Content uploaded by Jonathan Abramowitz
Author content
All content in this area was uploaded by Jonathan Abramowitz
Content may be subject to copyright.
Robust Dimensions of Anxiety Sensitivity: Development and Initial
Validation of the Anxiety Sensitivity Index—3
Steven Taylor
University of British Columbia
Michael J. Zvolensky
University of Vermont
Brian J. Cox
University of Manitoba
Brett Deacon
University of Wyoming
Richard G. Heimberg and Deborah Roth Ledley
Temple University
Jonathan S. Abramowitz
University of North Carolina at Chapel Hill
Robert M. Holaway
Temple University
Bonifacio Sandin
Universidad Nacional de Educacio´n a Distancia
Sherry H. Stewart
Dalhousie University
Meredith Coles
State University of New York at Binghamton
Winnie Eng
City University of New York
Erin S. Daly
Boston University and Veterans Affairs Boston Healthcare System
Willem A. Arrindell
University of Groningen
Martine Bouvard
Universite´ de Savoie
Samuel Jurado Cardenas
University of Mexico
Accumulating evidence suggests that anxiety sensitivity (fear of arousal-related sensations) plays an important
role in many clinical conditions, particularly anxiety disorders. Research has increasingly focused on how the
basic dimensions of anxiety sensitivity are related to various forms of psychopathology. Such work has been
hampered because the original measure—the Anxiety Sensitivity Index (ASI)—was not designed to be
multidimensional. Subsequently developed multidimensional measures have unstable factor structures or
measure only a subset of the most widely replicated factors. Therefore, the authors developed, via factor
analysis of responses from U.S. and Canadian nonclinical participants (n 2,361), an 18-item measure, the
ASI–3, which assesses the 3 factors best replicated in previous research: Physical, Cognitive, and Social
Concerns. Factorial validity of the ASI–3 was supported by confirmatory factor analyses of 6 replication
samples, including nonclinical samples from the United States and Canada, France, Mexico, the Netherlands,
and Spain (n 4,494) and a clinical sample from the United States and Canada (n 390). The ASI–3
displayed generally good performance on other indices of reliability and validity, along with evidence of
improved psychometric properties over the original ASI.
Keywords: anxiety sensitivity, Anxiety Sensitivity Index, anxiety disorders
Supplemental material: http://dx.doi.org/10.1037/1040-3590.19.2.176.supp
Steven Taylor, Department of Psychiatry, University of British Colum-
bia, Vancouver, Canada; Michael J. Zvolensky, Department of Psychology,
University of Vermont; Brian J. Cox, Department of Psychiatry, University
of Manitoba, Canada; Brett Deacon, Department of Psychology, University
of Wyoming; Richard G. Heimberg, Deborah Roth Ledley, and Robert M.
Holaway, Department of Psychology, Temple University; Jonathan S.
Abramowitz, Department of Psychology, University of North Carolina at
Chapel Hill; Bonifacio Sandin, Department of Psychology, Universidad
Nacional de Educacio´n a Distancia, Madrid, Spain; Sherry H. Stewart,
Departments of Psychiatry and Psychology, Dalhousie University, Halifax,
Nova Scotia, Canada; Meredith Coles, Department of Psychology, State
University of New York at Binghamton; Winnie Eng, Department of
Psychology, City University of New York; Erin S. Daly, Department of
Psychiatry, Boston University, and Veterans Affairs Boston Healthcare
System; Willem A. Arrindell, Department of Clinical Psychology, Univer-
sity of Groningen, the Netherlands; Martine Bouvard, Department of
Psychology, Universite´ de Savoie, Chambery, France; Samuel Jurado
Cardenas, Department of Psychology, University of Mexico, Mexico City.
Correspondence concerning this article, including requests for obtaining
copies of the ASI–3, should be addressed to Steven Taylor, Department of
Psychiatry, University of British Columbia, Vancouver, British Columbia
V6T 2A1, Canada. E-mail: taylor@unixg.ubc.ca
Psychological Assessment Copyright 2007 by the American Psychological Association
2007, Vol. 19, No. 2, 176 –188 1040-3590/07/$12.00 DOI: 10.1037/1040-3590.19.2.176
176
Anxiety sensitivity (AS) is the fear of arousal-related sensations,
arising from beliefs that the sensations have adverse consequences
such as death, insanity, or social rejection (Reiss & McNally,
1985). AS is conceptualized as a contributor to individual differ-
ences in general fearfulness and as a diathesis for various types of
anxiety disorders, including panic disorder, social anxiety disorder,
specific phobia, and posttraumatic stress disorder (Reiss & Mc-
Nally, 1985; Taylor, 1999). This is because AS is an anxiety
amplifier; when highly anxiety-sensitive people become anxious,
they become alarmed about their arousal-related sensations, which
further intensifies their anxiety. Consistent with this formulation is
evidence showing that AS is elevated in people with various types
of anxiety disorders compared with control participants, and a
person’s current level of AS predicts the risk of future anxiety
symptoms (see Taylor, 1999, for a review).
AS was originally conceived as a unidimensional construct, as
measured by the Anxiety Sensitivity Index (ASI; Peterson & Reiss,
1992). Many factor analyses of the ASI have been conducted, with
solutions ranging from one to four factors (Taylor, 1999). The
inconsistent factor solutions may have arisen for various reasons,
including differences across studies in factor selection criteria
(e.g., use of the eigenvalue 1 rule in some studies, which can
lead to factor overextraction) and the use of small samples in some
studies, which can yield unreliable findings (Taylor, 1999). The
fact that the ASI was not constructed to be multidimensional may
also have contributed to instability (lack of replicability) in the
factor structure, because some of the domains of AS (as described
below) were measured by only a few items, thereby reducing the
odds that a given factor would be reliably obtained across different
samples.
Despite the inconsistencies, the most commonly obtained factor
solution consists of three correlated factors labeled Physical Con-
cerns, Cognitive Concerns, and Social Concerns (Taylor, 1999).
To illustrate these factors, high scores on Physical Concerns are
associated with the belief that palpitations lead to cardiac arrest.
High scores on Cognitive Concerns are associated with the belief
that concentration difficulties lead to insanity. High scores on
Social Concerns are associated with the belief that publicly ob-
servable anxiety reactions (e.g., trembling) will elicit social rejec-
tion or ridicule. The three factors load on a single higher order
factor (Global AS). The factor structure does not appear to vary as
a function of gender or age, at least for the age ranges that have
been studied (i.e., adults and children as young as 7 years; e.g.,
Dehon, Weems, Stickle, Costa, & Berman, 2005; Stewart, Taylor,
& Baker, 1997).
There are several problems with the ASI for the multidimen-
sional assessment of AS. The ASI Physical Concerns subscale
(constructed on the basis of factor analysis) has eight items,
whereas the Cognitive and Social Concerns subscales each contain
only four items. Therefore, the latter two subscales might not have
adequate reliability, especially if they do not have strong content
validity. Related to the issue of content validity, particularly for the
purpose of measuring specific AS dimensions, is the fact that some
ASI items do not target specific dimensions. For example, the item
“It scares me when I am nauseous” could measure physical con-
cerns, or it could measure social concerns (i.e., concerns about the
social consequences of vomiting in front of others).
In attempts to overcome the problems with the ASI, Taylor and
Cox (1998a, 1998b) developed a 36-item revised scale—the ASI—
Revised (ASI–R)—and a 60-item Anxiety Sensitivity Profile. Un-
fortunately, both scales have unstable factor structures, with dif-
ferent studies obtaining different solutions (e.g., Deacon,
Abramowitz, Woods, & Tolin, 2003; Zvolensky et al., 2003).
Further research is needed to develop a psychometrically sound,
multidimensional measure of AS.
1
Accordingly, in the present
series of investigations we selected items from the ASI–R, which
is the most widely used multidimensional measure of AS, and
constructed and evaluated a revised scale, the ASI–3, which mea-
sures the three most commonly replicated dimensions: physical,
cognitive, and social concerns. This scale, like the original ASI,
was intended for use in clinical and nonclinical samples. Although
it is theoretically important to develop a scale that can reliably and
validly assess AS in clinical samples, it is equally important that
such a scale should be able to assess AS in nonclinical samples in
order to identify people who are theoretically at risk for developing
anxiety disorders or related problems. Accordingly, we evaluated
the ASI–3 with both types of samples.
Study 1 concerned the development of the ASI–3. Study 2
evaluated the factorial validity of the ASI–3 in six replication
samples and compared the results with those of the ASI. Study 3
examined reliability as internal consistency of each of the ASI–3
subscales, which were compared with those of the ASI subscales.
Although only one index of reliability was examined (internal
consistency), it provides a good estimate of reliability in general
because the sampling of item content is usually the major source of
measurement error for traitlike constructs (Nunnally & Bernstein,
1994). Study 4 examined convergent, discriminant, and criterion-
related (known groups) validities. The studies described in this
article were based on data collected, with informed consent, in
previous research (Bernstein et al., 2006; Deacon et al., 2003;
Roth, Coles, & Heimberg, 2002; Taylor & Cox, 1998a; Zvolensky
et al., 2003). None of those studies addressed the aims of the
present article.
Study 1: Construction of the ASI–3
A goal of this study was to construct the ASI–3 by selecting
ASI–R items that each measured only one of the domains of
physical, cognitive, or social concerns. A further goal was to
determine whether the items formed subscales corresponding to
separate but correlated factors. Item selection for the ASI–3 was
done by balancing two opposing goals: to have an overall scale
that is as efficient to administer as the original (16-item) ASI,
while also having subscales that each contain a sufficient number
of items to ensure adequate subscale reliability. Therefore, it was
decided to create three 6- to 7-item subscales.
Method
Participants. A sample of 4,720 young adults were recruited
from universities across the United States and Canada: Dalhousie
1
Blais et al. (2001) developed a modified version of the ASI by deleting
items that were not highly correlated with the total score. This resulted in
the deletion of all social concerns items, and so the scale measured only
two of the three most widely replicated factors (physical and cognitive
concerns). A comprehensive assessment of AS requires that all of the most
widely replicated factors be assessed. Given these concerns, we did not
specifically examine the Blais et al. ASI in the present study.
177
ROBUST DIMENSIONS OF ANXIETY SENSITIVITY
University (n 155), University of British Columbia (n 156),
University of Manitoba (n 167), University of Vermont (n
347), Temple University (n 1,912), and Northern Illinois Uni-
versity (n 1,983). Participants were all undergraduates, except
for those recruited from the University of Vermont site, who were
a mix of 70% college students and 30% nonstudent young adults
recruited from the community. The overall sample of 4,720 par-
ticipants was randomly split in two. The U.S.–Canadian Sub-
sample 1 (n 2,361) was used for the construction of the ASI–3
in the present study. Subsample 2 was one of the samples used in
subsequent studies described in this article.
2
The mean age of
Subsample 1 was 19.6 years (SD 3.4), and 66% were women.
The majority of participants were White (61%), with the remainder
being African American (19%), Asian (9%), Hispanic (3%), or
other (8%).
Measures and procedure. All participants completed a short
questionnaire assessing demographic features. U.S. participants
completed the 36-item ASI–R. Canadian participants completed a
42-item scale, consisting of the original 16-item ASI and the
36-item ASI–R (the ASI and ASI–R have 10 items in common).
Participants completed the measures in classroom or individual
settings for either course credit or an honorarium.
Scale construction and statistical methods. Items were ini-
tially selected, from the pool of 36 ASI–R items, with an emphasis
on content validity. An item was selected if its content unambig-
uously corresponded to only one of the domains of physical,
cognitive, or social concerns. Twelve items were eliminated be-
cause it was unclear which domain they measured. For example,
“It scares me when I feel faint” could measure physical concerns
(e.g., fear of fainting and injuring oneself) or social concerns (e.g.,
fear of making a spectacle of oneself by collapsing in front of
others). Similarly, the item “When my head is pounding, I worry
I could have a stroke” was eliminated because it could measure
cognitive concerns (fear of cognitive impairment resulting from a
stroke) or physical concerns (fear of stroke-related death or phys-
ical impairment). Of the remaining 24 ASI–R items, 3 were
eliminated because their wording was very similar to that of other
items. For each set of such redundant items, we retained the item
with the lowest Flesch–Kincaid reading level. The resulting pool
consisted of seven items for each of the three subscales, in which
each subscale assessed a latent factor (Physical, Cognitive, or
Social Concerns).
3
The degree of fit of the items to the three-factor model was
tested with LISREL 8.72 (Jo¨reskog & So¨rbom, 2005), by means of
polychoric correlations because the data were ordinal, and robust
weighted least-squares. As such, all of the resulting fit indices
were robust estimates. A model-fitting (confirmatory factor ana-
lytic) approach was used instead of a purely exploratory factor
analytic approach because we hypothesized, on the basis of pre-
vious research (e.g., Zinbarg, Barlow, & Brown, 1997), that a
three-factor solution would provide a good fit to the data.
4
How
-
ever, model modification indices were used in this study, and so
confirmatory factor analysis was used in an exploratory mode for
the purpose of item selection.
Factors were allowed to covary with one another because pre-
vious AS research has consistently shown that such factors are
correlated (e.g., Taylor & Cox, 1998a, 1998b; Taylor, Koch,
Woody, & McLean, 1996). Error terms (residuals) were not per-
mitted to be correlated. The model was simultaneously fitted to
two groups: women and men from Subsample 1. This was done to
test whether the factor structure was invariant across gender, as
suggested by previous studies of AS, and to make any model
modifications (item deletions) that equally emphasized model fit in
the samples of women and men.
A stepwise approach was adopted for the multigroup analyses:
(a) An individual three-factor confirmatory analysis was con-
ducted for each subgroup (i.e., separate analyses for women and
men in the present study), (b) tests of equal form across subgroups
(i.e., fitting the same number of factors to each gender), (c) tests of
invariance of factor loadings across subgroups (i.e., equal factor
loadings for the three-factor model across gender), (d) tests of
invariance of factor correlations, with factor loadings constrained
to be invariant, and finally, (e) tests of invariance of error vari-
ances (with loadings, error variances, and factor correlations con-
strained). We tested whether model fit deteriorated with successive
constraints, as assessed by changes in fit indices and, in the case of
nested models, by Satorra–Bentler corrected differences in chi-
square (i.e.,
2
diff
). Only the main analyses for this stepwise
approach are reported in this article. Here, we report the results for
the most stringent evaluation of the ASI–3, that is, the results for
Step (e). The results for the remaining steps are available in an
appendix (available online), which reports these and other results
pertinent to, but not essential for, an adequate psychometric eval-
uation of the ASI–3.
The selection of fit indices was based on the findings and
recommendations of Hu and Bentler (1998). They recommended
2
We chose to combine the U.S.–Canadian samples and then randomly
split the combined group in two because that should improve the general-
izability (external validity) of the results. That is, the scale construction
sample (Subsample 1) consisted of respondents from several different
regions and universities. This meant that the final 18 items of the ASI–3
were selected to be broadly applicable to many samples instead of being
fitted to one particular sample. The latter can lead to model overfitting (i.e.,
fitting the model to the specific characteristics of a given sample; Brown,
2006), which undermines the generalizability of the results. Accordingly,
Subsamples 1 and 2 were deliberately heterogeneous (in terms of our
U.S.–Canadian recruitment sites), with the goal of creating a scale that
would have a stable factor structure across other samples. As it turned out,
the results from these combined groups were no different from the findings
from other single-site nonclinical groups, that is, the Dutch, French, Mex-
ican, and Spanish groups. In other words, the homogeneity versus heter-
ogeneity of the samples had no effect on the pattern of results.
3
Statistical analyses supported this approach to item selection. Here, all
36 ASI–R items were assigned, on the basis of their content, to Physical,
Cognitive, or Social Concerns subscales. If an item was relevant to more
than one domain (e.g., physical and social concerns), then it was assigned
to more than one subscale. Each subscale was individually factor analyzed
with principal axis factor analysis, and the first factor was extracted. Thus,
three single-factor solutions were obtained, each representing one of the
three domains of physical, cognitive, or social concerns. The 18 items that
were ultimately retained in the ASI–3 all had strong (.50) loadings on
their respective factors. Collectively, the loadings for items included in the
ASI–3 were one standard deviation higher than those of items that were
excluded; respectively, the means (and standard deviations) were .64 (.07)
versus .56 (.09).
4
The main model that was tested, consisting of three correlated factors,
yields identical goodness-of-fit results as a hierarchical model in which the
three factors load on a single higher order factor (Brown, 2006).
178
TAYLOR ET AL.
using at least two indices, one of which is the standardized-root-
mean-square residual (SRMR). Of the other recommended indices,
we selected the root-mean-square error of approximation (RM-
SEA), the comparative fit index (CFI) and the Tucker–Lewis Index
(TLI). The SRMR was used because it is among the most sensitive
to misspecified factor correlations, and the other fit indices were
selected because they are among the most sensitive to misspecified
factor loadings (Hu & Bentler, 1998). To interpret whether a given
factor model provided a good fit to the data, we used Hu and
Bentler’s (1999) empirically derived cutoff values. These values
minimize errors in deciding whether a model provides a good fit to
the data. Good fit is indicated by SRMR ⱕ .08, RMSEA ⱕ .06,
CFI ⱖ .95, or TLI ⱖ .95. For descriptive purposes, chi-square
values are also reported for each confirmatory factor analysis.
For analyses in which two or more competing models were
tested, the relative goodness of fit was tested in several ways. If the
models were nested, then Satorra–Bentler corrected differences in
chi-square values were computed. The 90% confidence interval
(CI) for each RMSEA value was also computed to compare
competing factor models.
Results and Discussion
For the 21 items, the percentage of missing data for a given item
ranged from 0% to 0.3%. The three-factor model for the 21
items—fitted simultaneously to women and men according to Step
(e) described above—yielded two fit indices indicating a good fit:
CFI .972 and TLI .972. The values of the other indices fell
outside the threshold for good fit: SRMR .085, RMSEA .077
(RMSEA 90th percentile CI: .074, .079),
2
(417, N 2,361)
3,305.44, p .001. To improve model fit, we used LISREL model
modification indices to identify the worst-fitting item from each
subscale, that is, the item for which there would be the greatest
drop in chi-square if that item was allowed to cross-load on one or
both of the other factors. This method enabled us to identify items
that measured more than one latent dimension (i.e., items that were
not pure measures of a given factor). Three items were identified
and deleted, and the three-factor model was tested on the remain-
ing 18 items, in the manner described above. With this minor
modification, all four fit indices indicated that the three-factor
(18-item) model had a good fit to the data: CFI .986, TLI
.986, SRMR .060, and RMSEA .058 (90th percentile CI:
.055, .061),
2
(303, N 2,361) 1,487.79, p .001. The items
and their loadings appear in Table 1, which represents the model
simultaneously fitted to the groups of women and men. The three
factors were correlated .70 to .82. When subscale scores were
computed on the basis of the unit-weighted sum of item scores for
a given factor, the subscale correlations ranged from .53 to .62. In
the remaining studies reported in this article, subscale scores were
computed on the basis of unit-weighted sum of items.
A good fit to the data was also obtained for the three-factor
(18-item) model when data from women and men were combined
to form a single group: CFI .986, TLI .984, SRMR .051,
RMSEA .058 (90th percentile CI: .055, .061),
2
(132, N
2,361) 1,163.62, p .001. The three factors were correlated .70
to .81. The matrix of polychoric correlations for Subsample 1
(combining women and men) is shown for descriptive purposes in
Table 2. The mean interitem polychoric correlations for each
subscale were as follows: Physical Concerns (.48), Cognitive
Concerns (.63), and Social Concerns (.48). For each subscale for
this combined sample, coefficient alpha and the range of corrected
item-total Pearson product–moment correlations (r
it
) were as fol
-
lows: Physical Concerns .79, r
it
range .48 to .60; Cognitive
Concerns .84, r
it
range .48 to .67; Social Concerns .79,
r
it
range .46 to .72.
As a further test of whether the three-factor model provided an
optimal fit to the data, we compared it to the one- and two-factor
models, using the two-group (women vs. men) methods described
above. These one- and two-factor models have been obtained in
some previous factor analyses (e.g., Zvolensky et al., 2003), al-
though they have not been as widely replicated as the three-factor
model. For the one-factor model, all 18 items loaded on a single
dimension. The fit values were as follows: CFI .967, TLI
.967, SRMR .083, and RMSEA .088 (90th percentile CI:
.085, .091),
2
(306, N 2,361) 3,094.31, p .001. The
two-factor model consisted of two correlated factors: a Physical
Concerns factor and a combined Cognitive–Social Concerns fac-
tor. The fit values were CFI .971, TLI .971, SRMR .076,
and RMSEA .081 (90th percentile CI: .079, .084),
2
(305, N
2,361) 2,690.52, p .001. Thus, the three-factor model was
associated with better values than the one- and two-factor models
on all four fit indices. Whereas the three-factor model yielded a
good fit to the data on all four indices, the one- and two-factor
models yielded a good fit on only two of four indices. Inspection
of the CIs for the RMSEA indicated that the three-factor model
had a significantly better fit than both the one- and two-factor
models, as indicated by nonoverlapping CIs. The one- and two-
factor models did not differ from one another in their degree of fit
to the data.
The three-factor model also had a superior fit to the one- and
two-factor models in terms of Satorra–Bentler corrected differ-
ences in chi-square values: Factor 1 versus 2,
2
diff
(1, N
2,361) 40.99, p .001; Factor 2 versus 3,
2
diff
(2, N
2,361) 245.34, p .001; and Factor 1 versus 3,
2
diff
(3, N
2,361) 257.31, p .001.
In summary, in the present study an 18-item three-factor model
provided the best fit to the data, for both women and men. The
items from this scale form the ASI–3, which was derived from the
ASI–R and contains five items from the original ASI (one to two
items per subscale, as shown in Table 1). The ASI–3 has an overall
Flesch–Kincaid reading level of Grade 6.5, indicating that it would
be readily comprehended by the majority of adults.
Study 2: Factorial Validity
Factorial validity is a form of construct validity established
through factor analysis. In the case of the ASI–3, this form of
validity was investigated by determining whether the three-factor
model could be replicated across different samples. Two types of
tests of factorial validity were conducted, as part of a stepwise
multigroup confirmatory factor analysis. The more liberal one
involved determining whether the three factors provided a good fit
to the data in separate confirmatory factor analyses of each of six
replication samples, including a sample in which a two-group
(women vs. men) analysis was conducted (i.e., in Subsample 2; the
remaining samples were not large enough for such a two-group
analysis). A more rigorous, multigroup confirmatory factor anal-
ysis was also conducted, in which a common factor structure was
179
ROBUST DIMENSIONS OF ANXIETY SENSITIVITY
simultaneously fitted to all the samples in this article (i.e., Sub-
sample 1 and the six replication samples, without splitting gender
into distinct groups). In this multigroup analysis, the degree of fit
was evaluated for a model that fitted exactly the same factor
structure to each sample (i.e., groups were matched on the size of
factor loadings, factor correlations, and error variances). It was
predicted that the three-factor model would have a good fit to the
data, regardless of the language in which the scale was adminis-
tered and regardless of the patient status of the samples (clinical vs.
nonclinical participants). In addition, it was predicted that the
degree of fit for the three-factor model would be superior to the fit
of the one- and two-factor models. It was also predicted that the
three-factor model of the ASI–3 would show incremental factorial
validity over the three-factor model of the original ASI. That is, the
three-factor model for the ASI–3 was predicted to have better fit
indices than the corresponding model from the ASI, because the
Table 1
Study 1: U.S.–Canadian Subsample 1 (n 2,361)—Loadings (and Standard Errors) for Final, Multigroup Three-Factor Solution of
the ASI–3
Item no. Item
Factor 1:
Physical
Concerns
Factor 2:
Cognitive
Concerns
Factor 3:
Social
Concerns
4 When my stomach is upset, I worry that I might be seriously ill.
a
.79 (.02)
12 When I notice my heart skipping a beat, I worry that there is something seriously
wrong with me.
.76 (.02)
8 When I feel pain in my chest, I worry that I’m going to have a heart attack. .69 (.02)
7 When my chest feels tight, I get scared that I won’t be able to breathe properly. .68 (.02)
15 When my throat feels tight, I worry that I could choke to death. .67 (.02)
3 It scares me when my heart beats rapidly.
a
.66 (.02)
14 When my thoughts seem to speed up, I worry that I might be going crazy. .87 (.01)
18 When my mind goes blank, I worry there is something terribly wrong with me. .84 (.01)
10 When I feel “spacey” or spaced out I worry that I may be mentally ill. .83 (.02)
16 When I have trouble thinking clearly, I worry that there is something wrong with me. .83 (.01)
2 When I cannot keep my mind on a task, I worry that I might be going crazy.
a
.77 (.02)
5 It scares me when I am unable to keep my mind on a task.
a
.62 (.02)
9 I worry that other people will notice my anxiety. .85 (.01)
6 When I tremble in the presence of others, I fear what people might think of me. .79 (.01)
11 It scares me when I blush in front of people. .75 (.02)
13 When I begin to sweat in a social situation, I fear people will think negatively of me. .70 (.02)
17 I think it would be horrible for me to faint in public. .59 (.02)
1 It is important for me not to appear nervous.
a
.54 (.02)
Note. Factor model was simultaneously fitted to the samples of women and men, matching loadings, item errors, and factor correlations. ASI–3 Anxiety
Sensitivity Index–3.
a
Items from the original Anxiety Sensitivity Index.
Table 2
Study 1: Polychoric Correlations Among ASI–3 Items for Subsample 1 (Pooled Across the Samples of Women and Men; n 2,361)
ASI–3
item no. 1234567891011121314151617
2 .27 —
3 .43 .41 —
4 .26 .57 .51 —
5 .20 .52 .26 .43 —
6 .29 .47 .31 .45 .39 —
7 .20 .41 .44 .46 .33 .42 —
8 .21 .45 .46 .50 .35 .36 .51 —
9 .54 .41 .40 .41 .38 .65 .40 .34 —
10 .30 .65 .41 .55 .51 .46 .39 .52 .52 —
11 .38 .45 .43 .43 .33 .52 .31 .30 .62 .54 —
12 .29 .42 .55 .55 .33 .38 .50 .59 .44 .54 .43 —
13 .31 .36 .25 .36 .38 .62 .38 .32 .56 .43 .52 .36 —
14 .30 .65 .40 .61 .48 .49 .41 .51 .52 .75 .51 .55 .47 —
15 .08 .39 .29 .43 .39 .45 .61 .42 .38 .49 .36 .45 .39 .49 —
16 .39 .62 .48 .57 .52 .48 .35 .41 .56 .65 .49 .50 .38 .74 .37 —
17 .45 .23 .35 .31 .16 .40 .27 .22 .59 .30 .39 .33 .37 .35 .22 .42 —
18 .29 .68 .38 .58 .56 .53 .42 .50 .48 .67 .49 .50 .42 .69 .56 .70 .35
Note. See Table 1 for item wording. Given the large sample size, all correlations are significant at p .001 (two-tailed). ASI–3 Anxiety Sensitivity
Index–3.
180
TAYLOR ET AL.
former but not the latter was designed to be a multidimensional
scale.
Method
Participants. As mentioned in Study 1, the U.S.–Canadian
participants (Subsample 2; n 2,359) were mostly undergraduate
students. The mean age was 19.5 years (SD 3.7), and 67% were
women. The other nonclinical participants were undergraduates
recruited from the Universite´ de Lyon II (France; n 701, M
age 20.4 years, SD 4.2, 89% women), University of Mexico
(Mexico; n 418, M age 22.2 years, SD 1.8, 41% women),
University of Groningen (the Netherlands; n 536, M age 20.8
years, SD 2.4, 58% women), and the Universidad Nacional de
Educacio´n a Distancia (Spain; n 480, M age 26.4 years, SD
7.9, 67% women). For the clinical sample (from the United States
and Canada), the sample size was 390, mean age was 36.2 years
(SD 12.0), and 65% were women.
Ethnicity data were formally collected only for the U.S.–
Canadian samples. This was due to the archival nature of the study,
in which data were pooled across sites that differed in the number
of demographic variables they assessed. For U.S.–Canadian Sub-
sample 2, the majority of participants were White (62%), with the
remainder being African American (19%), Asian (7%), Hispanic
(3%), or other (9%). For the clinical sample, the majority were
White (95%), with the remainder being African American (1%),
Asian (2%), Hispanic (1%), or other (2%). The ethnicity of the
other samples largely reflected the country in which the data were
collected; participants from France, the Netherlands, and Spain
were mostly White, and those from Mexico were primarily His-
panic.
For the clinical sample, primary diagnoses (in terms of severity)
were panic disorder (n 143), obsessive–compulsive disorder
(n 104), social anxiety disorder (n 38), generalized anxiety
disorder (n 30), specific phobia (n 29), hypochondriasis (n
15), posttraumatic stress disorder (n 9), trichotillomania (n 5),
major depressive disorder (n 3), behavioral medicine conditions
(e.g., irritable bowel syndrome, pain disorder, Raynaud’s phenom-
enon; n 9), and miscellaneous other conditions (e.g., Tourette’s
syndrome, kleptomania, adjustment disorder; n 5). The propor-
tion of patients taking psychotropic medication, assessed only at
the University of British Columbia (UBC) site, was 48%. Medi-
cations were mostly benzodiazepines (e.g., lorazepam), selective
serotonin reuptake inhibitors (e.g., fluoxetine), or tricyclic antide-
pressants (e.g., imipramine).
Participants who served in Study 1 (Subsample 1) were also
included in those analyses in which the factor solutions for the
replication samples were compared with those of Subsample 1.
Measures and procedure. Participants completed a short ques-
tionnaire assessing demographic features. U.S. participants (clin-
ical and nonclinical) completed the 36-item ASI–R. All other
participants completed the 42-item ASI–R. Participants completed
the measures in classroom or individual settings.
All measures were administered in English, except for those
administered to the samples in France, the Netherlands, Mexico,
and Spain, in which cases the measures were translated into either
French, Dutch, or Spanish. Translations were consistent with con-
temporary guidelines and practices (Butcher & Pancheri, 1976;
Geisinger, 1994). Translators were doctoral-level psychologists
who were native speakers of the dominant language of a given
country and also fluent in English. Translators were knowledge-
able about the culture in which the scales were to be administered
and were also familiar with the nature and assessment of AS. The
scales were independently back-translated to ensure accuracy. All
translators had previous experience in translating scales.
The clinical sample was recruited from cognitive– behavioral
outpatient programs specializing primarily in anxiety disorders at
UBC Hospital (n 155) and the Mayo Clinic Anxiety/Obsessive–
Compulsive Disorder (OCD) program (n 235). All patients were
referred by physician. Patients were diagnosed with an unstruc-
tured interview by the referring physician, who forwarded the
diagnostic information in a referral letter. When patients attended
the cognitive– behavioral outpatient programs, they were rediag-
nosed with structured or unstructured clinical interviews according
to Diagnostic and Statistical Manual of Mental Disorders (DSM–
IV) criteria (American Psychiatric Association, 2000), with all
diagnoses reviewed for accuracy in weekly staff meetings. Sixty-
two percent of UBC patients were diagnosed with the Anxiety
Disorders Interview Schedule for DSM–IV (DiNardo, Brown, &
Barlow, 1994), and 60% of patients at the Mayo Clinic were
assessed with either the Structured Clinical Interview for DSM–IV
(First, Spitzer, Gibbon, & Williams, 1996) or the Mini-
International Neuropsychiatric Interview (Sheehan et al., 1998).
The remaining patients from both sites were diagnosed with an
unstructured clinical interview. Diagnosticians were either
doctoral-level psychologists or predoctoral interns working under
supervision. The reliability of diagnostic procedures has been
previously established for the UBC and Mayo clinics (e.g., Taylor
et al., 1996). Of the 390 clinical participants, 358 completed the
measures before the commencement of cognitive– behavioral treat-
ment. The remaining 32 completed the measures after they had
received a full or partial course of therapy. Inclusion of both types
of patients increased the range of scores for the factor analyses.
Statistical methods. The confirmatory factor analytic proce-
dures in this study were essentially the same as those for the first
study; polychoric correlations with asymptotic weights and robust
weighted least-squares were used, and error terms were not per-
mitted to be correlated. Given that this was a replication study,
model modification indices were not used. The fit indices and
criteria for interpreting these indices were the same as those used
in Study 1. For the ASI–3, the 18-item three-factor model was
tested in a multigroup (women vs. men) analysis for Subsample 2.
This model was tested in single-group analyses for each of the
other replication samples. The model was also tested in a multi-
group analysis, which was conducted across all seven samples (not
split by gender). The latter analysis was a highly stringent test of
the replicability of the ASI–3 factor model; it required that factor
correlations, factor loadings, and error variances were matched
across samples. These analyses were then repeated for the one- and
two-factor models of the ASI–3.
For the analyses comparing the ASI–3 with the ASI, the Cana-
dian participants from Subsamples 1 and 2 were included (n
478) because they—like the French, Dutch, Mexican, and Spanish
samples— completed the 42-item ASI–R, which contains items
that form the ASI–3 and ASI. U.S. participants from Subsamples
1 and 2 completed the 36-item ASI–R, which does not contain all
the ASI items. Similarly, the comparison of ASI–3 and ASI was
limited to Canadian patients (n 155) because they were the only
181
ROBUST DIMENSIONS OF ANXIETY SENSITIVITY
patients who completed the 42-item version of the ASI–R. For the
comparison of ASI–3 and ASI, the latter was scored to yield
subscales measuring physical, cognitive, and social concerns. The
selection of items for the three-factor model for the ASI was based
on previous research (Zinbarg et al., 1997). The items composing
each ASI subscale were as follows: Physical Concerns (3, 4, 6, 8,
9, 10, 11, 14), Cognitive Concerns (2, 12, 15, 16), and Social
Concerns (1, 5, 7, 13).
Results and Discussion
For the 18 ASI–3 items, the percentage of missing data for a
given item ranged from 0% to 0.4% for Subsample 2. The corre-
sponding ranges for the other samples were as follows: clinical
(0% to 0.8%), France (0%), Mexico (0% to 0.5%), the Netherlands
(0% to 0.6%), and Spain (0% to 0.4%). The ranges of item-level
missing data for the ASI were as follows: Canadian participants
from Subsamples 1 and 2 (0% to 0.4%), Canadian participants
from the clinical sample (0% to 1.9%), France (0%), Mexico (0%
to 1.0%), the Netherlands (0% to 0.4%), and Spain (0% to 0.6%).
Table 3 shows the main results of the confirmatory factor
analyses for the replication samples. With the exception of one
sample—the clinical sample—the three-factor model had a good
fit to the data for all fit indices in all analyses. For the clinical
sample, the model had a good fit on three out of four indices; the
exception was the RMSEA.
A series of analyses were conducted to determine why the RMSEA
did not produce an acceptable goodness of fit for the clinical sample.
RMSEA was not improved when eight multivariate outliers were
removed from the data set. RMSEA also did not vary across the site
of recruitment of the clinical samples (Mayo Clinic vs. UBC). We
also examined the model modification indices of the three-factor
model for the clinical sample. These indices indicated that no paths
from factors to items should be added or deleted. This is consistent
with other findings reported in this article, in which it was found that,
for all the samples including the clinical sample, the three-factor
solution had a better degree of fit than the one- and two-factor
solutions. Model modification indices for the three-factor model sug-
gested that goodness of fit would be improved if item uniqueness
terms were allowed to be correlated. This could indicate that more
factors should be specified for the clinical sample. To investigate this
possibility, we conducted an exploratory factor analysis, using prin-
cipal axis factor analysis and oblique rotation. Multiple rules were
used to determine the number of factors to extract: the eigenvalue
greater than one rule, visual inspection of the scree plot, and parallel
analysis. Each indicated a three-factor solution. The pattern of load-
ings indicated good simple structure, with every item of a given
ASI–3 subscale having strong (.50) loadings on only one factor and
small (.20) loadings on the other factors. In other words, the results
clearly supported a factor solution defined by physical, cognitive, and
social concerns. This suggests that the lack of fit on the RMSEA was
not because the type and number of factors for the clinical sample
were different from the factor structure in the other samples.
It is possible that the goodness of fit on the RMSEA for the
clinical sample may have been improved if we had been able to
control for extraneous variables contributing correlations among
the item uniqueness terms. Extraneous influences may include
psychotropic medications and other treatments, which have been
shown to influence AS scores (Taylor, 1999). A small proportion
(8%) of the clinical sample was assessed after partial or full
cognitive– behavioral treatment, whereas the remainder was as-
sessed prior to such treatment. RMSEA results did not change
when we controlled for the effects of this form of treatment.
Unfortunately, however, we were unable to control for the effects
of psychotropic medication. Medication data were not available for
the larger (Mayo Clinic) site. For the UBC site, 48% of patients
were on medication, although data on the dose and duration of
medication were not available. Further research is needed to in-
vestigate the factor structure of the ASI–3 in clinical samples,
particularly research that controls for extraneous variables influ-
encing AS scores such as psychotropic medication.
For the seven-group solution, in which the three-factor model was
simultaneously fitted to Subsample 1 and the six replication samples,
Table 3 shows that the model had a good fit for three of the four fit
indices. Again, the exception was the RMSEA. It is noteworthy,
however, that exactly the same factor model—in terms of loadings,
factor correlations, and error variances—provided a good fit to all
seven samples for the majority of the fit indices.
The results in Table 3 show that, across the various samples, the
three factors tended to be highly correlated with one another.
When the ASI–3 subscales were computed as the unit-weighted
sum of their items, the correlations among subscales tended to be
much lower than the correlations among their corresponding fac-
tors. For each replication sample, the ranges of correlations among
unit-weighted subscales were as follows: Subsample 2 (.50 to .59),
clinical (.41 to .53), France (.26 to .43), Mexico (.59 to .63), the
Netherlands (.44 to .54), and Spain (.37 to .54).
Fit indices for the one- and two-factor models of the ASI–3 for
the replication samples and for the seven-group solution were also
examined and compared with the three-factor model. Recall from
Study 1 that the two-factor model combines the items assessing
cognitive and social concerns into a single factor, with the items
assessing physical concerns forming a separate factor. In the
one-factor model, all items load on a single factor. The results
(appearing in tables in the appendix, available online) indicated
that, for most analyses, the one- and two-factor models poorly
fitted the data and that the goodness of fit was superior for the
three-factor model.
A goal in developing the ASI–3 was to develop a scale with a
more robust (replicable) factor structure than the original ASI.
Accordingly, we expected that the degree of fit of the three-factor
model of the ASI–3 should be better than that of the ASI. Table 4
shows the main results relevant to this prediction for participants
who had completed both scales. Three findings in the table are
noteworthy. First, the fit indices for the ASI–3 tend to be better
than those of the ASI; 26 of 28 indices in the table indicate a good
fit for the ASI–3, whereas a good fit is indicated for 20 of 28
indices for the ASI. Second, the magnitude of the fit indices is
generally better (i.e., higher for the CFI and TLI, and lower for the
SRMR and RMSEA) for the ASI–3 compared with the ASI. This
result holds for 27 of the 28 indices. Third, the ASI–3 and ASI had
nonoverlapping RMSEA CIs in five of seven instances. The latter
indicates that, for most analyses, the three-factor model of the
ASI–3 had a significantly better fit than the corresponding model
of the ASI.
A further series of confirmatory factor analyses examined the one-
and two-factor solutions of the ASI, with these models defined in the
same way that they were for the ASI–3. The results (appearing in the
182
TAYLOR ET AL.
appendix, available online) indicated that the three-factor model of the
ASI had a better fit than the one- and two-factor models. In other
words, of the factor models examined in the present article, the
best-fitting (three-factor) model for the ASI tended to have a poorer fit
to the data than did the corresponding model for the ASI–3. Overall,
the findings support the factorial validity of the ASI–3 and indicate
that this scale has stronger factorial validity than the ASI.
Study 3: Reliability as Internal Consistency
Cronbach’s coefficient alpha was computed for each of the
ASI–3 subscales. Coefficients greater than or equal to .70 were
defined as acceptable, and those greater than or equal to .80 were
defined as good (Nunnally & Bernstein, 1994). It was predicted
that these values would be as good as or better than the corre-
Table 3
Study 2: Fit Indices and Correlations Among Factors for the Three-Factor Model of the ASI–3 for Each Replication Sample and for
All Seven Samples (Scale Construction and Replication Samples)
Sample
Range of
correlations among
factors SRMR
RMSEA
(and 90% CI) CFI TLI
2
(df)
U.S.–Canadian Subsample 2: two-group
analysis (matching women and men to
the same factor model; n 2,359) .69–.78 .061 .055 (.052, .058) .986 .986 1,397.13 (303)
U.S.–Canadian clinical sample (n 390) .45–.61 .061 .090 (.083, .098) .965 .959 550.43 (132)
France (n 701) .32–.59 .054 .043 (.036, .049) .985 .982 298.65 (132)
Mexico (n 418) .74–.86 .053 .040 (.031, .049) .992 .991 220.52 (132)
The Netherlands (n 536) .59–.70 .051 .032 (.023, .040) .994 .994 202.90 (132)
Spain (n 480) .43–.68 .061 .056 (.049, .064) .985 .983 330.95 (132)
7-group solution: fitting the factor model
(factor loadings, item errors, and factor
correlations) to Subsample 1 and to all
6 replication samples (not split by
gender; N 7,245) .70–.79 .101 .078 (.076, .079) .962 .965 8,355.19 (1,158)
Note. ASI–3 Anxiety Sensitivity Index–3; SRMR standardized root-mean-square residual; RMSEA root-mean-square error of approximation;
CI confidence interval; CFI comparative fit index; TLI Tucker-Lewis index; Boldface type fit index indicating that the three-factor model
provided a good fit to the data (SRMR ⱕ .08, RMSEA ⱕ .06, CFI ⱖ .95, or TLI ⱖ .95).
Table 4
Study 2: ASI–3 Versus ASI: Fit Indices and Correlations Among Factors for the Three-Factor Model
Sample
ASI–3 ASI
Range of
correlations
among
factors SRMR
RMSEA
(and 90% CI) CFI TLI
2
(df)
Range of
correlations
among
factors SRMR
RMSEA
(and 90% CI) CFI TLI
2
(df)
Canadian
nonclinical
(n 478)
a
.63–.75 .055 .049 (.041,.049) .989 .987 282.83 (132) .65–.73 .076 .081 (.074,.090) .960 .952 420.47 (101)
Canadian clinical
(n 155)
a
.47–.73 .064 .074 (.060,.089) .979 .976 244.38 (132) .62–.81 .079 .093 (.077,.108) .966 .959 234.02 (101)
France (n 701) .32–.59 .054 .043 (.036,.049) .985 .982 298.65 (132) .57–.72 .066 .063 (.056,.070) .954 .946 380.95 (101)
Mexico (n 418) .74–.86 .053 .040 (.031,.049) .992 .991 220.57 (132) .58–.83 .068 .063 (.054,.072) .973 .968 268.56 (101)
The Netherlands
(n 536) .59–.70 .051 .032 (.023,.040) .994 .994 202.90 (132) .68–.82 .071 .065 (.057,.073) .978 .974 330.04 (101)
Spain (n 480) .43–.68 .061 .056 (.049,.064) .985 .983 330.95 (132) .56–.63 .066 .065 (.055,.074) .976 .972 264.56 (101)
6-group solution:
fitting all the
above samples
simultaneously
to 3-factor
model (n
2,768) .51–.69 .090 .047 (.044,.050) .987 .988 1,973.64 (987) .64–.73 .103 .068 (.065,.071) .964 .967 2,456.91 (781)
Note. Boldface type indicates fit index showing that the three-factor model provided a good fit to the data (SRMR ⱕ .08, RMSEA ⱕ.06, CFI ⱖ .95, or
TLI ⱖ .95). ASI–3 Anxiety Sensitivity Index–3; SRMR standardized root-mean-square residual; RMSEA root-mean-square error of approximation;
CFI comparative fit index; TLI Tucker-Lewis index; CI confidence interval.
a
Of the Canadian and U.S. samples, only the Canadian samples completed all the ASI items, so the U.S. samples were not used in the comparison of the
ASI–3 and ASI factor models.
183
ROBUST DIMENSIONS OF ANXIETY SENSITIVITY
sponding values of the ASI subscales. This is because the ASI was
not designed to be a multidimensional measure, and the Cognitive
and Social Concerns subscales of the ASI each contain only four
items. Although such short scales could, in principle, yield accept-
able alpha values, this would be most likely to occur if the items
had high content validity. Given the concerns raised earlier in this
article about the contents of some of the ASI items, it was expected
that four-item subscales would not be sufficiently long to yield
acceptable alphas.
Method
Participants. Participants were those used in Studies 1 and 2.
Samples were those in which both ASI–3 and ASI data were
available, that is, the Canadian clinical and nonclinical samples
and the samples from France, Mexico, the Netherlands, and Spain.
Measures and procedure. These were the same as in Studies 1
and 2.
Statistical methods. Coefficient alphas for the ASI–3 and ASI
subscales were compared by computing a 95th percentile CI
around each alpha, according to the methods described by Duh-
achek and Iacobucci (2004).
Recall that the ASI–3 and ASI can be conceptualized as having
hierarchic structures in which the three lower order factors load on
a common, general AS factor. Total scores on the ASI–3 and ASI
are useful for assessing the general factor, providing that the
general factor accounts for a substantial proportion of variance. To
investigate this for each of the six samples in which data on the
ASI–3 and ASI were available, we calculated the proportion of
variance due to the general factor (known as
H
; Zinbarg, Yovel,
Revelle, & McDonald, 2006), and we also calculated the propor-
tion of variance due to the lower order factors, separate from the
variance due to the general factor. This was done by performing a
Schmid–Leiman analysis of the loadings from the confirmatory
factor analyses reported earlier in this article, using the procedures
described by Zinbarg et al. (2006).
Results and Discussion
The proportions of missing data were the same as for Study 2.
Table 5 shows the coefficient alphas for the ASI–3 and ASI
subscales, along with the corresponding 95th percentile CIs. The
values for the ASI–3 were all in the range considered to be
acceptable or good (Nunnally & Bernstein, 1994), which supports
the intended use of these subscales as research instruments. The
ASI–3 and ASI did not differ in the magnitude of coefficients for
the Physical Concerns subscales. However, the ASI–3 generally
had significantly larger coefficients than did the ASI for the
Cognitive Concerns and Social Concerns subscales.
For each of the six samples, the proportion of variance due to the
general factor (
H
) of the ASI–3 was consistently slightly higher
than that of the ASI (respectively, M 0.36, SD 0.06, and M
0.33, SD 0.06). For each sample and scale, the total proportion
of variance from all three lower order factors was calculated, after
controlling for variance due to the general factor. Across all six
samples, the proportions of variance for ASI–3 were consistently
higher than those for the ASI (respectively, M 0.40, SD 0.06,
and M 0.33, SD 0.06). In summary, results indicate that for
the ASI–3, about 36% of the variance in scale scores is due to a
general AS factor, and 40% of variance is due to the lower order
factors (i.e., 76% total explained variance and 24% error variance).
The corresponding figures for the ASI were 33% and 33% (66%
explained variance, 34% error variance). These results support the
use of total and subscale scores of the ASI–3 and also show that
scores on the ASI–3 contain less error variance than those on the
ASI.
Study 4: Convergent, Discriminant, and Criterion-Related
Validities
Convergent validity of the ASI–3 was examined by intercorre-
lating the subscales of the ASI–3 and ASI. It was predicted that
similar subscales (e.g., ASI–3 and ASI Physical Concerns sub-
scales) would be highly correlated (r ⱖ .50; i.e., large correlations
according to Cohen, 1988). Discriminant validity for a measure of
a given construct would be supported when the measure is more
highly correlated with similar or theoretically related constructs
than when it is correlated with dissimilar or theoretically unrelated
constructs. Accordingly, it was predicted that similar subscales
(e.g., ASI–3 and ASI Physical Concerns) would be more highly
correlated than dissimilar subscales (e.g., ASI–3 Physical Con-
cerns correlated with ASI Cognitive Concerns or with ASI Social
Concerns). The ASI–3 and ASI contain five overlapping items,
which would inflate correlations between ASI–3 and ASI sub-
scales. Therefore, in conducting tests of convergent and discrimi-
nant validity, we removed the overlapping items from the ASI. We
did this because the ASI–3 subscales were the focus of investiga-
tion, so it would be inappropriate to eliminate the overlapping
items from the ASI–3.
Regarding criterion-related (known groups) validity, we tested
predictions about how scores on the ASI–3 subscales would differ
across different groups. U.S.–Canadian data from students (i.e.,
nonclinical controls [NC]) and pretreatment data from the four
largest clinical groups—panic disorder (PD), OCD, social anxiety
disorder (SANX), and generalized anxiety disorder (GAD)—
enabled us to test a prediction for each of the ASI–3 subscales. The
first prediction had two parts: (a) that the PD group would be
associated with higher scores on the Physical Concerns subscales
than would all other groups and (b) that the remaining clinical
groups would have higher scores than would the NC group. These
predictions were based on prior theory and research implicating
the importance of heightened physical concerns in PD and other
anxiety disorders (Taylor, 1999). Research indicates that physical
concerns are elevated in these disorders compared with NCs, with
PD tending to be associated with the highest scores (Taylor, 1999).
The second prediction also had two parts: (a) SANX, compared
with all other groups, would be associated with higher scores on
social concerns, and (b) the other clinical groups would score
higher on social concerns than would the NCs. The rationale was
that SANX, compared with other anxiety disorders and NCs, is
characterized by strong fear of negative evaluation (American
Psychiatric Association, 2000). Therefore, SANX should be asso-
ciated with higher scores than the other groups. In turn, GAD,
OCD, and PD should be associated with higher scores than NCs,
because social fears tend to be greater in these groups compared
with NCs (e.g., Rapee, Sanderson, & Barlow, 1988).
The third prediction was that each of the clinical groups would
score higher than NCs on the Cognitive Concerns subscale. The
184
TAYLOR ET AL.
rationale was that all four of the anxiety disorders have been found
to be associated with particular forms of cognitive concern. This
prediction was based on research that used measures of AS in PD
studies and on research on other disorders with the use of instru-
ments similar to the Cognitive Concerns subscale (e.g., measures
of beliefs about the importance of controlling one’s thoughts in
OCD research). To illustrate the various forms of cognitive con-
cerns, researchers have found PD to be associated with strong fears
of cognitive phenomena (e.g., derealization), associated with be-
liefs that these phenomena have catastrophic consequences, such
as permanent mental incapacitation (Taylor, 1999). Similarly,
SANX is associated with fears that one may not be able to perform
adequately in social situations, such as fears that one’s mind will
go blank (Clark & Wells, 1995). This is a situationally specific
form of cognitive concern, which should contribute to high scores
on the Cognitive Concerns subscale. Research has shown that
people with OCD, compared with control participants, are more
likely to overestimate the dangerousness of cognitive dyscontrol,
which is consistent with contemporary cognitive models of OCD
(Frost & Steketee, 2002), which propose that OCD arises, in part,
from such distorted beliefs about the harmful consequences of
cognitive dyscontrol. Similarly, theory and research indicate that
GAD is associated with “meta-worry,” that is, worry about the
deleterious effects of uncontrollable worry (Wells, 2005).
Empirical research has shown that one particular facet of AS—
physical concerns—is especially elevated in PD compared with
most other anxiety disorders (although it is not clear that people
with PD score any higher than those with posttraumatic stress
disorder; Taylor, 1999). It is unclear whether PD is associated with
higher scores, compared with other disorders, on the Cognitive and
Social Concerns subscales. It is for these reasons that we chose not
to include an a priori prediction about group differences in ASI–3
total score. Previous research on this issue has been based largely
on the ASI. Total scores on that scale are weighted toward the
assessment of physical concerns, which represent half of the items
of the scale. Therefore, previous findings that PD is associated
with the highest ASI total score (Taylor, 1999) might simply
reflect the ASI’s overemphasis on physical concerns. We also did
not compare the known groups validity of the ASI–3 with that of
the ASI. This was because there were too few clinical participants
who completed both measures (n 155), which meant that the
corresponding clinical groups would be too small for analysis.
Method
Participants. Participants were those used in Studies 1 and 2.
For the analyses involving group comparisons, the samples used
are shown in Table 6. For the other analyses, samples were those
Table 5
Study 3: Subscale Reliability for the ASI–3 Versus ASI: Coefficient Alphas and Their 95th Percentile Confidence Intervals
Sample
Physical Concerns Cognitive Concerns Social Concerns
ASI–3 ASI ASI–3 ASI ASI–3 ASI
Canadian nonclinical (n 478) .79 (.76, .82) .83 (.81, .86) .83 (.81, .86) .77 (.74, .80) .78 (.75, .81) .53 (.46, .60)
Canadian clinical (n 155) .86 (.83, .90) .89 (.86, .92) .91 (.89, .93) .84 (.80, .88) .86 (.83, .89) .60 (.50, .70)
France (n 701) .76 (.73, .79) .76 (.73, .79) .79 (.77, .82) .68 (.64, .72) .75 (.72, .78) .41 (.34, .49)
Mexico (n 418) .83 (.81, .86) .81 (.78, .84) .82 (.79, .85) .77 (.73, .81) .73 (.69, .77) .53 (.46, .60)
The Netherlands (n 536) .80 (.77, .82) .83 (.81, .85) .81 (.79, .84) .80 (.77, .83) .76 (.72, .79) .44 (.36, .52)
Spain (n 480) .84 (.81, .86) .82 (.80, .85) .87 (.85, .89) .81 (.78, .84) .84 (.81, .86) .66 (.61, .71)
Note. Confidence intervals are in parentheses. For pairwise comparisons between ASI-3 and ASI subscales, for a given sample and content domain (e.g.,
Physical Concerns), the significantly larger alpha is in boldface type, as indicated by nonoverlapping confidence intervals. ASI–3 Anxiety Sensitivity
Index–3.
Table 6
Study 4: Means (and Standard Deviations) on the ASI–3 Subscales
Sample Physical Concerns Cognitive Concerns Social Concerns Total score
Nonclinical groups
U.S.–Canada: women (n 3,153) 4.3 (4.2) 2.6 (3.8) 5.9 (4.7) 12.8 (10.5)
U.S.–Canada: men (n 1,567) 3.9 (4.2) 2.8 (3.8) 6.0 (4.8) 12.8 (10.8)
U.S.–Canada: total (n 4,720) 4.2 (4.2) 2.7 (3.8) 5.9 (4.7) 12.8 (10.6)
France (n 701) 5.0 (4.0) 2.8 (3.4) 8.5 (4.8) 16.4 (9.1)
Mexico (n 418) 5.5 (4.8) 3.5 (4.1) 6.1 (4.3) 15.2 (11.3)
The Netherlands (n 536) 2.7 (3.2) 1.7 (2.8) 5.7 (4.0) 10.7 (8.1)
Spain (n 480) 4.5 (3.9) 2.8 (3.7) 6.9 (4.7) 14.2 (9.8)
Selected clinical groups (pretreatment)
a
Panic disorder (n 120) 11.3 (6.7) 9.0 (6.4) 12.3 (5.8) 32.6 (14.3)
Obsessive-compulsive disorder (n 102) 8.3 (6.2) 7.7 (6.0) 10.3 (6.7) 26.3 (16.8)
Social anxiety disorder (n 38) 6.2 (4.5) 7.9 (6.1) 17.3 (4.8) 31.4 (11.9)
Generalized anxiety disorder (n 30) 8.1 (5.3) 8.9 (7.4) 10.5 (7.0) 27.5 (16.5)
a
Selected for inclusion in tests of criterion-related (known groups) validity. ASI–3 Anxiety Sensitivity Index–3.
185
ROBUST DIMENSIONS OF ANXIETY SENSITIVITY
in which both ASI–3 and ASI data were available, that is, the
Canadian clinical and nonclinical samples, as well as the samples
from France, Mexico, the Netherlands, and Spain.
Measures and procedure. These were the same as in Studies 1
and 2.
Statistical methods. For the correlational and group compari-
sons analyses, items were summed (unit-weighted) to form ASI–3
and ASI subscales. All correlations were Pearson product–moment
correlations unless stated otherwise. As mentioned, items that
overlapped with ASI–3 were omitted from the ASI subscales. To
disentangle the evaluation of convergent and discriminant validity
from effects due to subscale reliability, we disattenuated the cor-
relations in the validity tests for unreliability (Nunnally & Bern-
stein, 1994). That is, each correlation between a pair of measures
was divided by the square root of the product of the reliability ()
of each measure. Tests of differences between correlations were
computed according to the methods described by Meng,
Rosenthal, and Rubin (1992).
Regarding the tests of group differences in ASI–3 subscale
scores (i.e., known groups validity), conventional multiple com-
parisons (e.g., Tukey’s comparisons) were inappropriate because
our data violated the assumptions of equal sample sizes and
homogeneous variances. A further problem was that conventional
multiple comparisons yield intransitive results. That is, they often
yield ambiguous or logically impossible results. To illustrate, for a
three-group comparison of three means {1,2,3}, ordered from
highest to lowest, multiple comparisons such as the Tukey’s test
often yield results such as the following: 1 2, 2 3, and yet 1
3. This intransitivity adds confusion to the interpretation of the
findings. To circumvent all of these problems, we used the paired-
comparisons information criterion (PCIC; Dayton, 2003), which
computes the Akaike Information Criterion (AIC) for all logically
possible subsets of groups. For example, for three groups, ordered
from highest to lowest scores, there would be the following “mod-
els,” in which the commas indicate that a group differs from other
groups: {123} (null model in which there are no group differ-
ences), {1,23}, {12,3}, and {1,2,3}. The model with the smallest
AIC represents the best-fitting model. The PCIC does not involve
computing statistical significance tests, so the control of Type I
and II error is not an issue. The method does not produce intran-
sitive results and does not require that sample sizes or variances be
equal across groups. The AIC has a slight bias for selecting more
complicated models than the true model, and so an omnibus test
(e.g., multivariate analysis of variance) is used to test the null
hypothesis of no group differences. Monte Carlo studies indicate
that the protected PCIC is superior to other multiple comparison
procedures in correctly identifying patterns of group differences
(Dayton, 2003).
5
Results and Discussion
Preliminary analyses. The proportions of missing data were
the same as for Study 2. For descriptive purposes, ASI–3 and ASI
correlations with demographic variables were computed. Across
six samples (Canadian students, Canadian patients, and samples
from France, Mexico, the Netherlands, and Spain), the ASI–3 (and
ASI) total score correlations with age ranged from .09 to .05
(.11 to .08). Polyserial correlations with gender (scored
women 1, men 2) ranged from .16 to .07 (.22 to .08).
Correlations with education level,
6
available only for the Canadian
samples, ranged from .06 to .15 (.06 to .12). Correlations with
ethnicity (Non-White 1, White 2), available only for the
Canadian samples, ranged from .22 to .12 (.19 to .12). In
summary, the patterns of correlations for the ASI–3 and ASI were
similar to one another. Although some of the correlations were
statistically significant because of the large sample sizes, all cor-
relations were in the range that would be classified as small or
trivial, according to Cohen’s (1988) classification scheme. In other
words, the ASI–3 performed similarly to the ASI; both were
largely unrelated to demographic variables.
Convergent and discriminant validity. Table 7 shows that
when the subscales were corrected for less-than-perfect reliability,
each ASI–3 subscale measured essentially the same content do-
main (e.g., physical concerns) as its ASI counterpart, as indicated
by correlations that approach unity. These large correlations sup-
port the convergent validity of the ASI–3 subscales. Correlations
between similar subscales from the ASI–3 and ASI (e.g., Physical
Concerns) were, for almost all analyses, significantly larger than
the correlations between dissimilar subscales (e.g., ASI–3 Physical
Concerns correlated with either ASI Cognitive or Social Con-
cerns). This is shown by the results for the planned contrasts (Z
scores) in Table 7 and provides supporting evidence of the dis-
criminant validity of the ASI–3 subscales.
Criterion-related (known groups) validity. A multivariate
analysis of variance was conducted as an omnibus test, in which
the dependent variables were the three subscale scores on the
ASI–3, and the independent variable (group) was defined by the
five groups (PD, OCD, GAD, SANX, and NC). The group factor
for the omnibus test was significant, Pillai F(12, 15,012) 67.74,
p .001,
2
.05. The corresponding means (and standard
deviations) appear in Table 6. The best-fitting results of the PCIC
comparisons were as follows: for the ASI–3 Physical Concerns
subscale, PD (OCD, GAD) SANX NC (
2
.08). For
Cognitive Concerns, (PD, GAD) (OCD, SANX) NC (
2
.10). For Social Concerns, SANX PD (OCD, GAD) NC
(
2
.09). Thus, each of the three predictions was supported
regarding the criterion-related validity of the ASI–3 subscales.
General Discussion
As researchers increasingly focus their attention on dimensions
of AS, rather than simply looking at AS as a global construct, it is
important to develop robust, psychometrically sound measures of
5
PCIC was used instead of planned contrasts because of the violation of
the assumptions of equal sample size and homogeneity of variance. A
further reason for not using planned contrasts was because the latter are
performed in the context of null hypothesis significance testing, in which
the purpose of the contrasts is to reduce the odds of Type II error by
conducting comparisons only among particular pairs of means instead of
among all possible pairs. Type II and Type I error rates are not relevant to
PCIC (Dayton, 2003).
6
Education level was rated by respondents on the following scale: 1
Grade 6 or less;2 Grades 7–12, without graduating from high school;
3 high school or equivalent;4 partial college;5 graduated from a
2-year college program;6 graduated from a 4-year college or university
program;7 partial graduate or professional school; and 8 completed
graduate or professional school.
186
TAYLOR ET AL.
these dimensions. The present study developed and evaluated such
a scale, which measures the three most widely replicated dimen-
sions: physical, cognitive, and social concerns. This scale, called
the ASI–3, had a stable three-factor structure across gender and
across seven different samples, according to most fit indices. The
findings were especially encouraging given that they were ob-
tained from different populations (clinical vs. nonclinical) and
from different countries and different language versions of the
measures. The ASI–3 shares 5 of its 18 items with the ASI, with
1 to 2 overlapping items on each of the ASI–3’s 6-item subscales.
The overlap is a result of the scale construction process for the
ASI–3, in which we selected, on the basis of content validity and
psychometric properties, the best of the ASI items for inclusion in
the ASI–3.
Although the ASI–3 is only two items longer than the original
ASI, it tended to be a better measure of the AS dimensions, as
assessed by reliability as internal consistency and by factorial
validity. Evidence also supported the convergent, discriminant,
and criterion-related (known groups) validities of the ASI–3. The
findings suggest that the ASI–3 would be preferable to the ASI in
studies of the dimensions of AS. The ASI–3 subscales may also be
preferable to the ASI subscales in taxometric studies because of
the greater reliability of the former. Subscales are commonly used
as indicators in taxometric studies, and indicator reliability influ-
ences the taxometric validity of the indicators, that is, their ability
to distinguish between taxonic groups (Meehl, 1992).
The multisample data set used in this article has both strengths
and limitations. The sample was clinically and nationally diverse,
and to our knowledge it is the largest data set to have ever been
used in AS research. However, the items for the ASI–3 were
embedded in the larger pool of items forming the ASI–R, and so
it remains to be seen whether the results of the present study
generalize to situations in which the ASI–3 is administered as a
stand-alone instrument. Nevertheless, it is noteworthy that the
original ASI was also embedded among the ASI–R items, and the
performance of the ASI in the present study resembled that found
in many other studies (e.g., the identification of a three-factor
solution; Taylor, 1999).
Further research is needed to more fully evaluate the validity of
the ASI–3. Studies are required to test whether, across different
types of samples, the ASI–3 subscales are more strongly correlated
with theoretically related variables (e.g., anxiety-related variables)
than with theoretically unrelated variables (e.g., extraversion or
other traits that are not correlated with the tendency to experience
negative emotions). Laboratory studies, such as symptom-
provocation studies using carbon dioxide inhalation (which in-
duces intense dyspnea), could also be conducted to test the con-
vergent validity of the ASI–3. Studies of the ASI have shown that
scores on this scale predict anxiety evoked by carbon dioxide
inhalation (Taylor, 1999). The same is predicted for the ASI–3.
The ASI–3’s test–retest reliability remains to be studied. Pro-
spective studies are required to determine whether the ASI–3, like
the ASI, predicts the risk of psychopathology such as panic at-
tacks, and whether the prediction of psychopathology differs
among the ASI–3 subscales. If the ASI–3 proves to be sensitive to
treatment effects, then this short scale could be repeatedly admin-
Table 7
Study 4: Pearson Product–Moment Correlations Among ASI–3 and ASI Subscales, With (and Without) Correction for Attenuation
Sample and subscale
ASI–3 and ASI
Planned contrasts among correlations
that were corrected for attenuation
Physical Cognitive Social Contrast Z
*
Canadian nonclinical (n 478)
Physical .93 (.73) .69 (.49) .72 (.37) Phys vs. Cog, Soc 20.01
Cognitive .62 (.50) .99 (.73) .67 (.36) Cog vs. Phys, Soc 77.18
Social .62 (.49) .76 (.53) .99 (.54) Soc vs. Phys, Cog 72.15
Canadian clinical (n 155)
Physical .99 (.85) .66 (.53) .71 (.42) Phys vs. Cog, Soc 31.66
Cognitive .72 (.64) .94 (.78) .92 (.56) Cog vs. Phys, Soc 6.36
Social .59 (.50) .74 (.60) .99 (.67) Soc vs. Phys, Cog 46.41
France (n 701)
Physical .92 (.67) .55 (.35) .64 (.27) Phys vs. Cog, Soc 28.36
Cognitive .61 (.45) .98 (.64) .59 (.25) Cog vs. Phys, Soc 49.16
Social .47 (.34) .54 (.34) .99 (.52) Soc vs. Phys, Cog 106.70
Mexico (n 418)
Physical .92 (.72) .74 (.57) .74 (.40) Phys vs. Cog, Soc 17.25
Cognitive .77 (.60) .93 (.71) .65 (.35) Cog vs. Phys, Soc 20.39
Social .85 (.63) .74 (.53) .99 (.54) Soc vs. Phys, Cog 63.26
The Netherlands (n 536)
Physical .96 (.75) .58 (.44) .83 (.36) Phys vs. Cog, Soc 62.59
Cognitive .70 (.55) .83 (.64) .91 (.40) Cog vs. Phys, Soc 0.03
Social .59 (.45) .63 (.47) .99 (.48) Soc vs. Phys, Cog 206.48
Spain (n 480)
Physical .98 (.79) .61 (.45) .53 (.38) Phys vs. Cog, Soc 39.66
Cognitive .73 (.59) .86 (.64) .54 (.39) Cog vs. Phys, Soc 11.84
Social .59 (.47) .55 (.41) .92 (.65) Soc vs. Phys, Cog 21.44
Note. Item overlap between the ASI–3 and ASI was eliminated by omitting overlapping items from the ASI. ASI Anxiety Sensitivity Index; Phys
Physical; Cog Cognitive; Soc Social.
*
All ps .001, apart from Z 0.03, for which p .10.
187
ROBUST DIMENSIONS OF ANXIETY SENSITIVITY
istered during treatment studies to investigate the relative effects of
treatments on the three dimensions of AS and to study the medi-
ators and moderators of treatment-related changes in AS. Finally,
further studies are needed to fully investigate the cross-cultural
similarities and differences in AS.
References
American Psychiatric Association. (2000). Diagnostic and statistical man-
ual of mental disorders (4th ed., text rev.). Washington, DC: Author.
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.
Blais, M. A., Otto, M. W., Zucker, B. G., McNally, R. J., Schmidt, N. B.,
Fava, M., et al. (2001). The Anxiety Sensitivity Index: Item analysis and
suggestions for refinement. Journal of Personality Assessment, 77, 272–
294.
Brown, T. A. (2006). Confirmatory factor analysis for applied research.
New York: Guilford Press.
Butcher, J. N., & Pancheri, P. (1976). Handbook of cross-national MMPI
research. Minneapolis, MN: University of Minnesota Press.
Clark, D. M., & Wells, A. (1995). A cognitive model of social phobia. In
R. G. Heimberg, M. Liebowitz, D. Hope, & F. Schneier (Eds.), Social
phobia: Diagnosis, assessment, and treatment (pp. 69–93). New York:
Guilford Press.
Cohen, J. (1988). Statistical power analyses for the behavioral sciences
(2nd ed.). Hillsdale, NJ: Erlbaum.
Dayton, C. M. (2003). Information criteria for pairwise comparisons.
Psychological Methods, 8, 61–71.
Deacon, B. J., Abramowitz, J. A., Woods, C. M., & Tolin, D. F. (2003).
The Anxiety Sensitivity Index—Revised: Psychometric properties and
factor structure in two nonclinical samples. Behaviour Research and
Therapy, 41, 1427–1449.
Dehon, C., Weems, C. F., Stickle, T. R., Costa, N. M., & Berman, S. L.
(2005). A cross-sectional evaluation of the factorial invariance of anx-
iety sensitivity in adolescents and young adults. Behaviour Research and
Therapy, 43, 799 – 810.
DiNardo, P., Brown, T. A., & Barlow, D. H. (1994). Anxiety Disorders
Interview Schedule for DSM–IV. New York: Graywind.
Duhachek, A., & Iacobucci, D. (2004). Alpha’s standard error (ASE): An
accurate and precise confidence interval estimate. Journal of Applied
Psychology, 89, 792– 808.
First, M. B., Spitzer, R. L., Gibbon, M., & Williams, J. B. W. (1996).
Structured Clinical Interview for Axis I DSM–IV disorders—Patient
edition. New York: Biometrics Research Department, New York State
Psychiatric Institute.
Frost, R. O., & Steketee, G. (2002). Cognitive approaches to obsessions
and compulsions: Theory, assessment and treatment. Oxford, England:
Elsevier.
Geisinger, K. F. (1994). Cross-cultural normative assessment: Translation
and adaptation issues influencing the normative interpretation of assess-
ment instruments. Psychological Assessment, 6, 304 –312.
Hu, L., & Bentler, P. M. (1998). Fit indices in covariance structure
modeling: Sensitivity to underparameterized model misspecification.
Psychological Methods, 3, 424 – 453.
Hu, L., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance
structure analysis: Conventional criteria versus new alternatives. Struc-
tural Equation Modeling, 6, 1–55.
Jo¨reskog, K., & So¨rbom, D. (2005). LISREL 8.72 [manual and computer
program]. Lincolnwood, IL: Scientific Software International.
Meehl, P. E. (1992). Factors and taxa, traits and types, differences of
degree and differences in kind. Journal of Personality, 60, 117–174.
Meng, X.-L., Rosenthal, R., & Rubin, D. B. (1992). Comparing correlated
correlation coefficients. Psychological Bulletin, 111, 172–175.
Nunnally, J. C., & Bernstein, I. H. (1994). Psychometric theory (3rd ed.).
New York: McGraw-Hill.
Peterson, R. A., & Reiss, S. (1992). Anxiety Sensitivity Index revised
manual. Worthington, OH: IDS Publishing.
Rapee, R. M., Sanderson, W. C., & Barlow, D. H. (1988). Social phobia
features across the DSM–III–R anxiety disorders. Journal of Psychopa-
thology and Behavioral Assessment, 10, 287–299.
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.
Roth, D. A., Coles, M. E., & Heimberg, R. G. (2002). The relationship
between memories of childhood teasing and anxiety and depression in
adulthood. Journal of Anxiety Disorders, 16, 149 –164.
Sheehan, D. V., Lecrubier, Y., Sheehan, K. H., Amorim, P., Janavs, J.,
Weiller, E., et al. (1998). The Mini-International Neuropsychiatric In-
terview (M.I.N.I.): The development and validation of a structured
diagnostic psychiatric interview for DSM–IV and ICD–10. Journal of
Clinical Psychiatry, 59(Suppl. 20), 22–33.
Stewart, S. H., Taylor, S., & Baker, J. M. (1997). Gender differences in
dimensions of anxiety sensitivity. Journal of Anxiety Disorders, 11,
179 –200.
Taylor, S. (1999). Anxiety sensitivity: Theory, research, and treatment of
the fear of anxiety. Mahwah, NJ: Erlbaum.
Taylor, S., & Cox, B. J. (1998a). An expanded Anxiety Sensitivity Index:
Evidence for a hierarchic structure in a clinical sample. Journal of
Anxiety Disorders, 12, 463– 484.
Taylor, S., & Cox, B. J. (1998b). Anxiety sensitivity: Multiple dimensions
and hierarchic structure. Behaviour Research and Therapy, 36, 37–51.
Taylor, S., Koch, W. J., Woody, S., & McLean, P. (1996). Anxiety
sensitivity and depression: How are they related? Journal of Abnormal
Psychology, 105, 474 – 479.
Wells, A. (2005). The metacognitive model of GAD: Assessment of
meta-worry and relationship with DSM–IV generalized anxiety disorder.
Cognitive Therapy and Research, 29, 107–121.
Zinbarg, R. E., Barlow, D. H., & Brown, T. A. (1997). The hierarchical
structure and general factor saturation of the Anxiety Sensitivity Index:
Evidence and implications. Psychological Assessment, 9, 277–284.
Zinbarg, R. E., Yovel, I., Revelle, W., & McDonald, R. P. (2006). Esti-
mating generalizability to a latent variable common to all of a scale’s
indicators: A comparison of estimators for
H
. Applied Psychological
Measurement, 30, 121–144.
Zvolensky, M. J., Arrindell, W. A., Taylor, S., Bouvard, M., Cox, B. J.,
Stewart, S. H., et al. (2003). Anxiety sensitivity in six countries. Behav-
iour Research and Therapy, 41, 841– 859.
Received November 28, 2005
Revision received December 15, 2006
Accepted December 22, 2006 䡲
188
TAYLOR ET AL.
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.
