed among the random simulated data are, by definition,
nonsubstantive spurious factors and should not be extracted (e.g.,
Gorsuch, 1983; O’Connor, 2000). Parallel analysis supported a
two-factor solution. Thus, the PAF analysis was rerun, constrained
to a two-factor solution.
The PAF pattern matrices (loadings) and communalities for the
two-factor solution are shown in Table 3. First, after an oblique
rotation, the two-factor solution demonstrated sound simple struc-
ture. No items demonstrated complex loadings (i.e., loadings of ⬎
.30 on more than one factor). Furthermore, like the one-factor
solution, five hyperplane items (i.e., Items 1, 5, 7, 12, and 13) were
identified and removed from the final factor solution. Thus, all
Social Concerns items and one Psychological Concerns item were
removed from the factor solution. Eight items loaded onto Factor
I (i.e., Items 3, 4, 6, 8, 9, 10, 11, and 14)—with the exception of
Item 8, all of these items similarly loaded onto the alternative
one-factor solution. These eight items uniformly reflect physical or
disease concerns. Three items univocally loaded onto Factor II
(i.e., Items 2, 15, and 16). With exception of Item 2, these two
items similarly loaded onto the alternative one-factor solution.
These three items uniformly reflect mental illness or cognitive
dyscontrol concerns. Furthermore, two items that primarily loaded
onto Factor I, Items 9 and 14, demonstrated near salient loadings
on Factor II as well. Consequently, the two-factor solution was a
retainable solution based on contemporary standards for explor-
atory factor extraction in that it was psychometrically and theo-
retically interpretable.
Discussion
Results of the MAXCOV test, internal consistency tests, the
MAXEIG test, the inchworm consistency test, and comparisons
with Monte Carlo simulated taxonic and continuous data con-
verged on the conclusion that AS, as indexed by the ASI, is taxonic
in this large sample of young adults. MAXCOV and MAXEIG
plots were characteristically categorical or characteristic of latent
discontinuity. The base rate estimate was .11. These results are
consistent with the findings of previous taxometric work on AS
using the 16-item ASI in young adults from North America
(Schmidt et al., 2005; Zvolensky et al., in press), the CASI among
youth (Bernstein, Zvolensky, Weems, et al., 2005) and across
gender (Bernstein, Zvolensky, Stewart, et al., 2006), and the 16-
item ASI and 36-item ASI–R among adults from different regions
of the world (Bernstein, Leen-Feldner, et al., 2006; Bernstein,
Zvolensky, Kotov, et al., 2006). Overall, these taxometric results
replicate and extend empirical evidence for the existence of dis-
crete, putatively normative and anxiety-psychopathology-
vulnerability-conferring forms of AS. In addition, the Social Con-
Table 2
Loadings and Extraction Communalities for the Unidimensional Factor Solution of the Anxiety
Sensitivity Taxon
Anxiety Sensitivity Index item Loading h
2
Item 9: When I notice that my heart is beating rapidly, I worry that I might have a
heart attack.
.53 .28
Item 10: It scares me when I become short of breath. .50 .25
Item 11: When my stomach is upset, I worry that I might be seriously ill. .46 .21
Item 14: Unusual body sensations scare me. .45 .21
Item 15: When I am nervous, I worry that I might be mentally ill. .37 .14
Item 6: It scares me when my heart beats rapidly. .34 .12
Item 16: It scares me when I am nervous. .34 .11
Item 4: It scares me when I feel faint. .33 .11
Item 3: It scares me when I feel “shaky” (trembling). .33 .11
Item 13: Other people notice when I feel shaky. .23 .05
Item 2: When I cannot keep my mind on a task, I worry that I might be going crazy. .23 .05
Item 8: It scares me when I am nauseous. .16 .03
Item 12: It scares me when I am unable to keep my mind on a task. .14 .02
Item 7: It embarrasses me when my stomach growls. ⫺.16 .03
Item 1: It is important to me not to appear nervous. ⫺.23 .05
Item 5: It is important to me to stay in control of my emotions. ⫺.26 .07
Note. Salient loadings (⬎ .30) appear in bold.
84
BERNSTEIN ET AL.
cerns manifest indicator less clearly discriminated between
taxometrically identified latent classes of AS than did the other
manifest indicators. Finally, within-complement-class between-
indicator nuisance correlations were observed, whereas no such
nuisance correlations were observed within the taxon class.
Factor analyses yielded what may be informative results in the
field’s ongoing efforts to better understand the latent nature of AS.
CFAs indicated that the replicated three-factor hierarchical model
of AS (Zinbarg et al., 1997, 1999) demonstrated acceptable fit with
ASI data for the complement class but poor fit for the taxon class.
These CFAs thus demonstrate that the well-established three-
factor structure of AS (Zinbarg et al., 1997, 1999) may in actuality
reflect the factor structure of the normative form of AS but not the
taxon form. These findings are consistent with the observed
between-variable nuisance correlations among the complement,
but not the taxon, class and the conjecture that previous factor
analytic studies of AS violated the key assumption of population
heterogeneity by not distinguishing between the latent taxonic AS
classes. Indeed, EFA results indicate that the AS taxon demon-
strates a different factor solution. The AS taxon factor structure is
unique from that observed in previous factor analytic studies that
did not discriminate between latent AS classes and from the factor
structure of the complement class.
EFAs indicate alternative or competing factor solutions for
future factor analytic investigations to evaluate in independent
samples. Preliminary EFA findings suggested that the AS taxon
may demonstrate a unique factor solution, specifically, a unidi-
mensional factor structure or that the taxon may demonstrate a
two-factor solution. The observed AS taxon unidimensional factor
solution may be conceptualized as consistent with early theoretical
conceptualizations of AS as a core, singular fear-of-fear composed
of unidimensionally interrelated fears of experiencing anxiety
(Reiss & McNally, 1985). The two-factor solution is mostly con-
sistent with the factor solution among the complement class—
composed of a Physical Concerns factor and a Psychological
Concerns factor— but unlike the complement class has no Social
Concerns facet to its dimensional structure. Though it may be
premature to gauge the clinical significance of the different factor
structures of the latent classes observed here, it is possible that the
types of experiences or events that shape physical and psycholog-
ical concerns may be particularly relevant to understanding certain
anxiety problems. Future study into the etiology of AS itself while
recognizing the latent structure of the construct would be an
important and timely next research step.
The preliminary AS taxon unidimensional and two-factor solu-
tions did not retain items related to social concerns. These findings
are consistent with previous taxometric findings that indicate less
compelling evidence for the Social Concerns dimension as a valid
manifest indicator in discriminating between latent AS classes
(Bernstein, Zvolensky, Stewart, et al., 2006). Overall, the present
factor analytic findings provide further support for the meaning-
fulness of the taxometrically based inference of AS taxonicity.
These findings also demonstrate that the latent structural nature of
AS may be more accurately conceptualized as taxonic latent class
structure composed of two types or forms of AS, each of which,
however, is characterized by its own unique, systematically mean-
ingful latent continuity and dimensional structure.
A number of limitations need to be recognized. First, the ob-
served taxon may reflect latent taxonicity of AS or could hypo-
thetically reflect some other latent taxon that is only related to the
ASI manifest indicators intended to measure the construct. A more
complete, construct valid understanding of the latent AS taxon can
be achieved by future studies that more rigorously screen partici-
pants for possible psychopathology that may be taxonic and related
to AS. In this same general context, it is important to note the need
for further exploration of whether the observed findings are spe-
cific to AS. Here, it would be advisable to examine whether the
observed AS results extend more broadly to anxiety or negative
emotionality. There is emerging work on this exact topic (Kotov,
Schmidt, Lerew, Joiner, & Ialongo, 2005). Second, the EFAs do
not include the rigorous tests and fit indices afforded by CFA
analyses. Such CFA approaches should thus be applied to future
Table 3
Loadings and Extraction Communalities for the Two-Factor Solution of the Anxiety Sensitivity Taxon
Anxiety Sensitivity Index item
Loading
h
2
Factor I Factor II
Item 10: It scares me when I become short of breath. .64 .08 .39
Item 9: When I notice that my heart is beating rapidly, I worry that I might have a heart attack. .46 .17 .28
Item 11: When my stomach is upset, I worry that I might be seriously ill. .45 .09 .23
Item 4: It scares me when I feel faint. .45 .11 .19
Item 6: It scares me when my heart beats rapidly. .39 .02 .15
Item 8: It scares me when I am nauseous. .37 .27 .16
Item 14: Unusual body sensations scare me. .35 .20 .20
Item 3: It scares me when I feel “shaky” (trembling). .32 .04 .11
Item 15: When I am nervous, I worry that I might be mentally ill. .06 .84 .68
Item 2: When I cannot keep my mind on a task, I worry that I might be going crazy. .07 .49 .22
Item 16: It scares me when I am nervous. .10 .41 .20
Item 13: Other people notice when I feel shaky. .18 .11 .05
Item 12: It scares me when I am unable to keep my mind on a task. .00 .21 .05
Item 7: It embarrasses me when my stomach growls. .03 .22 .05
Item 1: It is important to me not to appear nervous. .13 .16 .05
Item 5: It is important to me to stay in control of my emotions. .18 .14 .06
Note. Salient loadings (⬎ .30) appear in bold.
85
LATENT STRUCTURE OF ANXIETY SENSITIVITY
data sets to test the fit of the EFA-based dimensional factor models
of the AS taxon found in the present investigation. We did not
complete these analyses in the present study to avoid overmodel-
ing of the data. Third, although large in overall size compared with
prior taxometric and factor analytic studies of AS, the sample was
not a normative probability sample of adults. Future study using
such epidemiological sampling techniques is important for repli-
cation, parameter estimation, and the derivation of a disseminable
“ASI Taxon Scale” that is psychometrically sound and generaliz-
able. These findings suggest that this scale or subscale be devel-
oped to (a) assign individual cases to the correct latent class with
the highest rate of sensitivity and specificity possible and (b)
assign individuals accurately to the widely dispersed value(s) on
the unique, latent continua within each class.
Fourth, although previous taxometric studies included analyses
related to differences between the AS taxon and complement
classes or forms of AS, no such analyses were possible in the
present investigation because of multisite sampling and cross-
sample differences in measurement. Future study should thus
examine the factor structure of each latent form of AS in relation
to theoretically relevant external criteria to test for class-level and
potentially for dimension-level differential specificity. Fifth, it is
important to note that one alternative account of the observed
differences in factor structure among the complement class com-
pared with the taxon class may be related to differential levels of
ASI score variability between classes. For example, lower ASI
score variability among the taxon class could in part account for
the observed factor structure differences between classes. How-
ever, this alternative hypothesis seems unlikely in the present
investigation, as there were high levels of variability for all AS
indicators and total scores among both the complement and taxon
classes. Future investigations could further test this alternative
account by oversampling taxon members and evaluating the factor
structures between the taxonic AS classes. Sixth, the present
sample was an aggregate of separate samples collected across
different regions and locations in North America. As none of the
sites were treatment clinics for psychological or medical problems,
it is probably unlikely that “institutional taxa” (Cattell, 1966)
played a role in the observed AS taxonicity. To further address this
matter, however, future research could usefully seek to replicate
and extend the present results by studying persons who have been
determined not to have psychopathology via structured clinical
interviews. Finally, the present investigation also was limited
insofar as manifest indices were derived from a single self-report
measure. Future investigations should replicate the present find-
ings using a multimethod approach.
Despite the limitations, there are a number of direct clinical
implications that emerge from the study. To the extent that there
are different and clinically significant forms of AS at the latent
level, it is important that researchers and clinicians incorporate
such findings into research and practice activities. For example, by
identifying those persons in an AS taxon class, rather than simply
using elevated levels of total or subscale ASI scores, it may be
possible to target them in a more efficient manner with prevention
programs targeted at cognitive vulnerability for anxiety psychopa-
thology. By not incorporating such information into the assessment
process, vital information is lost, rendering subsequent clinically
oriented activities potentially less effective. This type of work also
may hold promise for refining existing theory on cognitive vul-
nerability for anxiety problems.
References
Arbuckle, J. L. (1999). Amos 4.0 [Computer software]. Chicago: Small-
waters.
Barlow, D. H. (2002). Anxiety and its disorders (2nd ed.). New York:
Guilford.
Bauer, D. J., & Curran, P. J. (2004). The integration of continuous and
discrete latent variable models: Potential problems and promising op-
portunities. Psychological Methods, 9, 3–29.
Beauchaine, T. P. (2003). Taxometrics and developmental psychopathol-
ogy. Development and Psychopathology, 15, 501–527.
Bernstein, A., Leen-Feldner, E. W., Kotov, R., Schmidt, N. B., & Zvolen-
sky, M. J. (2006). A test of a panic-relevant diathesis–stress model using
a taxonic index of anxiety sensitivity. In A. J. Sanfelippo (Ed.), Panic
disorders: New research (pp. 15– 40). Hauppauge, NY: Nova Science
Publishing.
Bernstein, A., Zvolensky, M. J., Feldner, M. T., Lewis, S. F., Fauber, A. L.,
Leen-Feldner, E. W., & Vujanovic, A. A. (2005). Anxiety sensitivity
taxon and trauma: Discriminant associations for posttraumatic stress and
panic symptomatology among young adults. Depression and Anxiety,
22, 138 –149.
Bernstein, A., Zvolensky, M. J., Feldner, M. T., Lewis, S. F., & Leen-
Feldner, E. W. (2005). Anxiety sensitivity taxonicity: A concurrent test
of cognitive vulnerability for posttraumatic stress symptomatology
among young adults. Cognitive Behavior Therapy, 34, 229 –241.
Bernstein, A., Zvolensky, M. J., Kotov, R., Arrindell, W. A., Taylor, S.,
Sandin, B., et al. (2006). Taxonicity of anxiety sensitivity: A multi-
national analysis. Journal of Anxiety Disorders, 20, 1–22.
Bernstein, A., Zvolensky, M. J., Stewart, S. H., Comeau, M. N., &
Leen-Feldner, E. W. (2006). Anxiety sensitivity taxonicity across gender
among youth. Behaviour Research and Therapy, 44, 679 – 698.
Bernstein, A., Zvolensky, M. J., Weems, C., Stickle, T., & Leen-Feldner,
E. (2005). Taxonicity of anxiety sensitivity: An empirical test among
youth. Behaviour Research and Therapy, 43, 1131–1155.
Bollen, K. A. (1989). Structural equations with latent variables. New
York: Wiley.
Bouton, M. E. (2004). Context and behavioral processes in extinction.
Learning and Memory, 11, 485– 494.
Byrne, B. M. (2001). Structural equation modeling with AMOS: Basic
concepts, applications and programming. Mahwah, NJ: Erlbaum.
Cattell, R. (1966). The meaning and strategic use of factor analysis. In
R. B. Cattell (Ed.), Handbook of multivariate experimental psychology
(pp. 174 –243). Chicago: Rand McNally.
Floyd, F. J., & Widaman, K. F. (1995). Factor analysis in the development
and refinement of clinical assessment instruments. Psychological As-
sessment, 7, 286 –299.
Gorsuch, R. L. (1983). Factor analysis (2nd ed.). Hillsdale, NJ: Erlbaum.
Haslam, N., & Kim, H. C. (2002). Categories and continua: A review of
taxometric research. Genetic, Social, and General Psychology Mono-
graphs, 128, 271–320.
Horn, J. L. (1965). A rationale and test for the number of factors in factor
analysis. Psychometrika, 30, 179 –185.
Hu, L.-T., & Bentler, P. M. (1995). Evaluating model fit. In R. H. Hoyle
(Ed.), Structural equation modeling: Concepts, issues, and applications
(pp. 76 –99). Thousand Oaks, CA: Sage.
James, L. R., Mulaik, S. A., & Brett, J. M. (1982). Causal analysis:
Assumptions, models, and data. Beverly Hills, CA: Sage.
Jo¨reskog, K. G. (1971). Simultaneous factor analysis in several popula-
tions. Psychometrika, 36, 409 – 426.
Kaiser, H. F. (1961). A note on Guttman’s lower bound for the number of
common factors. Multivariate Behavioral Research, 1, 249 –276.
86
BERNSTEIN ET AL.
Kotov, R., Schmidt, N. B., Lerew, D. R., Joiner, T. E., & Ialongo, N. S.
(2005). Latent structure of anxiety: Taxometric exploration. Psycholog-
ical Assessment, 17, 369 –374.
McNally, R. J. (2002). Anxiety sensitivity and panic disorder. Biological
Psychiatry, 52, 938 –946.
O’Connor, B. P. (2000). SPSS and SAS programs for determining the
number of components using parallel analysis and Velicer’s MAP test.
Behavior Research Methods, Instrumentation, and Computers, 32, 396 –
402.
Peterson, R. A., & Plehn, K. (1999). Measuring anxiety sensitivity. In S.
Taylor (Ed.), Anxiety sensitivity: Theory, research, and treatment of the
fear of anxiety (pp. 61– 81). Mahwah, NJ: Erlbaum.
Pickles, A., & Angold, A. (2003). Natural categories or fundamental
dimensions: On carving nature at the joints and the rearticulation of
psychopathology. Development and Psychopathology, 15, 529 –551.
Reiss, S., & McNally, R. J. (1985). Expectancy model of fear. In S. Reiss
& R. R. Bootzin (Eds.), Theoretical issues in behavior therapy (pp.
107–121). San Diego, CA: Academic Press.
Reiss, S., Peterson, R. A., Gursky, M., & McNally, R. J. (1986). Anxiety
sensitivity, anxiety frequency, and the prediction of fearfulness. Behav-
iour Research and Therapy, 24, 1– 8.
Ruscio, J. (2004). Taxometric programs in the R language. Retrieved
January 2005, from http://www.etown.edu/psychology/Documents/
TaxProgDocR.PDF
Ruscio, J., & Ruscio, A. M. (2002). A structure-based approach to psy-
chological measurement: Matching measurement models to latent struc-
ture. Assessment, 9, 4 –16.
Ruscio, J., & Ruscio, A. M. (2004). Clarifying boundary issues in psycho-
pathology: The role of taxometrics in a comprehensive program of
structural research. Journal of Abnormal Psychology, 113, 24 –38.
Schmidt, N. B., Kotov, R., & Joiner, T. E., Jr. (2004). Taxometrics: Toward
a new diagnostic scheme for psychopathology. Washington, DC: Amer-
ican Psychological Association.
Schmidt, N. B., Kotov, R., Lerew, D. R., Joiner, T. E., & Ialongo, N. S.
(2005). Evaluating latent discontinuity in cognitive vulnerability to
panic: A taxometric investigation. Cognitive Therapy and Research, 29,
673– 691.
Schmidt, N. B., Lerew, D. R., & Joiner, T. E., Jr. (2000). Prospective
evaluation of the etiology of anxiety sensitivity: Test of a scar model.
Behaviour Research and Therapy, 38, 1083–1095.
Silverman, W. K., Fleisig, W., Rabian, B., & Peterson, R. A. (1991). Child
Anxiety Sensitivity Index. Journal of Clinical Child Psychology, 20,
162–168.
Silverman, W. K., Goedhart, A. W., Barrett, P., & Turner, C. (2003). The
facets of anxiety sensitivity represented in the Childhood Anxiety Sen-
sitivity Index: Confirmatory analyses of factor models from past studies.
Journal of Abnormal Psychology, 112, 364 –374.
Steiger, J. H. (1990). Structural model evaluation and modification: An
interval estimation approach. Multivariate Behavioral Research, 25,
173–180.
Stewart, S. H., Taylor, S., Jang, K. L., Cox, B. J., Watt, M. C., Fedoroff,
I. C., & Borger, S. C. (2001). Causal modeling of relations among
learning history, anxiety sensitivity, and panic attacks. Behaviour Re-
search and Therapy, 39, 443– 456.
Taylor, S., & Cox, B. J. (1998). An expanded anxiety sensitivity index:
Evidence for a hierarchic structure in a clinical sample. Journal of
Anxiety Disorders, 12, 463– 483.
Taylor, S., Rabian, B., & Fedoroff, I. (1999). Anxiety sensitivity: Progress,
prospects, and challenges. In S. Taylor (Ed.), Anxiety sensitivity: Theory,
research, and treatment of the fear of anxiety (pp. 339 –353). Mahwah,
NJ: Erlbaum.
Thurstone, L. L. (1935). On the mixture of distributions. Annals of Math-
ematical Statistics, 31, 55–73.
Waller, N. G., & Meehl, P. E. (1998). Multivariate taxometric procedures.
Thousand Oaks, CA: Sage.
Zinbarg, R. E., Brown, T. A., & Barlow, D. H. (1997). Hierarchical
structure and general factor structure saturation of the Anxiety Sensitiv-
ity Index: Evidence and implications. Psychological Assessment, 9,
277–284.
Zinbarg, R. E., Mohlman, J., & Hong, N. N. (1999). Dimensions of anxiety
sensitivity. In S. Taylor (Ed.), Anxiety sensitivity: Theory, research, and
treatment of the fear of anxiety (pp. 83–114). Mahwah, NJ: Erlbaum.
Zvolensky, M. J., Forsyth, J. P., Bernstein, A., & Leen-Feldner, E. W. (in
press). A concurrent test of the anxiety sensitivity taxon: Its relation to
bodily vigilance and perceptions of control over anxiety-related events
in a sample of young adults. Journal of Cognitive Psychotherapy.
Received March 28, 2005
Revision received October 12, 2006
Accepted October 13, 2006 䡲
87
LATENT STRUCTURE OF ANXIETY SENSITIVITY
A preview of this full-text is provided by American Psychological Association.
Content available from Psychological Assessment
This content is subject to copyright. Terms and conditions apply.