- Email: [email protected]
Structure of the Anxiety Sensitivity Index psychometrics and factor structure in a community sample
Anxiety Disorders 16 (2002) 33±49
Structure of the Anxiety Sensitivity Index psychometrics and factor structure in a community sample Norman B. Schmidta,*, Thomas E. Joinerb a
Department of Psychology, The Ohio State University, 245 Townshend Hall, Columbus, OH 43210, USA b Department of Psychology, Florida State University, Tallahassee, FL, USA
Received 13 January 2000; received in revised form 14 June 2000; accepted 30 June 2000
Abstract Converging evidence suggests that anxiety sensitivity (i.e., threatening beliefs regarding autonomic arousal) is a risk factor for anxiety pathology. Speci®cation of premorbid risk factors requires exclusion of individuals with a history of spontaneous panic to ensure that anxiety sensitivity is not merely a consequence or concomitant of the experience of panic. However, the psychometrics and dimensional nature of anxiety sensitivity in such a sample is undetermined. The present study evaluated the factor structure of the Anxiety Sensitivity Index (ASI), a measure of anxiety sensitivity, in a community sample (N 233) with no history of psychiatric illness or spontaneous panic. Exploratory factor analyses (EFA) suggested a two- or three-factor solution (I, Fear of Mental Catastrophe; II, Fear of Cardiopulmonary Sensations; III, Fear of Vasovagal Sensations). Con®rmatory factor analyses (CFA) comparing alternative models indicated that a hierarchical two-factor solution (I, Fear of Mental Catastrophe; II, Fear of Cardiopulmonary Sensations) best accounted for the data. This model generalized well to a nonclinical college sample (N 809). # 2002 Elsevier Science Inc. All rights reserved. Keywords: Anxiety sensitivity; Factor analysis; Anxiety; Community sample
*
Corresponding author. Tel.: 1-614-292-2687. E-mail address: [email protected] (N.B. Schmidt).
0887-6185/02/$ ± see front matter # 2002 Elsevier Science Inc. All rights reserved. PII: S 0 8 8 7 - 6 1 8 5 ( 0 1 ) 0 0 0 8 7 - 1
34
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
1. Introduction Anxiety sensitivity has played an important role in the conceptualization and research on anxiety pathology (McNally, 1990). Anxiety sensitivity refers to the extent to which an individual believes that autonomic arousal can have aversive or harmful consequences (Reiss & McNally, 1985). For example, individuals with high anxiety sensitivity may believe that shortness of breath signals suffocation or that heart palpitations indicate a heart attack, whereas those with low anxiety sensitivity experience these sensations as unpleasant but nonthreatening. Consistent with cognitive theories of anxiety, the anxiety sensitivity conceptualization posits that cognitive misappraisal is critical for generation of anxiety. However, anxiety sensitivity is distinguished from other cognitive conceptualizations because anxiety sensitivity is believed to be a stable trait-like characteristic that may precede development of panic attacks. Individual differences in anxiety sensitivity are hypothesized to emerge from a variety of experiences that ultimately lead to acquisition of beliefs about potentially aversive consequences of arousal. Such experiences may include hearing others express fear of such sensations, receiving misinformation about the harmfulness of certain sensations, witnessing a catastrophic event such as the fatal heart attack of a loved one, and so forth. Thus, anxiety sensitivity constitutes a disposition to developing anxiety and does not require the experience of clinically signi®cant anxiety or panic in its own development (Reiss, 1991). One of the important and unique predictions of the anxiety sensitivity conceptualization is that anxiety sensitivity should act as a risk factor for development of panic attacks as well as related anxiety disorders. Both laboratory-based experimental studies as well as prospective quasi-experimental studies have supported this prediction. Experimental studies using nonclinical samples with no history of panic attacks have indicated that anxiety sensitivity is predictive of fearful responding in the context of hyperventilation, caffeine, and carbon dioxide inhalation (Donnell & McNally, 1990; Harrington, Schmidt, & Telch, 1996; Rapee & Medoro, 1994; Schmidt & Telch, 1994; Telch, Silverman, & Schmidt, 1996). Prospective evaluation of nonclinical samples has determined that anxiety sensitivity is a risk factor for development of spontaneous panic attacks (Schmidt, Lerew, & Jackson, 1997) as well as anxiety disorders (Maller & Reiss, 1992). Taken together, laboratory and prospective studies provide converging evidence for anxiety sensitivity as a risk factor in the development of anxiety pathology. Anxiety sensitivity was originally proposed as a unitary construct (Reiss & McNally, 1985). Empirical evaluation of this proposition using the most widely used measure of anxiety sensitivity, the Anxiety Sensitivity Index (ASI), has yielded inconsistent ®ndings. Several early factor analytic studies provided support for the unitary nature of anxiety sensitivity (Reiss, Peterson, Gursky, & McNally, 1986; Taylor, Koch, & Crockett, 1991). Others found that the ASI is multifactorial (Telch, Shermis, & Lucas, 1989; Wardle, Ahmad, & Hayward,
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
35
1990). Recently, Blais et al. (in press) determined that the ASI is best described by a two-factor solution consisting of fears of physical sensations related to anxiety and Fear of Mental Catastrophe. Two studies using con®rmatory factor analytic (CFA) techniques have attempted to clarify the factor structure of the ASI. CFA techniques provide a method for evaluating competing factor solutions by examining goodness-of-®t indices for various models. Taylor, Koch, McNally, and Crockett (1992) concluded that the ASI was best viewed as a single-factor measure even though an orthogonal four-factor model (i.e., Telch, Shermis, & Lucas, 1989) resulted in a similar goodness-of-®t. Taylor et al. argued that the four-factor solutions obtained by Telch, Shermis, and Lucas (1989). and Wardle et al. (1990) were artifacts of the orthogonality constraint because signi®cant interfactor correlations resulted when this constraint was removed. The high level of interfactor association suggested that it was more appropriate to regard these four factors as facets of a single construct, and that it was most parsimonious to view the ASI as a single-factor measure. Cox, Parker, and Swinson (1996) evaluated both orthogonal and oblique rotations of several four-factor models to compare them with a unidimensional model of the ASI. Cox et al. (1996) found that the oblique solutions for multidimensional models, in particular the Peterson and Heilbronner (1987) model, provided better ®t indices than did the unidimensional model. In using oblique models, Cox et al. acknowledge that correlations among factors may compromise the view that the ASI is multidimensional. However, these authors felt that factor intercorrelations were suf®ciently low to support the position that the ASI was multidimensional. In sum, both Exploratory factor analyses (EFA) and CFA studies have produced discrepant ®ndings. In addition, CFA studies have generally indicated that even the best-®tting models provide only a marginally adequate ®t for the data. None of the models tested in the Taylor et al. (1992) study met criteria standards for adequacy and even the best-®tting model in the Cox et al. (1996) study (i.e., Peterson and Heilbronner oblique solution) produced ®t indices that barely met criteria standards. This suggests a need for evaluation of alternative structural models. One notable omission in the growing factor analytic studies of the ASI is evaluation of anxiety sensitivity in community samples. The majority of the studies evaluating nonclinical samples have utilized college populations, which are not necessarily representative of the general population. The only exception is Wardle et al. (1990), who used a relatively small sample (N 120) that is likely to produce an unreliable factor structure. In addition, the Wardle et al. control sample was not assessed for history of psychiatric illness or panic attacks. Because use of the ASI in nonclinical samples is likely to rise, as risk factor studies are conducted, it is imperative to understand the psychometrics and dimensional nature in a community sample. Moreover, determination of premorbid risk factors requires the exclusion of individuals with a history of
36
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
spontaneous panic to ensure that the structure of anxiety sensitivity is not merely a consequence or concomitant of the experience of panic. Anxiety sensitivity is likely to increase following the experience of spontaneous panic (Schmidt, Lerew, & Joiner, 2000) and because of this potential for change, there is reason to wonder whether the underlying dimensional nature may also be transformed. It is conceivable that different factor structures will emerge for clinical samples, for nonclinical samples with a history of spontaneous panic, and for those with no history of spontaneous panic, but this has yet to be evaluated. The principal aim of the present study was to evaluate the factor structure of the ASI in a community sample with no history of psychiatric illness or spontaneous panic. Both EFA and CFA were utilized. EFA provided an initial description of the factor structure of the instrument. CFA compared alternative factor models derived from the present study as well as the literature. In addition, the best®tting model was ®tted to a nonclinical college sample to assess its generalizability. 2. Method 2.1. Participants and procedure Two samples were utilized. The ®rst sample consisted of individuals (N 233) from the Washington, DC metropolitan community. These participants were solicited via newspaper advertisements requesting volunteers for various psychological and medical assessments. Interested participants were given an initial phone screen interview to rule out a history of spontaneous panic and psychiatric illness. Participants completing the phone screen were further screened with a face-to-face structured diagnostic interview (i.e., SCID-I/P; First, Spitzer, Gibbon, & Williams, 1994) conducted by graduate students in clinical psychology with extensive training in SCID administration and completed the ASI along with other measures and assessments and were paid US$40. The second sample consisted of undergraduate students (N 809) from a southwestern university. The ASI was group administered to students enrolled in introductory psychology classes who received research credit for their participation. Mean age of the student participants was 18.0 (S:D: 0:7) with a majority being Caucasian (80%). No data are available on the clinical history of these students. 2.2. Measure The ASI is a 16-item self-report instrument that assesses threatening beliefs regarding arousal symptoms (Peterson & Reiss, 1987). Each item is rated on a scale ranging from 0 (very little) to 4 (very much). Scale items are presented in Table 1.
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
37
Table 1 Corrected item±total correlations for ASI items ASI items
Item-to-scale correlations ASI-16
1. It is important to me not to appear nervous. .40 2. When I cannot keep my mind on a task, I worry that .45 I might be going crazy. 3. It scares me when I feel `shaky'' (trembling). .57 4. It scares me when I feel faint. .63 5. It is important to me to stay in control of my emotions. .28 6. It scares me when my heart beats rapidly. .59 7. It embarrasses me when my stomach growls. .29 8. It scares me when I am nauseous. .60 9. When I notice that my heart is beating rapidly, I worry that .43 I might have a heart attack. 10. It scares me when I become short of breath. .59 11. When my stomach is upset, I worry that I might be seriously ill. .53 12. It scares me when I am unable to keep my mind on a task. .63 13. Other people notice when I feel shaky. .49 14. Unusual body sensations scare me. .58 15. When I am nervous, I worry that I might be mentally ill. .43 16. It scares me when I am nervous. .61 M 10.7 S.D. 7.6 Coef®cient a .86
ASI-13 ASI-10 ± .48
± .50
.58 .63 ± .59 ± .62 .47
± ± ± .56 ± .61 .51
.63 .56 .65 .48 .58 .46 .65 5.7 6.3 .88
.58 .62 .64 .43 ± .53 .67 4.0 4.7 .85
Total N 233.
2.3. Analytic procedure Psychometric evaluation of the ASI was determined through internal consistency (i.e., coef®cient a) as well as corrected item±total correlations. Items whose corrected item±total correlations were less than .30 were considered questionable (Nunnally & Bernstein, 1994). A comprehensive factor analytic strategy was employed, as any single factor analytic technique can be questioned (Floyd & Widaman, 1995; Watson, Clark, & Harkness, 1994). EFAs were conducted to provide an initial description of the factor structure of the 16-item ASI. Both principal components analysis (PCA) and principal axis factoring (PAF) analyses were conducted using orthogonal and oblique rotations to enhance interpretability. PCA was used, consistent with Nunnally's (1978) recommendation that, in an exploratory analysis, PCA (with unities in the diagonals) is a reasonable analytic strategy. PAF provided a comparative strategy and because it may be better suited to describing underlying latent variables relative to PCA (Widaman, 1993). Factor estimation was based on the following criteria: (1) Kaiser's (1961) criterion to retain factors with unrotated eigenvalues of approximately one or greater; (2) the screen test (Cattell, 1966);
38
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
and (3) the interpretability of resulting factor structures (Gorsuch, 1983), which involves examining solutions with different extraction criteria to determine the point at which trivial or redundant factors emerge (e.g., Schmidt, Joiner, Young, & Telch, 1995). CFAs were based on maximum-likelihood analyses using EQS 5.4 (Bentler, 1995). Consistent with recommendation, a variety of goodness-of-®t measures for the overall model was evaluated (Marsh, Balla, & McDonald, 1988). The nonsigni®cant w2 statistic (assessing whether residual differences between the ®tted population covariance matrix and the observed covariance matrix converge to zero) is the most frequently used index. However, the w2 statistic is problematic as the likelihood of rejecting a model increases with sample size even when the model provides a very good ®t (Bentler & Bonett, 1980). Marsh et al. (1988) note that the w2 statistic will reject virtually all models with very large samples. Three stand-alone indicators (i.e., based on the results of the a priori model ®tting to data) were evaluated including the LISREL's Goodness-of-Fit Index (GFI: values >0.84), which provides a measure of the relative amount of variance and covariance accounted for by the model, the LISREL's Adjusted Goodness-ofFit Index (AGFI: values >0.80), which controls for the degrees of freedom in the model, and the root mean squared residual (RMSR: values <0.10), which is a measure of the difference between the observed and reconstructed correlations. Three incremental ®t indices (i.e., based on the difference between a target model and an alternative model) were also evaluated including the Bentler±Bonett Normed Fit Index (NFI: values >0.90), which represents one of the most widely used ®t indices, the Bentler±Bonett Nonnormed Fit Index (NNFI: values >0.90), which avoids the underestimation of ®t sometimes noted for the NFI in small samples (Bentler, 1990), and the Comparative Fit Index (CFI: values >0.90), which is adjusted for model parsimony and avoids the sampling variability of other incremental ®t indices. 3. Results 3.1. Item analysis of the ASI in a community sample The corrected item±total correlations for each of the 16 ASI items and coef®cient a are presented in Table 1. Coef®cient a for the ASI-16 was .86 but Items 5 and 7 produced questionable item±total correlations. Zero-order correlations among the individual ASI items are presented in Table 2. Particular attention was paid to item±item correlations below .20, because an item that is not at least moderately related with other items measuring the same construct will perform poorly in a factor analysis (Floyd & Widaman, 1995). As would be predicted from low item±total correlations, Items 5 and 7 produced both the fewest signi®cant correlations with other items and the fewest number of correlations greater than .20. In addition, Item 1 had seven correlations <.20.
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
39
Table 2 Intercorrelations among ASI items Item
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
± .19 .27 .30 .49 .22 .10 .19 .06 .18 .19 .28 .26 .22 .15 .24
± .32 .23 .06 .19 .14 .32 .26 .15 .41 .54 .34 .27 .56 .51
± .62 .19 .33 .16 .34 .17 .42 .27 .48 .40 .42 .20 .39
± .21 .50 .21 .51 .24 .55 .16 .40 .33 .54 .11 .36
± .25 .04 .18 .04 .18 .11 .15 .20 .20 .09 .16
± .21 .46 .43 .61 .31 .33 .27 .38 .16 .45
± .27 .31 .17 .19 .25 .19 .27 .11 .13
± .36 .57 .41 .43 .26 .34 .32 .45
± .50 .47 .34 .14 .31 .22 .29
± .35 .38 .17 .39 .14 .36
± .50 .38 .31 .53 .48
± .38 .43 .53 .54
± .41 .41 .42
± .21 .39
± .61
±
N 233.
In sum, preliminary analyses indicated that Items 1, 5, and 7 were weak and should likely be removed. 3.2. PCA and PAF of the ASI Findings from the PCA are summarized in Table 3. Both orthogonal and oblique rotations resulted in a similar pattern of factor loadings. PCA indicated four factors, although the order of the ®rst two factors and the last two factors was reversed depending on orthogonal or oblique rotation. The overall solution accounted for 63% of the variance. Factor 1 accounted for 36.5% of the variance (eigenvalue 5:83), Factor 2 accounted for 10.9% (eigenvalue 1:74), Factor 3 accounted for 9% (eigenvalue 1:44), and Factor 4 accounted for 6.6% (eigenvalue 1:0). Based on items with the greatest salient loadings, these factors could be described as Fear of Cardiopulmonary Sensations, Fear of Mental Catastrophe, Fear of Vasovagal Sensations, and Loss of Control Fears. Every item except Item 7 was loaded (>.39) on at least one factor and Items 11 and 13 were loaded on two factors with the orthogonal rotation. Not surprisingly, the oblique rotation resulted in higher cross-factor overlap with all items producing salient loadings on more than one factor. It is notable that the Loss of Control Fears factor was composed of two weak items (1 and 5) with a borderline eigenvalue (1.0) suggesting that this factor should be eliminated. Similarly, the screen test indicated that, of these four factors, two or possibly three should be retained. PAF analyses using both orthogonal and oblique rotations also yielded fourfactor solutions that were very similar to the corresponding PCA solutions. Examination of items with salient loadings suggests only two changes from
40
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
Table 3 Factor loadings for ASI items ASI items
PCA (orthogonal) I
1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16.
Nervous Going crazy ``Shaky'' Faint Control Heart beats Stomach Nauseous Heart attack Short of breath Stomach Mind Others notice Body sensations Mentally ill Nervous
.17 .76 .22 .02 .03 .08 .12 .28 .24 .02 .65 .64 .52 .22 .86 .68
II .03 .07 .13 .36 .08 .68 .38 .60 .79 .77 .45 .29 .01 .31 .08 .29
PCA (oblique) III
IV
I
.18 .18 .78 .80 .08 .30 .16 .29 .07 .34 .04 .36 .46 .62 .01 .26
.80 .00 .11 .15 .87 .25 .08 .12 .05 .13 .09 .08 .18 .08 .05 .13
.11 .22 .30 .49 .12 .73 .38 .68 .79 .81 .54 .44 .15 .43 .22 .43
II
III .24 .77 .35 .20 .09 .23 .19 .41 .33 .19 .71 .72 .58 .35 .86 .75
.83 .12 .28 .32 .87 .35 .00 .25 .03 .25 .18 .24 .31 .24 .15 .27
IV .28 .29 .81 .85 .18 .44 .23 .43 .09 .47 .13 .49 .53 .69 .12 .40
Factor labels: I (orthogonal) and II oblique Fear of Mental Catastrophe; II (orthogonal) and I oblique Cardiopulmonary Fears; III (orthogonal) and IV oblique Vasovagal Fears; IV (orthogonal) and III oblique loss of control fears.
the pattern of loadings described in Table 3. The orthogonal rotation produced a salient loading for Item 11 on the Cardiopulmonary Fears factor (.45) and the oblique rotation lost a salient loading for Item 3 on the Fear of Mental Catastrophe factor. In general, evaluation of eigenvalues and the screen plot suggests two main factors (Cardiopulmonary and Mental Catastrophe) along with two more questionable factors (Vasovagal and Loss of Control). A summary of the PCA and PAF analyses generally suggest two factors. Stability of the Loss of Control Fears and Vasovagal Fears factors is questionable. Loss of Control Fears is questionable because it is constituted by two weak items (i.e., based on interitem correlations and item±whole correlations) and Vasovagal Fears has high overlap with other factors in the oblique solution. In addition, the eigenvalues greater than 1 rule may overextract factors (Floyd & Widaman, 1995) and both factors possess borderline eigenvalues (1.0) depending upon orthogonal or oblique solution. Based on the information from the correlational and PCA/PAF analyses, it was determined that Items 1, 5, and 7 should be removed from the scale because of their low item±total correlations (Table 1), low interitem correlations (Table 2), and, in the case of Item 7, its failure to load on any factor. 3.3. Factor analyses on the revised ASI As per the suggestion of Floyd and Widaman (1995), items that performed poorly were deleted and a similar set of analyses was conducted with the 13-item
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
41
ASI (ASI-13). The corrected item±total correlations are reported in Table 1, along with the new alpha coef®cient. The internal consistency of the scale improved after the three questionable items were eliminated (a :88). PCA and PAF analyses of the ASI-13 using both orthogonal and oblique rotations were similar in terms of yielding three-factor solutions with similar salient loadings (>.40). PCA and PAF analyses with orthogonal rotation resulted in Items 2, 11, 12, 13, 15, and 16 loading onto a Fear of Mental Catastrophe factor (eigenvalue 5:50, accounting for 42.3% of the variance), Items 6, 8, 9, 10, and 11 loading onto a Cardiopulmonary Fears factor (eigenvalue 1:71, accounting for 13.2% of the variance) and Items 3, 4, 13, and 14 loading onto a Vasovagal Fears factor (eigenvalue 1:20, accounting for 9.2% of the variance). Oblique rotations indicated substantial factor overlap with moderate factor intercorrelations (Factors I±II: r :32; Factors I±III: r :15; Factors I±IV: r :34; Factors II±III: r :19; Factors II±IV: r :29; Factors III±IV: r :29). The PCA indicated that 8 of 13 items, and the PAF indicated that 10 of 13 items, loaded on more than one factor (>.40) with the oblique rotation. In general, analyses using orthogonal rotations were consistent with three-factor solutions as all three factors were interpretable and accounted for suf®cient variance. Oblique solutions were more consistent with analyses using the ASI-16 as they were suggestive of two-factor solutions due to dif®culty in interpreting the Vasovagal Fears factor due to the high level of item complexity (i.e., items loading on multiple factors). 3.4. Con®rmatory factor analyses 3.4.1. Unifactorial models Four separate unifactorial models were constructed based on the original fullscale ASI (ASI-16), the ASI with Items 7 and 13 was removed as suggested by Taylor et al.'s (1992) con®rmatory analyses (ASI-14), the ASI with the ®ve items (i.e., 1, 5, 7, 8, 13) was removed as suggested by Blais et al.'s (in press) analyses (ASI-11), and the ASI with the three items (i.e., 1, 5, 7) was removed as suggested by the analyses reported above (ASI-13). Goodness-of-®t indices for all models are presented in Table 4. It is notable that all latent variable models yielded signi®cant w2 statistics, which is likely to result from the large sample size (Marsh et al., 1988). The overall ®t of each unifactorial model was comparable across most ®t measures (e.g., CFI range: .69±.71) suggesting that all models accounted for the data about equally well. Each model, however, produced a poor ®t according to all criteria except the RMSR suggesting that a unifactorial model does not adequately account for the data. 3.4.2. Multifactorial models Four multifactorial models were tested with both orthogonal and oblique solutions (see Table 4). The Peterson and Heilbronner (1987) and Telch, Shermis, and Lucas (1989) four-factor models were selected based on their superior
42
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
Table 4 Goodness-of-®t indices for latent variable models Model
Fit indices Stand alone w
2
NFI
NNFI CFI
SB
RCFI
.08a .09a .05a .06a
.65 .66 .68 .67
.65 .63 .65 .61
.69 .69 .71 .69
262.1 219.0 164.9 126.1
.68 .68 .71 .70
.70 .80a
.17 .04a
.69 .82
.68 .83
.73 .87
275.3 .64 155.6 .86
Telch, Shermis, and Lucas (1989) Orthogonal 511.9 101 .78 Oblique 397.2 96 .82
.71 .75
.15 .07a
.69 .76
.68 .75
.73 .80
274.3 .65 182.2 .82
Blais et al. (in press) Orthogonal 417.9 Oblique 329.7
.65 .65
.11 .05a
.64 .72
.58 .67
.67 .74
139.0 .65 109.1 .76
Schmidt and Joiner (three-factor) Orthogonal 428.5 65 .80 Oblique 240.1 62 .87a
.72 .81a
.16 .04a
.70 .83
.68 .83
.73 .87
156.7 .72 120.4 .83
Schmidt and Joiner (two-factor) Orthogonal 185.2 35 Oblique 121.8 34 102.3 26 Oblique (8)b Oblique (13)c 113.5 26 Oblique (8,13)d 94.4 19
.82a .86a .85a .84a .83a
.11 .03a .03a .03a .03a
.81 .88 .88 .88 .88
.79 .88 .87 .86 .85
.84 .91a .91a .90a .90a
Unifactorial ASI-16 ASI-14 ASI-13 ASI-11
df
Incremental
570.1 104 504.5 77 458.9 65 387.0 44
GFI
AGFI RMSR
.73 .72 .72 .72
.64 .62 .61 .58
Multifactorial Peterson and Heilbronner (1987) Orthogonal 506.4 100 .78 Oblique 292.6 94 .86a
44 43
.77 .77
.88a .91a .91a .91a .91a
96.5 62.2 52.1 57.5 48.4
.65 .86 .82 .82 .76
GFI: LISREL's Goodness-of-Fit Index; AGFI: LISREL's Adjusted Goodness-of-Fit Index; RMSR: Root Mean Squared Residual; NFI: Bentler±Bonet Normed Fit Index; NNFI: Benter±Bonet Nonnormed Fit Index; CFI: Comparative Fit Index; SB: Satorra±Bentler rescaled w2 statistic; RCFI: Robust Comparative Fit Index. a Indicates meets threshold for adequate ®t (GFI > :84; AGFI > :79; RMSR < :10; NFI, NNFI, CFI > :89). b Item 8 removed from model. c Item 13 removed from model. d Items 8 and 13 removed from model.
performance in previous CFA studies. The Blais et al. (in press) two-factor model, which has not been evaluated using CFA, and the three- and two-factor models obtained from the EFA were also examined. The Peterson and Heilbronner (1987) orthogonal model did not yield an adequate ®t. Criteria standards were not met for the w2 statistic and other ®t indices (w2 506:4, df 100; CFI :73). However, the oblique model produced a substantially better ®t according to all ®t indices. Despite a signi®cant w2 statistic and
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
43
a CFI that did not meet criteria standards (w2 292:6, df 94; CFI :87), the GFI, AGFI, and RMSR indicated adequate ®t (GFI :86, AGFI :80, RMSR :04). The Telch, Shermis, and Lucas (1989) orthogonal and oblique models were not found to adequately ®t the data. The orthogonal model produced no ®t indices meeting criteria standards (w2 511:9, df 101; CFI :73). The oblique model produced a relatively better ®t and did yield an adequate RMSR (.07) but no other ®t indices met criteria standards (w2 397:2, df 96; CFI :80). The Blais et al. (in press) orthogonal and oblique models also were not found to adequately ®t the data. The orthogonal model produced no ®t indices meeting criteria standards (w2 417:9, df 44; CFI :67). The oblique model produced a relatively better ®t and did yield an adequate RMSR (.05) but no other ®t indices met criteria standards (w2 329:7, df 43; CFI :74). Both the three- and two-factor models derived from preliminary PCA and PAF analyses were evaluated. The three-factor orthogonal model did not yield an adequate ®t (w2 428:5, df 65; CFI :73). However, the oblique three-factor model produced a pattern of ®t similar to the Peterson and Heilbronner four-factor oblique solution. Despite a signi®cant w2 statistic and a CFI that did not meet criteria standards (w2 240:1, df 62; CFI :87), the GFI, AGFI, and RMSR indicated adequate ®t (GFI :87, AGFI :81, RMSR :04). The two-factor model produced ®t indices with markedly better ®t relative to all other orthogonal solutions. The two-factor orthogonal model met criteria standards for the GFI (.88) and AGFI (.82) despite a signi®cant w2 and inadequate CFI (w2 185:2, df 35; CFI :84). The two-factor oblique model produced the best overall ®t of all models tested (see Fig. 1). This model met criteria standards for the GFI (.91), AGFI (.86), RMSR (.03), and CFI (.91) despite a signi®cant w2 statistic (w2 121:8, df 34).
Fig. 1. Two-factor oblique model of the ASI in a community sample with standardized path coefficients.
44
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
Three additional analyses were conducted in an attempt to further re®ne the two-factor oblique model. The model was evaluated after the removal of Item 8 (i.e., nausea) from the Cardiopulmonary Fears factor, removal of Item 13 (i.e., others notice when I'm shaky) from the Fear of Mental Catastrophe factor, and after removal of both items. The rationale for these tests was based on the face validity of the items in factors appearing to tap a somewhat different area (e.g., gastrointestinal vs. cardiac and pulmonary items) as well as the fact that Items 8 and 13 had the weakest loadings of all salient items in the PAF and PCA. These revised models yielded highly comparable ®t indices relative to the original twofactor oblique solution suggesting that nothing is gained from removal of these items (see Table 4). The Lagrange Multiplier (LM) and Wald (W) tests provide tests of the statistical necessity of sets of parameters that may be added or deleted from a model (Bentler, 1995). The W test indicated that no paths should be deleted and the multivariate LM test indicated adding several parameters. However, inclusion of any of these parameters created uninterpretable path coef®cients (standardized values >1.0). In sum, a re®ned model was not generated from substantive hypotheses or respeci®cation based on LM or W tests suggesting that it is best to retain the twofactor 10-item (ASI-10) scale. Table 1 shows item±whole correlations for the twofactor 10-item scale. Reanalysis of internal consistency reliability for the ASI-10 indicated adequate internal consistency (overall coef®cient a :86; Factor I coef®cient a :82; Factor II coef®cient a :80). The ASI-10 was highly correlated with the overall ASI-16 (r :92). 3.4.3. Generalizability of latent variable models to a nonclinical college student sample Generalizability of the derived 10-item (ASI-10) two-factor model was assessed by comparing it with the other oblique models (i.e., Peterson & Heilbronner, Telch et al., Blais et al.) in a large sample (N 809) of college undergraduates. The twofactor oblique model derived from the community sample provided an adequate ®t to the data as indicated by several ®t indices (GFI :94, AGFI :91, RMSR :05, NFI :89, NNFI :87, CFI :90). This level of ®t is similar to that obtained for the community sample. The two-factor model provides a somewhat better ®t relative to all other models including the Peterson and Heilbronner four-factor model (GFI :92, AGFI :89, RMSR :05, NFI :86, NNFI :85, CFI : 88), the Telch et al. four-factor model (GFI :92, AGFI :89, RMSR :05, NFI :87, NNFI :87, CFI :89), and the Blais et al. two-factor model (GFI :91, AGFI :86, RMSR :06, NFI :85, NNFI :82, CFI :86). 3.4.4. Additional statistical considerations Multivariate normality is a general requirement for maximum-likelihood CFA that assumes that each variable possesses zero skewness and zero kurtosis (Cole, 1987). Unfortunately, the majority of data collected in behavioral research do not follow univariate or multivariate normal distributions (Micceri, 1989). Two
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
45
different approaches have been suggested to address multivariate nonnormality (Curran, West, & Finch, 1996). One approach is to use methods of estimation (i.e., asymptotic distribution free) that do not assume multivariate normality such that variables possessing nonzero kurtoses pose no special problems. A second approach adjusts the maximum-likelihood w2 estimate for the presence of nonzero kurtosis (i.e., Satorra±Bentler (SB) rescaled w2). Evaluation of multivariate and univariate kurtosis and skewness suggests signi®cant multivariate kurtosis (Mardia's coefficient 240:2) as well as signi®cant kurtosis for Items 2, 9, 11, 15, and 16. Both approaches were evaluated for all CFA models. All models were reanalyzed using EQS's AGLS estimation (i.e., the equivalent of asymptotic distribution free). The overall pattern of goodness-of-®t across all models remained relatively stable with AGLS remodeling, although there tended to be substantial increases in the LISREL indices (e.g., two-factor oblique model: GFI :99, AGFI :98, RMSR :03, NFI :87, NNFI :87, CFI :90; Peterson & Heilbronner oblique model: GFI :99, AGFI :98, RMSR :04, NFI :82, NNFI :83, CFI :86).The adjusted SB w2 statistic as well as a corresponding CFI (i.e., the Robust CFI) are provided in Table 4. The SB statistic yielded lower w2 values but all continued to be statistically signi®cant. The Robust CFI was generally comparable to the CFI, although some orthogonal models produced substantially poorer ®ts according to this index. 4. Discussion Because of the ambiguity that can arise in the interpretation of both EFA and CFA (Floyd & Widaman, 1995), a comprehensive analytic strategy was employed to gain convergence as to the best description of the structure of anxiety sensitivity in a community sample. Utility of this methodology was evidenced in that exploratory analyses suggested either a two- or three-factor solution, whereas con®rmatory analyses clearly indicated the superiority of the two-factor solution described by Fears of Mental Catastrophe and Fears of Cardiopulmonary Sensations. Factor intercorrelations, as well as the superior ®t of an oblique solution, indicate a hierarchical model described by a superordinate anxiety sensitivity factor with two interrelated ®rst-order factors. Anxiety sensitivity researchers have struggled to obtain a consensus in terms of the ASI's factor structure but ®ndings from factor analytic studies, including the present report, suggest common themes. The factors extracted in both the exploratory and con®rmatory analyses were descriptively similar to those in previous studies. In particular, most studies have extracted two separate factors specifying fear of sensations linked with physical catastrophe and fears of sensations linked with mental catastrophe (Blais et al., in press; Peterson & Heilbronner, 1987; Telch, Shermis, & Lucas, 1989; Wardle et al., 1990). For example, the derived two-factor solution corresponds well in terms of items with the highest salient loadings, with the two-factor model (i.e., fear of somatic
46
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
sensations, fear of loss of mental control) obtained by Blais et al. (in press). Findings across these studies indicate that the anxiety sensitivity construct may be described by two interrelated factors that generally tap physical and mental domains. It was hypothesized that beliefs about autonomic arousal may change in the context of relevant life experiences (e.g., having a spontaneous panic attack) that would create differences in the structure of anxiety sensitivity across clinical and nonclinical samples. In fact, the structure of the ASI was highly similar for individuals with no history of spontaneous panic compared to a nonclinical college sample, which undoubtedly included a signi®cant percentage of individuals with a history of spontaneous panic. Moreover, this two-factor model was similar to Blais et al.'s (in press) two-factor solution in a largely psychiatric sample. The major distinction between the community sample and Blais et al.'s mixed sample is that the salience of particular anxiety sensitivity dimensions, as indexed by the primary factor obtained in each study, appears to shift from a focus on fears of mental catastrophe to fears of physical catastrophe in the clinical sample. Thus, the underlying factor structure of anxiety sensitivity, as measured by the ASI, appears to be relatively stable across nonclinical and clinical samples even as the absolute level of anxiety sensitivity increases. Presence of two ®rst-order factors indicates two potential risk dimensions that may contribute to the development of anxiety pathology. Individuals with mental fears may be more apt to experience panic in the context of unexplained or unexpected signs of mental catastrophe (e.g., derealization) whereas those with physical fears may panic when bodily cues (e.g., heart palpitations) mimic physical concerns. Ultimately, these subgroups may produce different anxiety disorders or discernable subtypes of one type of anxiety pathology. For example, the expression of anxiety sensitivity differs among patients with panic disorder (Schmidt, 1999). Although patients with panic disorder may be generally fearful of strong bodily perturbations, patients may often evidence particular fears of one or two speci®c arousal cues. For example, panic disordered patients' beliefs about the consequences of panic-related arousal tend to cluster in three domains: physical, social, and loss of control (Telch, Brouillard, Telch, Agras, & Taylor, 1989). Dimensionality of the ASI is one of the controversial topics within the anxiety sensitivity literature (Taylor et al., 1992). A multidimensional ASI is potentially problematic for the unidimensional construct of anxiety sensitivity. However, Lilienfeld, Turner, and Jacob's (1993) proposition that anxiety sensitivity may be hierarchical allows for both superfactor and multidimensional models (see also Watson et al., 1994). This view is gaining acceptance (Cox et al., 1996; Lilienfeld, Turner, & Jacob, 1996; Zinbarg, Barlow, & Brown, 1997) and is consistent with the present ®ndings. Another important aspect of the dimensionality debate has to do with the degree to which multifactor solutions are intercorrelated. For example, Taylor et al. (1992) argued that four-factor solutions (e.g., Telch, Shermis, & Lucas, 1989) showed high levels of factor-intercorrelation and should therefore be
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
47
best viewed as unidimensional. On the other hand, Cox et al. (1996) note that none of the unidimensional models provides adequate ®t for the data. Indeed, all of the unidimensional models tested in the present study also provided poor ®t for the data. On the other hand, oblique multidimensional models ®t the data much better. Cox et al. comment that it is not uncommon for multidimensional measures of personality traits to be highly intercorrelated. Similarly, subscale indices on intellectual tests also possess high levels of intercorrelation (Carroll, 1993; Horn & Noll, 1997). Analogous to the saturation of intelligence tests with the higher order factor, g, the ®rst-order ASI factors can be considered to be saturated by the higher order dimension of anxiety sensitivity. In sum, anxiety sensitivity appears to be unidimensional at a higher order, but multidimensional at a lower order. Importantly, it is worth reiterating that most ASI models thus far evaluated in the literature provide poor ®t to the data. The best-®tting model in the present study provided an adequate ®t but it is likely that this could be improved. Because of these issues, and in concurrence with recommendations of Cox et al. (1996), it is suggested that the ASI item pool be expanded in order to construct more reliable subscales. Until subscales are expanded, caution should be used in interpreting the subscale scores from the derived factors because of instability that may result from scales with only four or six items. Given the data implicating anxiety sensitivity as a risk factor for anxiety pathology, evaluation of anxiety sensitivity before onset of panic attacks and formal anxiety disorder syndromes is critical. Interestingly, the overall level of anxiety sensitivity in the community sample was substantially lower than the college sample and other nonclinical samples reported in the literature (Telch, Shermis, & Lucas, 1989). This difference is likely to be due to exclusionary criteria (i.e., history of spontaneous panic and psychiatric disorders) for the community sample, which will be present in undergraduate populations. Further evaluation is needed, however, as this disparity may be due to other inherent differences among community versus college nonclinical samples. Acknowledgments The authors thank Margaret Koselka, Jack Trakowski, and Matt Wineman for their assistance with structured diagnostic interviews, and Manuel Orellana for assistance with data management. References Bentler, P. M. (1990). Comparative ®t indices in structural models. Psychological Bulletin, 107, 238±246. Bentler, P. M. (1995). EQS structural equations program manual. Encino, CA: Multivariate Software. Bentler, P. M., & Bonett, D. G. (1980). Signi®cance tests and goodness of ®t in the analysis of covariance structures. Psychological Bulletin, 88, 588±606.
48
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
Blais, M. A., Otto, M. W., Zucker, B. G., McNally, R. J., Schmidt, N. B., Fava, M., & Pollack, M. H. (in press). The Anxiety Sensitivity Index: I. Item analysis and suggestions for re®nement. Journal of Personality Assessment. Carroll, J. B. (1993). Human cognitive abilities: a survey of factor-analytic studies. New York: Cambridge University Press. Cattell, R. B. (1966). The screen test for the number of factors. Multivariate Behavioral Research, 1, 245±276. Cole, D. A. (1987). Utility of con®rmatory factor analysis in test validation research. Journal of Consulting and Clinical Psychology, 55, 584±594. Cox, B. J., Parker, J. D. A., & Swinson, R. P. (1996). Anxiety sensitivity: con®rmatory evidence for a multidimensional construct. Behaviour Research and Therapy, 34, 591±598. Curran, P. J., West, S. G., & Finch, J. F. (1996). The robustness of test statistics to nonnormality and speci®cation error in con®rmatory factor analysis. Psychological Methods, 1, 16±29. Donnell, C. D., & McNally, R. J. (1990). Anxiety sensitivity and panic attacks in a nonclinical population. Behaviour Research and Therapy, 28, 83±85. First, M. B., Spitzer, R. L., Gibbon, M., & Williams, J. B. (1994). Structured clinical interview for DSM-IV Patient Edition. New York: Biometrics Research Department (SCID-N/P Version 2.0). Floyd, F. J., & Widaman, K. F. (1995). Factor analysis in the development and re®nement of clinical assessment instruments. Psychological Assessment, 7, 286±299. Gorsuch, R. (1983). Factor analysis (2nd ed.). Hillsdale, NJ: Erlbaum. Harrington, P. H., Schmidt, N. B., & Telch, M. J. (1996). Prospective evaluation of panic potentiation following 35% CO2 challenge in a nonclinical sample. American Journal of Psychiatry, 153, 823±825. Horn, J. L., & Noll, J. (1997). Human cognitive capabilities: Gf±Gc theory. In: D. P. Flanagan, J. L. Harrison, & P. L. Harrison (Eds.), Contemporary intellectual assessment: theories, tests, and issues (pp. 52±91). New York: Guilford. Kaiser, H. F. (1961). A note on Guttman's lower bound for the number of common factors. Multivariate Behavioral Research, 1, 249±276. Lilienfeld, S. O., Turner, S. M., & Jacob, R. G. (1993). Anxiety sensitivity: an examination of theoretical and methodological issues. Advances in Behaviour Research and Therapy, 15, 147±182. Lilienfeld, S. O., Turner, S. M., & Jacob, R. G. (1996). Further comments on the nature and measurement of anxiety sensitivity: a reply to Taylor (1995b). Journal of Anxiety Disorders, 10, 411±424. Maller, R. G., & Reiss, S. (1992). Anxiety sensitivity in 1984 and panic attacks in 1987. Journal of Anxiety Disorders, 6, 241±247. Marsh, H. W., Balla, J. R., & McDonald, R. P. (1988). Goodness-of-®t indices in con®rmatory factor analyses: the effect of sample size. Psychological Bulletin, 103, 391±410. McNally, R. J. (1990). Psychological approaches to panic disorder: a review. Psychological Bulletin, 108, 403±419. Micceri, T. (1989). The unicorn, the normal curve, and other improbable creatures. Psychological Bulletin, 105, 156±166. Nunnally, J. (1978). Psychometric theory. New York: McGraw-Hill. Nunnally, J. C., & Bernstein, I. H. (1994). Psychometric theory (3rd ed.). New York: McGraw-Hill. Peterson, R. A., & Heilbronner, R. L. (1987). The Anxiety Sensitivity Index: construct validity and factor analytic structure. Journal of Anxiety Disorders, 1, 117±121. Peterson, R. A., & Reiss, S. (1987). Test manual for the Anxiety Sensitivity Index. Orland Park, IL: International Diagnostic Systems. Rapee, R. M., & Medoro, L. (1994). Fear of physical sensations and trait anxiety as mediators of the response to hyperventilation in nonclinical subjects. Journal of Abnormal Psychology, 103, 693±699. Reiss, S. (1991). Expectancy model of fear, anxiety, and panic. Clinical Psychology Review, 11, 141± 153.
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
49
Reiss, S., & McNally, R. J. (1985). The expectancy model of fear. In: S. Reiss & R. R. Bootzin (Eds.), Theoretical issues in behavior therapy (pp. 107±121). New York: Academic Press. Reiss, S., Peterson, R. A., Gursky, D. M., & McNally, R. J. (1986). Anxiety sensitivity, anxiety frequency and the prediction of fearfulness. Behaviour Research and Therapy, 24, 1±8. Schmidt, N. B. (1999). Examination of differential anxiety sensitivities in panic disorder: a test of anxiety sensitivity subdomains in predicting fearful responding to a 35% CO2 challenge. Cognitive Therapy and Research, 23, 3±19. Schmidt, N. B., Joiner, T.E., Jr., Young, J., & Telch, M. J. (1995). The schema questionnaire: investigation of psychometric properties and the hierarchical structure of a measure of maladaptive schemas. Cognitive Therapy and Research, 19, 295±321. Schmidt, N. B., Lerew, D., & Jackson, R. J. (1997). The role of anxiety sensitivity in the pathogenesis of panic: prospective evaluations of spontaneous panic attacks during acute stress. Journal of Abnormal Psychology, 106, 355±364. 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±1109. Schmidt, N. B., & Telch, M. J. (1994). The role of safety information and fear of bodily sensations in moderating responses to a hyperventilation challenge. Behavior Therapy, 25, 197±208. Taylor, S., Koch, W. J., & Crockett, D. J. (1991). Anxiety sensitivity, trait anxiety, and the anxiety disorders. Journal of Anxiety Disorders, 6, 293±311. Taylor, S., Koch, W. J., McNally, R. J., & Crockett, D. J. (1992). Conceptualizations of anxiety sensitivity. Psychological Assessment, 4, 245±250. Telch, M. J., Brouillard, M., Telch, C. F., Agras, W. S., & Taylor, C. B. (1989). Role of cognitive appraisal in panic-related avoidance. Behavior Research and Therapy, 27, 373±383. Telch, M. J., Shermis, M. D., & Lucas, J. A. (1989). Anxiety sensitivity: unitary personality trait or domain-speci®c appraisals? Journal of Anxiety Disorders, 3, 25±32. Telch, M. J., Silverman, A., & Schmidt, N. B. (1996). The relationship between anxiety sensitivity and perceived control in a caffeine challenge. Journal of Anxiety Disorders, 10, 21±35. Wardle, J., Ahmad, T., & Hayward, P. (1990). Anxiety sensitivity in agoraphobia. Journal of Anxiety Disorders, 4, 325±333. Watson, D., Clark, L. A., & Harkness, A. R. (1994). Structures of personality and their relevance to psychopathology. Journal of Abnormal Psychology, 103, 18±31. Widaman, K. F. (1993). Common factor-analysis versus principal component analysisÐdifferential bias in representing model parameters. Multivariate Behavioral Research, 28, 263±311. 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±281.
Structure of the Anxiety Sensitivity Index psychometrics and factor structure in a community sample Norman B. Schmidta,*, Thomas E. Joinerb a
Department of Psychology, The Ohio State University, 245 Townshend Hall, Columbus, OH 43210, USA b Department of Psychology, Florida State University, Tallahassee, FL, USA
Received 13 January 2000; received in revised form 14 June 2000; accepted 30 June 2000
Abstract Converging evidence suggests that anxiety sensitivity (i.e., threatening beliefs regarding autonomic arousal) is a risk factor for anxiety pathology. Speci®cation of premorbid risk factors requires exclusion of individuals with a history of spontaneous panic to ensure that anxiety sensitivity is not merely a consequence or concomitant of the experience of panic. However, the psychometrics and dimensional nature of anxiety sensitivity in such a sample is undetermined. The present study evaluated the factor structure of the Anxiety Sensitivity Index (ASI), a measure of anxiety sensitivity, in a community sample (N 233) with no history of psychiatric illness or spontaneous panic. Exploratory factor analyses (EFA) suggested a two- or three-factor solution (I, Fear of Mental Catastrophe; II, Fear of Cardiopulmonary Sensations; III, Fear of Vasovagal Sensations). Con®rmatory factor analyses (CFA) comparing alternative models indicated that a hierarchical two-factor solution (I, Fear of Mental Catastrophe; II, Fear of Cardiopulmonary Sensations) best accounted for the data. This model generalized well to a nonclinical college sample (N 809). # 2002 Elsevier Science Inc. All rights reserved. Keywords: Anxiety sensitivity; Factor analysis; Anxiety; Community sample
*
Corresponding author. Tel.: 1-614-292-2687. E-mail address: [email protected] (N.B. Schmidt).
0887-6185/02/$ ± see front matter # 2002 Elsevier Science Inc. All rights reserved. PII: S 0 8 8 7 - 6 1 8 5 ( 0 1 ) 0 0 0 8 7 - 1
34
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
1. Introduction Anxiety sensitivity has played an important role in the conceptualization and research on anxiety pathology (McNally, 1990). Anxiety sensitivity refers to the extent to which an individual believes that autonomic arousal can have aversive or harmful consequences (Reiss & McNally, 1985). For example, individuals with high anxiety sensitivity may believe that shortness of breath signals suffocation or that heart palpitations indicate a heart attack, whereas those with low anxiety sensitivity experience these sensations as unpleasant but nonthreatening. Consistent with cognitive theories of anxiety, the anxiety sensitivity conceptualization posits that cognitive misappraisal is critical for generation of anxiety. However, anxiety sensitivity is distinguished from other cognitive conceptualizations because anxiety sensitivity is believed to be a stable trait-like characteristic that may precede development of panic attacks. Individual differences in anxiety sensitivity are hypothesized to emerge from a variety of experiences that ultimately lead to acquisition of beliefs about potentially aversive consequences of arousal. Such experiences may include hearing others express fear of such sensations, receiving misinformation about the harmfulness of certain sensations, witnessing a catastrophic event such as the fatal heart attack of a loved one, and so forth. Thus, anxiety sensitivity constitutes a disposition to developing anxiety and does not require the experience of clinically signi®cant anxiety or panic in its own development (Reiss, 1991). One of the important and unique predictions of the anxiety sensitivity conceptualization is that anxiety sensitivity should act as a risk factor for development of panic attacks as well as related anxiety disorders. Both laboratory-based experimental studies as well as prospective quasi-experimental studies have supported this prediction. Experimental studies using nonclinical samples with no history of panic attacks have indicated that anxiety sensitivity is predictive of fearful responding in the context of hyperventilation, caffeine, and carbon dioxide inhalation (Donnell & McNally, 1990; Harrington, Schmidt, & Telch, 1996; Rapee & Medoro, 1994; Schmidt & Telch, 1994; Telch, Silverman, & Schmidt, 1996). Prospective evaluation of nonclinical samples has determined that anxiety sensitivity is a risk factor for development of spontaneous panic attacks (Schmidt, Lerew, & Jackson, 1997) as well as anxiety disorders (Maller & Reiss, 1992). Taken together, laboratory and prospective studies provide converging evidence for anxiety sensitivity as a risk factor in the development of anxiety pathology. Anxiety sensitivity was originally proposed as a unitary construct (Reiss & McNally, 1985). Empirical evaluation of this proposition using the most widely used measure of anxiety sensitivity, the Anxiety Sensitivity Index (ASI), has yielded inconsistent ®ndings. Several early factor analytic studies provided support for the unitary nature of anxiety sensitivity (Reiss, Peterson, Gursky, & McNally, 1986; Taylor, Koch, & Crockett, 1991). Others found that the ASI is multifactorial (Telch, Shermis, & Lucas, 1989; Wardle, Ahmad, & Hayward,
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
35
1990). Recently, Blais et al. (in press) determined that the ASI is best described by a two-factor solution consisting of fears of physical sensations related to anxiety and Fear of Mental Catastrophe. Two studies using con®rmatory factor analytic (CFA) techniques have attempted to clarify the factor structure of the ASI. CFA techniques provide a method for evaluating competing factor solutions by examining goodness-of-®t indices for various models. Taylor, Koch, McNally, and Crockett (1992) concluded that the ASI was best viewed as a single-factor measure even though an orthogonal four-factor model (i.e., Telch, Shermis, & Lucas, 1989) resulted in a similar goodness-of-®t. Taylor et al. argued that the four-factor solutions obtained by Telch, Shermis, and Lucas (1989). and Wardle et al. (1990) were artifacts of the orthogonality constraint because signi®cant interfactor correlations resulted when this constraint was removed. The high level of interfactor association suggested that it was more appropriate to regard these four factors as facets of a single construct, and that it was most parsimonious to view the ASI as a single-factor measure. Cox, Parker, and Swinson (1996) evaluated both orthogonal and oblique rotations of several four-factor models to compare them with a unidimensional model of the ASI. Cox et al. (1996) found that the oblique solutions for multidimensional models, in particular the Peterson and Heilbronner (1987) model, provided better ®t indices than did the unidimensional model. In using oblique models, Cox et al. acknowledge that correlations among factors may compromise the view that the ASI is multidimensional. However, these authors felt that factor intercorrelations were suf®ciently low to support the position that the ASI was multidimensional. In sum, both Exploratory factor analyses (EFA) and CFA studies have produced discrepant ®ndings. In addition, CFA studies have generally indicated that even the best-®tting models provide only a marginally adequate ®t for the data. None of the models tested in the Taylor et al. (1992) study met criteria standards for adequacy and even the best-®tting model in the Cox et al. (1996) study (i.e., Peterson and Heilbronner oblique solution) produced ®t indices that barely met criteria standards. This suggests a need for evaluation of alternative structural models. One notable omission in the growing factor analytic studies of the ASI is evaluation of anxiety sensitivity in community samples. The majority of the studies evaluating nonclinical samples have utilized college populations, which are not necessarily representative of the general population. The only exception is Wardle et al. (1990), who used a relatively small sample (N 120) that is likely to produce an unreliable factor structure. In addition, the Wardle et al. control sample was not assessed for history of psychiatric illness or panic attacks. Because use of the ASI in nonclinical samples is likely to rise, as risk factor studies are conducted, it is imperative to understand the psychometrics and dimensional nature in a community sample. Moreover, determination of premorbid risk factors requires the exclusion of individuals with a history of
36
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
spontaneous panic to ensure that the structure of anxiety sensitivity is not merely a consequence or concomitant of the experience of panic. Anxiety sensitivity is likely to increase following the experience of spontaneous panic (Schmidt, Lerew, & Joiner, 2000) and because of this potential for change, there is reason to wonder whether the underlying dimensional nature may also be transformed. It is conceivable that different factor structures will emerge for clinical samples, for nonclinical samples with a history of spontaneous panic, and for those with no history of spontaneous panic, but this has yet to be evaluated. The principal aim of the present study was to evaluate the factor structure of the ASI in a community sample with no history of psychiatric illness or spontaneous panic. Both EFA and CFA were utilized. EFA provided an initial description of the factor structure of the instrument. CFA compared alternative factor models derived from the present study as well as the literature. In addition, the best®tting model was ®tted to a nonclinical college sample to assess its generalizability. 2. Method 2.1. Participants and procedure Two samples were utilized. The ®rst sample consisted of individuals (N 233) from the Washington, DC metropolitan community. These participants were solicited via newspaper advertisements requesting volunteers for various psychological and medical assessments. Interested participants were given an initial phone screen interview to rule out a history of spontaneous panic and psychiatric illness. Participants completing the phone screen were further screened with a face-to-face structured diagnostic interview (i.e., SCID-I/P; First, Spitzer, Gibbon, & Williams, 1994) conducted by graduate students in clinical psychology with extensive training in SCID administration and completed the ASI along with other measures and assessments and were paid US$40. The second sample consisted of undergraduate students (N 809) from a southwestern university. The ASI was group administered to students enrolled in introductory psychology classes who received research credit for their participation. Mean age of the student participants was 18.0 (S:D: 0:7) with a majority being Caucasian (80%). No data are available on the clinical history of these students. 2.2. Measure The ASI is a 16-item self-report instrument that assesses threatening beliefs regarding arousal symptoms (Peterson & Reiss, 1987). Each item is rated on a scale ranging from 0 (very little) to 4 (very much). Scale items are presented in Table 1.
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
37
Table 1 Corrected item±total correlations for ASI items ASI items
Item-to-scale correlations ASI-16
1. It is important to me not to appear nervous. .40 2. When I cannot keep my mind on a task, I worry that .45 I might be going crazy. 3. It scares me when I feel `shaky'' (trembling). .57 4. It scares me when I feel faint. .63 5. It is important to me to stay in control of my emotions. .28 6. It scares me when my heart beats rapidly. .59 7. It embarrasses me when my stomach growls. .29 8. It scares me when I am nauseous. .60 9. When I notice that my heart is beating rapidly, I worry that .43 I might have a heart attack. 10. It scares me when I become short of breath. .59 11. When my stomach is upset, I worry that I might be seriously ill. .53 12. It scares me when I am unable to keep my mind on a task. .63 13. Other people notice when I feel shaky. .49 14. Unusual body sensations scare me. .58 15. When I am nervous, I worry that I might be mentally ill. .43 16. It scares me when I am nervous. .61 M 10.7 S.D. 7.6 Coef®cient a .86
ASI-13 ASI-10 ± .48
± .50
.58 .63 ± .59 ± .62 .47
± ± ± .56 ± .61 .51
.63 .56 .65 .48 .58 .46 .65 5.7 6.3 .88
.58 .62 .64 .43 ± .53 .67 4.0 4.7 .85
Total N 233.
2.3. Analytic procedure Psychometric evaluation of the ASI was determined through internal consistency (i.e., coef®cient a) as well as corrected item±total correlations. Items whose corrected item±total correlations were less than .30 were considered questionable (Nunnally & Bernstein, 1994). A comprehensive factor analytic strategy was employed, as any single factor analytic technique can be questioned (Floyd & Widaman, 1995; Watson, Clark, & Harkness, 1994). EFAs were conducted to provide an initial description of the factor structure of the 16-item ASI. Both principal components analysis (PCA) and principal axis factoring (PAF) analyses were conducted using orthogonal and oblique rotations to enhance interpretability. PCA was used, consistent with Nunnally's (1978) recommendation that, in an exploratory analysis, PCA (with unities in the diagonals) is a reasonable analytic strategy. PAF provided a comparative strategy and because it may be better suited to describing underlying latent variables relative to PCA (Widaman, 1993). Factor estimation was based on the following criteria: (1) Kaiser's (1961) criterion to retain factors with unrotated eigenvalues of approximately one or greater; (2) the screen test (Cattell, 1966);
38
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
and (3) the interpretability of resulting factor structures (Gorsuch, 1983), which involves examining solutions with different extraction criteria to determine the point at which trivial or redundant factors emerge (e.g., Schmidt, Joiner, Young, & Telch, 1995). CFAs were based on maximum-likelihood analyses using EQS 5.4 (Bentler, 1995). Consistent with recommendation, a variety of goodness-of-®t measures for the overall model was evaluated (Marsh, Balla, & McDonald, 1988). The nonsigni®cant w2 statistic (assessing whether residual differences between the ®tted population covariance matrix and the observed covariance matrix converge to zero) is the most frequently used index. However, the w2 statistic is problematic as the likelihood of rejecting a model increases with sample size even when the model provides a very good ®t (Bentler & Bonett, 1980). Marsh et al. (1988) note that the w2 statistic will reject virtually all models with very large samples. Three stand-alone indicators (i.e., based on the results of the a priori model ®tting to data) were evaluated including the LISREL's Goodness-of-Fit Index (GFI: values >0.84), which provides a measure of the relative amount of variance and covariance accounted for by the model, the LISREL's Adjusted Goodness-ofFit Index (AGFI: values >0.80), which controls for the degrees of freedom in the model, and the root mean squared residual (RMSR: values <0.10), which is a measure of the difference between the observed and reconstructed correlations. Three incremental ®t indices (i.e., based on the difference between a target model and an alternative model) were also evaluated including the Bentler±Bonett Normed Fit Index (NFI: values >0.90), which represents one of the most widely used ®t indices, the Bentler±Bonett Nonnormed Fit Index (NNFI: values >0.90), which avoids the underestimation of ®t sometimes noted for the NFI in small samples (Bentler, 1990), and the Comparative Fit Index (CFI: values >0.90), which is adjusted for model parsimony and avoids the sampling variability of other incremental ®t indices. 3. Results 3.1. Item analysis of the ASI in a community sample The corrected item±total correlations for each of the 16 ASI items and coef®cient a are presented in Table 1. Coef®cient a for the ASI-16 was .86 but Items 5 and 7 produced questionable item±total correlations. Zero-order correlations among the individual ASI items are presented in Table 2. Particular attention was paid to item±item correlations below .20, because an item that is not at least moderately related with other items measuring the same construct will perform poorly in a factor analysis (Floyd & Widaman, 1995). As would be predicted from low item±total correlations, Items 5 and 7 produced both the fewest signi®cant correlations with other items and the fewest number of correlations greater than .20. In addition, Item 1 had seven correlations <.20.
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
39
Table 2 Intercorrelations among ASI items Item
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
± .19 .27 .30 .49 .22 .10 .19 .06 .18 .19 .28 .26 .22 .15 .24
± .32 .23 .06 .19 .14 .32 .26 .15 .41 .54 .34 .27 .56 .51
± .62 .19 .33 .16 .34 .17 .42 .27 .48 .40 .42 .20 .39
± .21 .50 .21 .51 .24 .55 .16 .40 .33 .54 .11 .36
± .25 .04 .18 .04 .18 .11 .15 .20 .20 .09 .16
± .21 .46 .43 .61 .31 .33 .27 .38 .16 .45
± .27 .31 .17 .19 .25 .19 .27 .11 .13
± .36 .57 .41 .43 .26 .34 .32 .45
± .50 .47 .34 .14 .31 .22 .29
± .35 .38 .17 .39 .14 .36
± .50 .38 .31 .53 .48
± .38 .43 .53 .54
± .41 .41 .42
± .21 .39
± .61
±
N 233.
In sum, preliminary analyses indicated that Items 1, 5, and 7 were weak and should likely be removed. 3.2. PCA and PAF of the ASI Findings from the PCA are summarized in Table 3. Both orthogonal and oblique rotations resulted in a similar pattern of factor loadings. PCA indicated four factors, although the order of the ®rst two factors and the last two factors was reversed depending on orthogonal or oblique rotation. The overall solution accounted for 63% of the variance. Factor 1 accounted for 36.5% of the variance (eigenvalue 5:83), Factor 2 accounted for 10.9% (eigenvalue 1:74), Factor 3 accounted for 9% (eigenvalue 1:44), and Factor 4 accounted for 6.6% (eigenvalue 1:0). Based on items with the greatest salient loadings, these factors could be described as Fear of Cardiopulmonary Sensations, Fear of Mental Catastrophe, Fear of Vasovagal Sensations, and Loss of Control Fears. Every item except Item 7 was loaded (>.39) on at least one factor and Items 11 and 13 were loaded on two factors with the orthogonal rotation. Not surprisingly, the oblique rotation resulted in higher cross-factor overlap with all items producing salient loadings on more than one factor. It is notable that the Loss of Control Fears factor was composed of two weak items (1 and 5) with a borderline eigenvalue (1.0) suggesting that this factor should be eliminated. Similarly, the screen test indicated that, of these four factors, two or possibly three should be retained. PAF analyses using both orthogonal and oblique rotations also yielded fourfactor solutions that were very similar to the corresponding PCA solutions. Examination of items with salient loadings suggests only two changes from
40
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
Table 3 Factor loadings for ASI items ASI items
PCA (orthogonal) I
1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16.
Nervous Going crazy ``Shaky'' Faint Control Heart beats Stomach Nauseous Heart attack Short of breath Stomach Mind Others notice Body sensations Mentally ill Nervous
.17 .76 .22 .02 .03 .08 .12 .28 .24 .02 .65 .64 .52 .22 .86 .68
II .03 .07 .13 .36 .08 .68 .38 .60 .79 .77 .45 .29 .01 .31 .08 .29
PCA (oblique) III
IV
I
.18 .18 .78 .80 .08 .30 .16 .29 .07 .34 .04 .36 .46 .62 .01 .26
.80 .00 .11 .15 .87 .25 .08 .12 .05 .13 .09 .08 .18 .08 .05 .13
.11 .22 .30 .49 .12 .73 .38 .68 .79 .81 .54 .44 .15 .43 .22 .43
II
III .24 .77 .35 .20 .09 .23 .19 .41 .33 .19 .71 .72 .58 .35 .86 .75
.83 .12 .28 .32 .87 .35 .00 .25 .03 .25 .18 .24 .31 .24 .15 .27
IV .28 .29 .81 .85 .18 .44 .23 .43 .09 .47 .13 .49 .53 .69 .12 .40
Factor labels: I (orthogonal) and II oblique Fear of Mental Catastrophe; II (orthogonal) and I oblique Cardiopulmonary Fears; III (orthogonal) and IV oblique Vasovagal Fears; IV (orthogonal) and III oblique loss of control fears.
the pattern of loadings described in Table 3. The orthogonal rotation produced a salient loading for Item 11 on the Cardiopulmonary Fears factor (.45) and the oblique rotation lost a salient loading for Item 3 on the Fear of Mental Catastrophe factor. In general, evaluation of eigenvalues and the screen plot suggests two main factors (Cardiopulmonary and Mental Catastrophe) along with two more questionable factors (Vasovagal and Loss of Control). A summary of the PCA and PAF analyses generally suggest two factors. Stability of the Loss of Control Fears and Vasovagal Fears factors is questionable. Loss of Control Fears is questionable because it is constituted by two weak items (i.e., based on interitem correlations and item±whole correlations) and Vasovagal Fears has high overlap with other factors in the oblique solution. In addition, the eigenvalues greater than 1 rule may overextract factors (Floyd & Widaman, 1995) and both factors possess borderline eigenvalues (1.0) depending upon orthogonal or oblique solution. Based on the information from the correlational and PCA/PAF analyses, it was determined that Items 1, 5, and 7 should be removed from the scale because of their low item±total correlations (Table 1), low interitem correlations (Table 2), and, in the case of Item 7, its failure to load on any factor. 3.3. Factor analyses on the revised ASI As per the suggestion of Floyd and Widaman (1995), items that performed poorly were deleted and a similar set of analyses was conducted with the 13-item
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
41
ASI (ASI-13). The corrected item±total correlations are reported in Table 1, along with the new alpha coef®cient. The internal consistency of the scale improved after the three questionable items were eliminated (a :88). PCA and PAF analyses of the ASI-13 using both orthogonal and oblique rotations were similar in terms of yielding three-factor solutions with similar salient loadings (>.40). PCA and PAF analyses with orthogonal rotation resulted in Items 2, 11, 12, 13, 15, and 16 loading onto a Fear of Mental Catastrophe factor (eigenvalue 5:50, accounting for 42.3% of the variance), Items 6, 8, 9, 10, and 11 loading onto a Cardiopulmonary Fears factor (eigenvalue 1:71, accounting for 13.2% of the variance) and Items 3, 4, 13, and 14 loading onto a Vasovagal Fears factor (eigenvalue 1:20, accounting for 9.2% of the variance). Oblique rotations indicated substantial factor overlap with moderate factor intercorrelations (Factors I±II: r :32; Factors I±III: r :15; Factors I±IV: r :34; Factors II±III: r :19; Factors II±IV: r :29; Factors III±IV: r :29). The PCA indicated that 8 of 13 items, and the PAF indicated that 10 of 13 items, loaded on more than one factor (>.40) with the oblique rotation. In general, analyses using orthogonal rotations were consistent with three-factor solutions as all three factors were interpretable and accounted for suf®cient variance. Oblique solutions were more consistent with analyses using the ASI-16 as they were suggestive of two-factor solutions due to dif®culty in interpreting the Vasovagal Fears factor due to the high level of item complexity (i.e., items loading on multiple factors). 3.4. Con®rmatory factor analyses 3.4.1. Unifactorial models Four separate unifactorial models were constructed based on the original fullscale ASI (ASI-16), the ASI with Items 7 and 13 was removed as suggested by Taylor et al.'s (1992) con®rmatory analyses (ASI-14), the ASI with the ®ve items (i.e., 1, 5, 7, 8, 13) was removed as suggested by Blais et al.'s (in press) analyses (ASI-11), and the ASI with the three items (i.e., 1, 5, 7) was removed as suggested by the analyses reported above (ASI-13). Goodness-of-®t indices for all models are presented in Table 4. It is notable that all latent variable models yielded signi®cant w2 statistics, which is likely to result from the large sample size (Marsh et al., 1988). The overall ®t of each unifactorial model was comparable across most ®t measures (e.g., CFI range: .69±.71) suggesting that all models accounted for the data about equally well. Each model, however, produced a poor ®t according to all criteria except the RMSR suggesting that a unifactorial model does not adequately account for the data. 3.4.2. Multifactorial models Four multifactorial models were tested with both orthogonal and oblique solutions (see Table 4). The Peterson and Heilbronner (1987) and Telch, Shermis, and Lucas (1989) four-factor models were selected based on their superior
42
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
Table 4 Goodness-of-®t indices for latent variable models Model
Fit indices Stand alone w
2
NFI
NNFI CFI
SB
RCFI
.08a .09a .05a .06a
.65 .66 .68 .67
.65 .63 .65 .61
.69 .69 .71 .69
262.1 219.0 164.9 126.1
.68 .68 .71 .70
.70 .80a
.17 .04a
.69 .82
.68 .83
.73 .87
275.3 .64 155.6 .86
Telch, Shermis, and Lucas (1989) Orthogonal 511.9 101 .78 Oblique 397.2 96 .82
.71 .75
.15 .07a
.69 .76
.68 .75
.73 .80
274.3 .65 182.2 .82
Blais et al. (in press) Orthogonal 417.9 Oblique 329.7
.65 .65
.11 .05a
.64 .72
.58 .67
.67 .74
139.0 .65 109.1 .76
Schmidt and Joiner (three-factor) Orthogonal 428.5 65 .80 Oblique 240.1 62 .87a
.72 .81a
.16 .04a
.70 .83
.68 .83
.73 .87
156.7 .72 120.4 .83
Schmidt and Joiner (two-factor) Orthogonal 185.2 35 Oblique 121.8 34 102.3 26 Oblique (8)b Oblique (13)c 113.5 26 Oblique (8,13)d 94.4 19
.82a .86a .85a .84a .83a
.11 .03a .03a .03a .03a
.81 .88 .88 .88 .88
.79 .88 .87 .86 .85
.84 .91a .91a .90a .90a
Unifactorial ASI-16 ASI-14 ASI-13 ASI-11
df
Incremental
570.1 104 504.5 77 458.9 65 387.0 44
GFI
AGFI RMSR
.73 .72 .72 .72
.64 .62 .61 .58
Multifactorial Peterson and Heilbronner (1987) Orthogonal 506.4 100 .78 Oblique 292.6 94 .86a
44 43
.77 .77
.88a .91a .91a .91a .91a
96.5 62.2 52.1 57.5 48.4
.65 .86 .82 .82 .76
GFI: LISREL's Goodness-of-Fit Index; AGFI: LISREL's Adjusted Goodness-of-Fit Index; RMSR: Root Mean Squared Residual; NFI: Bentler±Bonet Normed Fit Index; NNFI: Benter±Bonet Nonnormed Fit Index; CFI: Comparative Fit Index; SB: Satorra±Bentler rescaled w2 statistic; RCFI: Robust Comparative Fit Index. a Indicates meets threshold for adequate ®t (GFI > :84; AGFI > :79; RMSR < :10; NFI, NNFI, CFI > :89). b Item 8 removed from model. c Item 13 removed from model. d Items 8 and 13 removed from model.
performance in previous CFA studies. The Blais et al. (in press) two-factor model, which has not been evaluated using CFA, and the three- and two-factor models obtained from the EFA were also examined. The Peterson and Heilbronner (1987) orthogonal model did not yield an adequate ®t. Criteria standards were not met for the w2 statistic and other ®t indices (w2 506:4, df 100; CFI :73). However, the oblique model produced a substantially better ®t according to all ®t indices. Despite a signi®cant w2 statistic and
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
43
a CFI that did not meet criteria standards (w2 292:6, df 94; CFI :87), the GFI, AGFI, and RMSR indicated adequate ®t (GFI :86, AGFI :80, RMSR :04). The Telch, Shermis, and Lucas (1989) orthogonal and oblique models were not found to adequately ®t the data. The orthogonal model produced no ®t indices meeting criteria standards (w2 511:9, df 101; CFI :73). The oblique model produced a relatively better ®t and did yield an adequate RMSR (.07) but no other ®t indices met criteria standards (w2 397:2, df 96; CFI :80). The Blais et al. (in press) orthogonal and oblique models also were not found to adequately ®t the data. The orthogonal model produced no ®t indices meeting criteria standards (w2 417:9, df 44; CFI :67). The oblique model produced a relatively better ®t and did yield an adequate RMSR (.05) but no other ®t indices met criteria standards (w2 329:7, df 43; CFI :74). Both the three- and two-factor models derived from preliminary PCA and PAF analyses were evaluated. The three-factor orthogonal model did not yield an adequate ®t (w2 428:5, df 65; CFI :73). However, the oblique three-factor model produced a pattern of ®t similar to the Peterson and Heilbronner four-factor oblique solution. Despite a signi®cant w2 statistic and a CFI that did not meet criteria standards (w2 240:1, df 62; CFI :87), the GFI, AGFI, and RMSR indicated adequate ®t (GFI :87, AGFI :81, RMSR :04). The two-factor model produced ®t indices with markedly better ®t relative to all other orthogonal solutions. The two-factor orthogonal model met criteria standards for the GFI (.88) and AGFI (.82) despite a signi®cant w2 and inadequate CFI (w2 185:2, df 35; CFI :84). The two-factor oblique model produced the best overall ®t of all models tested (see Fig. 1). This model met criteria standards for the GFI (.91), AGFI (.86), RMSR (.03), and CFI (.91) despite a signi®cant w2 statistic (w2 121:8, df 34).
Fig. 1. Two-factor oblique model of the ASI in a community sample with standardized path coefficients.
44
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
Three additional analyses were conducted in an attempt to further re®ne the two-factor oblique model. The model was evaluated after the removal of Item 8 (i.e., nausea) from the Cardiopulmonary Fears factor, removal of Item 13 (i.e., others notice when I'm shaky) from the Fear of Mental Catastrophe factor, and after removal of both items. The rationale for these tests was based on the face validity of the items in factors appearing to tap a somewhat different area (e.g., gastrointestinal vs. cardiac and pulmonary items) as well as the fact that Items 8 and 13 had the weakest loadings of all salient items in the PAF and PCA. These revised models yielded highly comparable ®t indices relative to the original twofactor oblique solution suggesting that nothing is gained from removal of these items (see Table 4). The Lagrange Multiplier (LM) and Wald (W) tests provide tests of the statistical necessity of sets of parameters that may be added or deleted from a model (Bentler, 1995). The W test indicated that no paths should be deleted and the multivariate LM test indicated adding several parameters. However, inclusion of any of these parameters created uninterpretable path coef®cients (standardized values >1.0). In sum, a re®ned model was not generated from substantive hypotheses or respeci®cation based on LM or W tests suggesting that it is best to retain the twofactor 10-item (ASI-10) scale. Table 1 shows item±whole correlations for the twofactor 10-item scale. Reanalysis of internal consistency reliability for the ASI-10 indicated adequate internal consistency (overall coef®cient a :86; Factor I coef®cient a :82; Factor II coef®cient a :80). The ASI-10 was highly correlated with the overall ASI-16 (r :92). 3.4.3. Generalizability of latent variable models to a nonclinical college student sample Generalizability of the derived 10-item (ASI-10) two-factor model was assessed by comparing it with the other oblique models (i.e., Peterson & Heilbronner, Telch et al., Blais et al.) in a large sample (N 809) of college undergraduates. The twofactor oblique model derived from the community sample provided an adequate ®t to the data as indicated by several ®t indices (GFI :94, AGFI :91, RMSR :05, NFI :89, NNFI :87, CFI :90). This level of ®t is similar to that obtained for the community sample. The two-factor model provides a somewhat better ®t relative to all other models including the Peterson and Heilbronner four-factor model (GFI :92, AGFI :89, RMSR :05, NFI :86, NNFI :85, CFI : 88), the Telch et al. four-factor model (GFI :92, AGFI :89, RMSR :05, NFI :87, NNFI :87, CFI :89), and the Blais et al. two-factor model (GFI :91, AGFI :86, RMSR :06, NFI :85, NNFI :82, CFI :86). 3.4.4. Additional statistical considerations Multivariate normality is a general requirement for maximum-likelihood CFA that assumes that each variable possesses zero skewness and zero kurtosis (Cole, 1987). Unfortunately, the majority of data collected in behavioral research do not follow univariate or multivariate normal distributions (Micceri, 1989). Two
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
45
different approaches have been suggested to address multivariate nonnormality (Curran, West, & Finch, 1996). One approach is to use methods of estimation (i.e., asymptotic distribution free) that do not assume multivariate normality such that variables possessing nonzero kurtoses pose no special problems. A second approach adjusts the maximum-likelihood w2 estimate for the presence of nonzero kurtosis (i.e., Satorra±Bentler (SB) rescaled w2). Evaluation of multivariate and univariate kurtosis and skewness suggests signi®cant multivariate kurtosis (Mardia's coefficient 240:2) as well as signi®cant kurtosis for Items 2, 9, 11, 15, and 16. Both approaches were evaluated for all CFA models. All models were reanalyzed using EQS's AGLS estimation (i.e., the equivalent of asymptotic distribution free). The overall pattern of goodness-of-®t across all models remained relatively stable with AGLS remodeling, although there tended to be substantial increases in the LISREL indices (e.g., two-factor oblique model: GFI :99, AGFI :98, RMSR :03, NFI :87, NNFI :87, CFI :90; Peterson & Heilbronner oblique model: GFI :99, AGFI :98, RMSR :04, NFI :82, NNFI :83, CFI :86).The adjusted SB w2 statistic as well as a corresponding CFI (i.e., the Robust CFI) are provided in Table 4. The SB statistic yielded lower w2 values but all continued to be statistically signi®cant. The Robust CFI was generally comparable to the CFI, although some orthogonal models produced substantially poorer ®ts according to this index. 4. Discussion Because of the ambiguity that can arise in the interpretation of both EFA and CFA (Floyd & Widaman, 1995), a comprehensive analytic strategy was employed to gain convergence as to the best description of the structure of anxiety sensitivity in a community sample. Utility of this methodology was evidenced in that exploratory analyses suggested either a two- or three-factor solution, whereas con®rmatory analyses clearly indicated the superiority of the two-factor solution described by Fears of Mental Catastrophe and Fears of Cardiopulmonary Sensations. Factor intercorrelations, as well as the superior ®t of an oblique solution, indicate a hierarchical model described by a superordinate anxiety sensitivity factor with two interrelated ®rst-order factors. Anxiety sensitivity researchers have struggled to obtain a consensus in terms of the ASI's factor structure but ®ndings from factor analytic studies, including the present report, suggest common themes. The factors extracted in both the exploratory and con®rmatory analyses were descriptively similar to those in previous studies. In particular, most studies have extracted two separate factors specifying fear of sensations linked with physical catastrophe and fears of sensations linked with mental catastrophe (Blais et al., in press; Peterson & Heilbronner, 1987; Telch, Shermis, & Lucas, 1989; Wardle et al., 1990). For example, the derived two-factor solution corresponds well in terms of items with the highest salient loadings, with the two-factor model (i.e., fear of somatic
46
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
sensations, fear of loss of mental control) obtained by Blais et al. (in press). Findings across these studies indicate that the anxiety sensitivity construct may be described by two interrelated factors that generally tap physical and mental domains. It was hypothesized that beliefs about autonomic arousal may change in the context of relevant life experiences (e.g., having a spontaneous panic attack) that would create differences in the structure of anxiety sensitivity across clinical and nonclinical samples. In fact, the structure of the ASI was highly similar for individuals with no history of spontaneous panic compared to a nonclinical college sample, which undoubtedly included a signi®cant percentage of individuals with a history of spontaneous panic. Moreover, this two-factor model was similar to Blais et al.'s (in press) two-factor solution in a largely psychiatric sample. The major distinction between the community sample and Blais et al.'s mixed sample is that the salience of particular anxiety sensitivity dimensions, as indexed by the primary factor obtained in each study, appears to shift from a focus on fears of mental catastrophe to fears of physical catastrophe in the clinical sample. Thus, the underlying factor structure of anxiety sensitivity, as measured by the ASI, appears to be relatively stable across nonclinical and clinical samples even as the absolute level of anxiety sensitivity increases. Presence of two ®rst-order factors indicates two potential risk dimensions that may contribute to the development of anxiety pathology. Individuals with mental fears may be more apt to experience panic in the context of unexplained or unexpected signs of mental catastrophe (e.g., derealization) whereas those with physical fears may panic when bodily cues (e.g., heart palpitations) mimic physical concerns. Ultimately, these subgroups may produce different anxiety disorders or discernable subtypes of one type of anxiety pathology. For example, the expression of anxiety sensitivity differs among patients with panic disorder (Schmidt, 1999). Although patients with panic disorder may be generally fearful of strong bodily perturbations, patients may often evidence particular fears of one or two speci®c arousal cues. For example, panic disordered patients' beliefs about the consequences of panic-related arousal tend to cluster in three domains: physical, social, and loss of control (Telch, Brouillard, Telch, Agras, & Taylor, 1989). Dimensionality of the ASI is one of the controversial topics within the anxiety sensitivity literature (Taylor et al., 1992). A multidimensional ASI is potentially problematic for the unidimensional construct of anxiety sensitivity. However, Lilienfeld, Turner, and Jacob's (1993) proposition that anxiety sensitivity may be hierarchical allows for both superfactor and multidimensional models (see also Watson et al., 1994). This view is gaining acceptance (Cox et al., 1996; Lilienfeld, Turner, & Jacob, 1996; Zinbarg, Barlow, & Brown, 1997) and is consistent with the present ®ndings. Another important aspect of the dimensionality debate has to do with the degree to which multifactor solutions are intercorrelated. For example, Taylor et al. (1992) argued that four-factor solutions (e.g., Telch, Shermis, & Lucas, 1989) showed high levels of factor-intercorrelation and should therefore be
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
47
best viewed as unidimensional. On the other hand, Cox et al. (1996) note that none of the unidimensional models provides adequate ®t for the data. Indeed, all of the unidimensional models tested in the present study also provided poor ®t for the data. On the other hand, oblique multidimensional models ®t the data much better. Cox et al. comment that it is not uncommon for multidimensional measures of personality traits to be highly intercorrelated. Similarly, subscale indices on intellectual tests also possess high levels of intercorrelation (Carroll, 1993; Horn & Noll, 1997). Analogous to the saturation of intelligence tests with the higher order factor, g, the ®rst-order ASI factors can be considered to be saturated by the higher order dimension of anxiety sensitivity. In sum, anxiety sensitivity appears to be unidimensional at a higher order, but multidimensional at a lower order. Importantly, it is worth reiterating that most ASI models thus far evaluated in the literature provide poor ®t to the data. The best-®tting model in the present study provided an adequate ®t but it is likely that this could be improved. Because of these issues, and in concurrence with recommendations of Cox et al. (1996), it is suggested that the ASI item pool be expanded in order to construct more reliable subscales. Until subscales are expanded, caution should be used in interpreting the subscale scores from the derived factors because of instability that may result from scales with only four or six items. Given the data implicating anxiety sensitivity as a risk factor for anxiety pathology, evaluation of anxiety sensitivity before onset of panic attacks and formal anxiety disorder syndromes is critical. Interestingly, the overall level of anxiety sensitivity in the community sample was substantially lower than the college sample and other nonclinical samples reported in the literature (Telch, Shermis, & Lucas, 1989). This difference is likely to be due to exclusionary criteria (i.e., history of spontaneous panic and psychiatric disorders) for the community sample, which will be present in undergraduate populations. Further evaluation is needed, however, as this disparity may be due to other inherent differences among community versus college nonclinical samples. Acknowledgments The authors thank Margaret Koselka, Jack Trakowski, and Matt Wineman for their assistance with structured diagnostic interviews, and Manuel Orellana for assistance with data management. References Bentler, P. M. (1990). Comparative ®t indices in structural models. Psychological Bulletin, 107, 238±246. Bentler, P. M. (1995). EQS structural equations program manual. Encino, CA: Multivariate Software. Bentler, P. M., & Bonett, D. G. (1980). Signi®cance tests and goodness of ®t in the analysis of covariance structures. Psychological Bulletin, 88, 588±606.
48
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
Blais, M. A., Otto, M. W., Zucker, B. G., McNally, R. J., Schmidt, N. B., Fava, M., & Pollack, M. H. (in press). The Anxiety Sensitivity Index: I. Item analysis and suggestions for re®nement. Journal of Personality Assessment. Carroll, J. B. (1993). Human cognitive abilities: a survey of factor-analytic studies. New York: Cambridge University Press. Cattell, R. B. (1966). The screen test for the number of factors. Multivariate Behavioral Research, 1, 245±276. Cole, D. A. (1987). Utility of con®rmatory factor analysis in test validation research. Journal of Consulting and Clinical Psychology, 55, 584±594. Cox, B. J., Parker, J. D. A., & Swinson, R. P. (1996). Anxiety sensitivity: con®rmatory evidence for a multidimensional construct. Behaviour Research and Therapy, 34, 591±598. Curran, P. J., West, S. G., & Finch, J. F. (1996). The robustness of test statistics to nonnormality and speci®cation error in con®rmatory factor analysis. Psychological Methods, 1, 16±29. Donnell, C. D., & McNally, R. J. (1990). Anxiety sensitivity and panic attacks in a nonclinical population. Behaviour Research and Therapy, 28, 83±85. First, M. B., Spitzer, R. L., Gibbon, M., & Williams, J. B. (1994). Structured clinical interview for DSM-IV Patient Edition. New York: Biometrics Research Department (SCID-N/P Version 2.0). Floyd, F. J., & Widaman, K. F. (1995). Factor analysis in the development and re®nement of clinical assessment instruments. Psychological Assessment, 7, 286±299. Gorsuch, R. (1983). Factor analysis (2nd ed.). Hillsdale, NJ: Erlbaum. Harrington, P. H., Schmidt, N. B., & Telch, M. J. (1996). Prospective evaluation of panic potentiation following 35% CO2 challenge in a nonclinical sample. American Journal of Psychiatry, 153, 823±825. Horn, J. L., & Noll, J. (1997). Human cognitive capabilities: Gf±Gc theory. In: D. P. Flanagan, J. L. Harrison, & P. L. Harrison (Eds.), Contemporary intellectual assessment: theories, tests, and issues (pp. 52±91). New York: Guilford. Kaiser, H. F. (1961). A note on Guttman's lower bound for the number of common factors. Multivariate Behavioral Research, 1, 249±276. Lilienfeld, S. O., Turner, S. M., & Jacob, R. G. (1993). Anxiety sensitivity: an examination of theoretical and methodological issues. Advances in Behaviour Research and Therapy, 15, 147±182. Lilienfeld, S. O., Turner, S. M., & Jacob, R. G. (1996). Further comments on the nature and measurement of anxiety sensitivity: a reply to Taylor (1995b). Journal of Anxiety Disorders, 10, 411±424. Maller, R. G., & Reiss, S. (1992). Anxiety sensitivity in 1984 and panic attacks in 1987. Journal of Anxiety Disorders, 6, 241±247. Marsh, H. W., Balla, J. R., & McDonald, R. P. (1988). Goodness-of-®t indices in con®rmatory factor analyses: the effect of sample size. Psychological Bulletin, 103, 391±410. McNally, R. J. (1990). Psychological approaches to panic disorder: a review. Psychological Bulletin, 108, 403±419. Micceri, T. (1989). The unicorn, the normal curve, and other improbable creatures. Psychological Bulletin, 105, 156±166. Nunnally, J. (1978). Psychometric theory. New York: McGraw-Hill. Nunnally, J. C., & Bernstein, I. H. (1994). Psychometric theory (3rd ed.). New York: McGraw-Hill. Peterson, R. A., & Heilbronner, R. L. (1987). The Anxiety Sensitivity Index: construct validity and factor analytic structure. Journal of Anxiety Disorders, 1, 117±121. Peterson, R. A., & Reiss, S. (1987). Test manual for the Anxiety Sensitivity Index. Orland Park, IL: International Diagnostic Systems. Rapee, R. M., & Medoro, L. (1994). Fear of physical sensations and trait anxiety as mediators of the response to hyperventilation in nonclinical subjects. Journal of Abnormal Psychology, 103, 693±699. Reiss, S. (1991). Expectancy model of fear, anxiety, and panic. Clinical Psychology Review, 11, 141± 153.
N.B. Schmidt, T.E. Joiner / Anxiety Disorders 16 (2002) 33±49
49
Reiss, S., & McNally, R. J. (1985). The expectancy model of fear. In: S. Reiss & R. R. Bootzin (Eds.), Theoretical issues in behavior therapy (pp. 107±121). New York: Academic Press. Reiss, S., Peterson, R. A., Gursky, D. M., & McNally, R. J. (1986). Anxiety sensitivity, anxiety frequency and the prediction of fearfulness. Behaviour Research and Therapy, 24, 1±8. Schmidt, N. B. (1999). Examination of differential anxiety sensitivities in panic disorder: a test of anxiety sensitivity subdomains in predicting fearful responding to a 35% CO2 challenge. Cognitive Therapy and Research, 23, 3±19. Schmidt, N. B., Joiner, T.E., Jr., Young, J., & Telch, M. J. (1995). The schema questionnaire: investigation of psychometric properties and the hierarchical structure of a measure of maladaptive schemas. Cognitive Therapy and Research, 19, 295±321. Schmidt, N. B., Lerew, D., & Jackson, R. J. (1997). The role of anxiety sensitivity in the pathogenesis of panic: prospective evaluations of spontaneous panic attacks during acute stress. Journal of Abnormal Psychology, 106, 355±364. 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±1109. Schmidt, N. B., & Telch, M. J. (1994). The role of safety information and fear of bodily sensations in moderating responses to a hyperventilation challenge. Behavior Therapy, 25, 197±208. Taylor, S., Koch, W. J., & Crockett, D. J. (1991). Anxiety sensitivity, trait anxiety, and the anxiety disorders. Journal of Anxiety Disorders, 6, 293±311. Taylor, S., Koch, W. J., McNally, R. J., & Crockett, D. J. (1992). Conceptualizations of anxiety sensitivity. Psychological Assessment, 4, 245±250. Telch, M. J., Brouillard, M., Telch, C. F., Agras, W. S., & Taylor, C. B. (1989). Role of cognitive appraisal in panic-related avoidance. Behavior Research and Therapy, 27, 373±383. Telch, M. J., Shermis, M. D., & Lucas, J. A. (1989). Anxiety sensitivity: unitary personality trait or domain-speci®c appraisals? Journal of Anxiety Disorders, 3, 25±32. Telch, M. J., Silverman, A., & Schmidt, N. B. (1996). The relationship between anxiety sensitivity and perceived control in a caffeine challenge. Journal of Anxiety Disorders, 10, 21±35. Wardle, J., Ahmad, T., & Hayward, P. (1990). Anxiety sensitivity in agoraphobia. Journal of Anxiety Disorders, 4, 325±333. Watson, D., Clark, L. A., & Harkness, A. R. (1994). Structures of personality and their relevance to psychopathology. Journal of Abnormal Psychology, 103, 18±31. Widaman, K. F. (1993). Common factor-analysis versus principal component analysisÐdifferential bias in representing model parameters. Multivariate Behavioral Research, 28, 263±311. 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±281.