ANNOTATED

DEPRESSION AND ANXIETY 32:509–517 (2015)

Research Article IDENTIFYING PANIC DISORDER SUBTYPES

USING FACTOR MIXTURE MODELING

Thomas Pattyn, M.D.,1,2∗ Filip Van Den Eede, M.D., Ph.D.,2,1 Femke Lamers, Ph.D.,3,4 Dick Veltman, M.D., Ph.D.,4 Bernard G. Sabbe, M.D., Ph.D.,1,5 and Brenda W. Penninx, Ph.D.4

Background: The clinical presentation of panic disorder (PD) is known to be highly heterogeneous, complicating research on its etiology, neurobiological path- ways, and treatment. None of the attempts to identify PD subtypes have been inde- pendently reproduced, rendering the current literature inconclusive. Methods: Using a data-driven, case-centered approach (factor mixture modeling) on a broad range of anxiety symptoms assessed with the Beck anxiety inventory, the present study identifies PD disorder subtypes in a large (n = 658), well- documented mixed-population sample from the Netherlands Study of Depression and Anxiety (NESDA), with subtypes being validated and detailed using a va- riety of clinical characteristics. Results: A three-class, one-factor model proved superior to all other possible models (Bayesian information criterion = 13,200; Lo-Mendel-Rubin = 0.0295; bootstrapped likelihood ratio test �0.0001), with the first class, a cognitive-autonomic subtype, accounting for 29.8%, the second class, the autonomic subtype, for 29.9%, and a third class, the aspecific subtype, for 40.3% of the population. The cognitive-autonomic and autonomic subtypes showed significant differences compared to the aspecific subtype (e.g., comor- bidity and suicide attempts) but on severity differed between themselves only. Conclusion: Three qualitatively different PD subtypes were identified: a severe cognitive-autonomic subtype, a moderate autonomic subtype, and a mild aspe- cific subtype. Qualitative and quantitative differences were related to severity and clinical properties such as comorbidity, suicide attempts, sleep, and sense of mastery. Depression and Anxiety 32:509–517, 2015. C© 2015 Wiley Periodicals, Inc.

1Collaborative Antwerp Psychiatric Research Institute (CAPRI), University of Antwerp, Antwerp, Belgium 2Department of Psychiatry, Campus Antwerp University Hos- pital, Antwerp, Belgium 3National Institute of Mental Health, Bethesda, Maryland, USA 4Department of Psychiatry, EMGO Institute of Health and Care Research, Neuroscience Campus Amsterdam, VU University Medical Center, Amsterdam, The Netherlands 5Department of Psychiatry, Campus Psychiatric Hospital Duffel, Duffel, Belgium

The infrastructure for the NESDA (www.nesda.nl) is funded through by the Geestkracht programme of the Netherlands Organization for Health Research and Development (Zon-Mw, grant no. 10-000- 1002) and is supported by the following participating universities and mental health care organizations: VU University Medical Cen- ter, GGZ inGeest, Arkin, Leiden University Medical Center, GGZ Rivierduinen, University Medical Center Groningen, Lentis, GGZ

Friesland, GGZ Drenthe, Scientific Institute for Quality of Health- care (IQ healthcare), the Netherlands Institute for Health Services Research (NIVEL) and the Netherlands Institute of Mental Health and Addiction (Trimbos). We are very much grateful for the bequest Cassiers for panic research on panic disorder (CAPRI) grant no. 5739.

∗Correspondence to: Thomas Pattyn, Collaborative Antwerp Psy- chiatric Research Institute (CAPRI), University of Antwerp, Univer- siteitsplein 1 R3.22, 2610 Antwerp, Belgium. E-mail: [email protected] Received for publication 30 October 2014; Revised 5 February 2015; Accepted 21 April 2015

DOI 10.1002/da.22379 Published online 26 May 2015 in Wiley Online Library (wileyonlinelibrary.com).

C© 2015 Wiley Periodicals, Inc.

510 Pattyn et al.

Key words: anxiety disorder; classification; factor analysis; latent class analy- sis; panic; factor mixture modeling; subtype; Beck anxiety inventory; NESDA; nosology

Abbreviations

PD = panic disorder EFA = exploratory factor analysis LCA = latent class analysis

FMM = factor mixture modeling BAI = Beck anxiety inventory ASI = anxiety sensitivity index

FEAR = fear questionnaire WHODAS = World Health Organization Disability

Assessment Schedule

INTRODUCTION In the DSM-IV and DSM-5, the most recent version, panic disorder (PD) is defined as having recurrent and unexpected panic attacks or episodes of intense fear or discomfort with at least four of 13 ad- ditional symptoms,[1, 2] thus covering a broad variety of clinical presentations. PD is also known to have a high lifetime and current comorbidity with psychiatric disor- ders such as depression and other anxiety disorders.[3] Furthermore, the literature shows substantial inconsis- tencies in neurobiological findings for PD, e.g., with regard to HPA-axis functioning,[4] electrophysiological activity,[5] regional brain morphology, and neurotrans- mitter receptor density.[6] All of the above indicates that the pathogenesis and manifestations of PD are likely to be far more heterogeneous than originally assumed, which inherently complicates treatment but also clinical research.

To date, theoretical research has frequently focused on respiratory and cognitive aspects of PD. Major the- oretical models of panic are the cognitive theories of Beck, Clark and Bandura,[7–10] and Klein’s false suffo- cation alarm.[11] A key feature of cognitive models is cognitive misinterpretation, such as catastrophic think- ing, that occurs in response to a normal physiological sensation. In contrast, in Klein’s model, a biologically erroneous stimulus, for example an abnormal CO2 sen- sitivity, prompts a cognitively correct interpretation.[10] More recently, a context-sensitivity panic-vulnerability model.[12] has been proposed as an explanatory model for the propensity to experience panic in situations that pose no objective threat to the individual (cf., cognitive mis- interpretation or Klein’s false suffocation alarm model).

Employing a more data-driven, variable-centered approach by applying a factor analysis to 14 DSM-III- R panic symptoms to identify PD subtypes in a clin- ical sample, Briggs et al.[13] concluded that, in a two- dimensional model, the presence of respiratory items or their absence was the most distinguishing criterion. More recently, Roberson-Nay and Kendler.[14] adopted

exploratory factor analysis (EFA), latent class analysis (LCA), and factor mixture modeling (FMM), the latter two both case-centered approaches, to identify possible subtypes in different samples mostly using DSM-III and DSM-III-R panic symptoms. Results revealed at least two subtypes in all samples, with the strongest subtype discriminator being high or low loadings on respiratory items, prompting the authors to propose a respiratory and a nonrespiratory subtype with a possible third mild respiratory subtype.[14]

The aim of the present study was to investigate PD subtypes in a large and well-documented, contemporary, mixed-population sample using EFA, LCA, and FMM on a broad range of panic symptoms derived from the 21 items of the Beck anxiety inventory (BAI). Second, we aimed to validate these subtypes using other anxi- ety indicators and to identify distinctive demographic and clinical characteristics for the various subtypes. The study was carried out in accordance with the latest ver- sion of the Helsinki Declaration and was approved by the Medical Ethics Committee of the Antwerp Univer- sity Hospital.

MATERIALS AND METHODS SAMPLE

Data were derived from the Netherlands Study of Depression and Anxiety (NESDA), an ongoing multisite naturalistic cohort study. The sample comprises 2,981 participants aged between 18 and 65 years who were recruited between September 2004 and Febru- ary 2007 from the general community, general (family) practices, and mental health facilities to ensure the representation of depression and anxiety psychopathology in various developmental stages and different treatment settings. Detailed descriptions of NESDA’s design can be found elsewhere.[15] Psychiatric diagnoses (DSM-IV) were made us- ing the Composite International Diagnostic Interview, version 2.1.[16]

Although a key feature of PD, panic attacks occur episodically, varying from daily to once a month, while for a sound PD diagnosis, patients must also be presenting with a persistent concern or change in behav- ior for at least one month. Thus, to ensure inclusion of all types of patients, we selected 670 patients with a current baseline diagnosis of PD with or without agoraphobia with a 6-month recency criterion. Twelve patients were excluded because of incomplete information on anxiety symptoms, resulting in a final sample of 658 participants.

BECK ANXIETY INVENTORY As it was our objective to include a broad range of panic-related

anxiety symptoms in our analysis, we opted for the BAI, a 21-item self- report instrument that assesses the overall severity of anxiety[17] and is designed to improve discrimination between depression and anxiety. Respondents are asked to rate how much they have been bothered by each symptom in the last week on a 4-point Likert scale (1 = not at all; 2 = mildly; 3 = moderately; 4 = severely). The items were dichotomously coded as absent = 1/2 or present = 3/4. We used all 21 items for the EFA, LCA, and FMM. All DSM items, excluding chest

Depression and Anxiety

Research Article: Identifying Panic Disorder Subtypes 511

pain and derealization, were well represented in the BAI, with some being represented by two or more BAI items (DSM: trembling, BAI: hands trembling/wobbliness in legs). The BAI additionally includes five items not represented in the DSM-IV PD criteria: being unable to relax, feeling terrified, nervous, scared, and fear of worst happening. The reliability and validity of the BAI have been well established in general populations and in a psychiatric sample [17–19].

VALIDATION AND DESCRIPTION OF IDENTIFIED SUBTYPES

Four groups of variables were evaluated to validate the identified PD subtypes and to detail subtype-specific features:

Demographics. Age, sex, and educational level (in years). Anxiety Indicators. The anxiety sensitivity index (ASI) [20] as-

sessing the fear of anxiety-related sensations, the Fear Questionnaire (FEAR) [21] assessing the severity of phobic avoidance, and two vali- dated BAI subscales[22]: a 14-item somatic (1–3, 6–8, 11–13, 15, 18–21) and a 7-item subjective subscale (4, 5, 9, 10, 14, 16, 17).

Clinical Variables. Comorbidity, that is, current comorbid af- fective disorders (major depressive disorder, social phobia, general anx- iety disorder), number of suicide attempts, frequent use of antidepres- sants, benzodiazepine usage (at least 50% of the time), the WHODAS functioning score (World Health Organization Disability Assessment Schedule II)[23] as an overall disability index, the mastery scale score[24]

indicating the extent to which a person perceives himself to be in con- trol of events or situations, and the three subscale scores of the Mood and Anxiety Symptom Questionnaire-30,[25] with the negative affect subscale gauging emotional states such as fear, anger, or guilt; the positive affect subscale assessing moods such as feeling active, excited, delighted, enthusiastic, interested; and the somatic arousal subscale evaluating physiological arousal symptoms such as trembling, shaking, dizziness, sweating, and heart racing, having some overlap with BAI items.

Physical Health Indicators. Current smoking; sleep (insomnia rating scale scores)[26]; physical activity as assessed with the Inter- national Physical Activity Questionnaire[27] and expressed per 1,000 metabolic equivalent min/week; the number of somatic diseases for which treatment is being received including respiratory disease, os- teoarthritis, cancer, gastrointestinal disease, liver disease, epilepsy, thy- roid disease, diabetes, and cardiovascular disease.

STATISTICAL ANALYSIS To identify PD subtypes we used FMM, a statistical method that

combines principles of EFA and LCA.[28] While EFA is a variable- centered approach that models the data in a set of latent continuous factors, LCA uses a case-centered approach that allows latent categor- ical factors to be identified and assigns cases to these classes. FMM allows latent categorical classes (subtypes) to be identified while con- sidering a continuous factor (dimensionality) in one integrated model. A more comprehensive explanation of these techniques can be found elsewhere.[14] , [28–33] In the current study, we adopted the Mplus 5.1 software[29] for all three methods.

Our stepwise analysis started with EFA to examine panic symp- tom dimensionality and factor structures (Fig. 1, step 1). Statistical interpretation was inferred from a combination of the Kaiser criterion (eigenvalues >1), qualitative interpretation of factor loadings, and a screeplot. A limitation of EFA is its assumption of homogeneity, which may not be met in a PD population. In the second step, LCAs were conducted to examine possible categorical panic classes. An important limitation of this technique is its inability to integrate dimensionality (e.g., panic severity) in its model (Fig. 1, step 2). Therefore, in the third and final step we ran a series of FMMs, integrating the principles of EFA and LCA, to examine all class-factor combinations withheld

with EFA and LCA, with FMM model types ranging from restrictive models to models of increasing flexibility (Fig. 1, step 3).

In the first FMM type, the full invariant type, item threshold, and factor loadings were class invariant, factor variance was set to be equal across classes, and factor means fixed to zero. Two partially invariant FMM types were examined; one was defined with class noninvari- ant factor loadings and factor variance (the factor-invariant type) and the second defined by class noninvariant thresholds (the threshold- invariant type). A fourth, noninvariant, type was examined, where all parameters were class noninvariant and allowed to be freely estimated (see Fig. 1, step 3). LCA and FMM results were evaluated by model- fit statistics, designating how well the data statistically fits the model, such as the Bayesian information criterion (lower is preferred),[30, 31]

Lo-Mendell-Rubin’s chi-square difference test (<0.05 is preferred),[32]

the bootstrapped likelihood ratio test ( �0.05 is preferred),[33] and the theoretical interpretability of the model.

Subtypes were described using SPSS 20 on demographics, anxiety indicators, clinical variables, and physical health indices. Appropri- ate parametric tests (chi-square for categorical variables, analysis of variance for continuous variables) and nonparametric tests (Kruskal- Wallis) were performed. Post hoc analyses were adjusted by means of Bonferroni corrections (analysis of variance) and Dunn test (Kruskal- Wallis) at a significance level of 0.05.

RESULTS SAMPLE

Our final sample included 658 PD patients with a mean age of 41 years, of whom 70.5% was female.

EXPLORATORY FACTOR ANALYSIS Following the Kaiser criterion, a maximum of five

factors could be extracted from the data (see Supple- mentary Tables, online). Further investigation using the screeplot technique and qualitative interpretation of the factor loadings showed high factor loadings on all BAI items when a single factor was extracted, suggesting a severity dimensionality. When two or more factors were extracted, qualitatively distinct factors emerged, sug- gesting distinct categorical classes and indicating the irrelevance of a multifactorial structure. Importantly, between-item correlations may be due to mean differ- ences between possible classes, meaning that observed factor structures may not directly translate into classes. Nonetheless, up to three factors were included in the FMM analysis.

LATENT CLASS ANALYSIS A four-class model showed the best model-fit statistics

(Bayesian information criterion = 13,500; Lo-Mendel- Rubin = 0.0001; bootstrapped likelihood ratio test �0.0001) (Table 1, upper section). A five-class model yielded a better (lower) Bayesian information crite- rion value of 13,475 and a similar bootstrapped likeli- hood ratio test (<0.0001) but the P-value of the Lo- Mendell-Rubin was no longer significant ( =0.44), indi- cating that the five-class model did not improve model fit. Interpretation of the four-class model revealed two qualitatively distinct classes and two comparable classes but with lower item probabilities, merely delineating a

Depression and Anxiety

512 Pattyn et al.

Figure 1. Overview of analytical strategy.

severe and less severe category within these two qualita- tively distinct classes. Up to four classes were included in the FMM.

FACTOR MIXTURE MODELING FMM models up to five classes and three factors

were examined, each with four different model types as explained above (one invariant, two semi-invariant, and one noninvariant type). All multifactorial models yielded poorer model-fit statistics than unifactorial models. Results from the unifactorial, factor-invariant, multi-

ple classes models are presented in Table 1, lower section.

A three-class, one-factor model yielded the best fit (Bayesian information criterion = 13,200; Lo-Mendel- Rubin = 0.0295; bootstrapped likelihood ratio test �0.0001). The results listed in Fig. 2 were therefore taken from the three-class, one-factor model, which was overall the best fitting data-driven solution. In the fig- ure, the item probabilities for the three-class one-factor model are graphically depicted. Mean posterior prob- abilities for most likely class membership were 0.895 for class 1, 0.830 for class 2, and 0.836 for class 3,

TABLE 1. Latent class and factor mixture fit indices

Latent class analysis Number of classes Pars BIC LMR VLMR BLRT Logl

1 21 15904.960 / / / -7884.343 2 43 13874.255 0.0000 0.0000 0.0000 -6797.610 3 65 13582.977 0.0000 0.0000 0.0000 -6580.589 4 87 13500.513 0.0001 0.0001 0.0000 -6467.976 5 109 13475.329 0.4375 0.4341 0.0000 -6384.003

Partially invariant factor mixture modela

Number of classes Pars BIC LMR VLMR BLRT Logl 1 43 13382.183 / / / -6551.574 2 67 13243.409 0.0099 0.0096 0.0000 -6404.316 3 91 13200.570 0.0295 0.0289 0.0000 -6305.026 4 115 13249.323 0.5012 0.4988 0.0000 -6251.532 5 NC NC NC NC NC NC

BIC, Bayesian information criterion; V(LMR), (Voung) Lo-Mendell-Rubin likelihood ratio test; BLRT, bootstrapped likelihood ratio test; NC, not calculated because earlier defined models were found to be nonsignificant. Lower BIC values indicate better fit. (V)LMR and BTLR P-values <0.05 are acceptable. aPartially invariant unifactorial model; factor loadings constrained to be equal across classes, thresholds are free; factor means fixed at zero in the first group and free in others.

Depression and Anxiety

Research Article: Identifying Panic Disorder Subtypes 513

Figure 2. Three-class one-factor mixture model.

indicating a good classification quality. Class 1 accounted for 29.8% (n = 196) of the sample and had the highest item probabilities on almost all items but was mainly differentiated from the other two classes by having very high item probabilities on cognitive items such as be- ing scared, terrified, fear of losing control, fear of dying, and fear of the worst happening. We therefore labeled this class “the cognitive-autonomic subtype”. Class 2 ac- counted for 29.9% (n = 197) and was characterized by the lowest probabilities on respiratory items such as feel- ing of choking and difficulty breathing while also having low probabilities on cognitive items but similar levels of autonomic item probabilities (e.g., feeling hot, indi- gestion, and face flushed) as class 1 and was therefore termed “the autonomic subtype”. Accounting for 40.3% (n = 265) of the sample, class 3 was the largest class, had overall low item probabilities and was mainly differenti- ated from the other two classes by having low autonomic and neurological item probabilities on items such as indi- gestion, numbness, hands/legs trembling, feeling shaky, and dizziness.

DESCRIPTION OF SUBTYPES Demographics, anxiety indicators, and clinical char-

acteristics are shown in Tables 2 and 3. Classes did not differ as to age, sex, or educational level. Examination of other additional anxiety indicators showed a clear or- der among the classes, where class 1 scored higher than class 2 on every indicator, with class 3 having the lowest scores on all anxiety indicators, indicating that classes

differed on anxiety-severity indices as well as on panic- symptom dimensions. Post hoc analysis showed that to- tal BAI scores differed statistically and clinically among all three subtypes, with class 1 being classified as severe (>25), class 2 as moderate (between 16 and 25), and class 1 as mild (between 8 and 15), while on the somatic sub- scale class 1 and class 2 only differed from class 3 but not from each other. Highest anxiety sensitivity and fear of phobias was found in class 1 with statistically signifi- cantly lower scores for class 2 and the lowest scores for class 3 consistent with item probabilities for cognitive items.

All classes also revealed a high current psychiatric co- morbidity. Class 1 differed from the two other classes by having the highest current agoraphobia rates (70.9% vs. 59.9 and 60.8%, respectively). Class 3 distinguished itself from the other two classes with lower scores on all three Mood and Anxiety Symptom Questionnaire scales, better functioning, fewer suicide attempts, and a higher sense of mastery. The use of benzodiazepines was sig- nificantly greater in classes 1 and 2 compared to class 3, but no between-class differences were found in the use of antidepressants. However, the prevalence of current comorbid depressive disorder was significantly higher in classes 1 and 2 compared to class 3 (Table 3).

With regard to physical health indices, no differences were found for current smoking, physical activity, and number of somatic diseases. Only the subjective sleep index signaled significantly more sleeping problems (insomnia rating scale >8) in classes 1 and 2 relative to class 3.

Depression and Anxiety

514 Pattyn et al.

TABLE 2. Demographic characteristics and anxiety indicators

Class 1 (n = 196) cognitive-autonomic

Class 2 (n = 197) autonomic

Class 3 (n = 265) aspecific

Overall P-value

Demographic characteristics Age, mean (SD, year) 39.43 (11.94) 41.12 (12.22) 41.95 (11.46) 0.760 Female 71% 68% 73% 0.510 Education, mean (SD, year) 11.19 (3.41) 11.24 (3.44) 11.60 (3.12) 0.340

Anxiety indices BAI total score, mean (SD) 29.30 (9.80) 20.97 (7.76) 15.41 (11.31) <0.001a , b , c

Somatic subscale, mean (SD) 15.73 (7.49) 13.52 (6.05) 9.56 (7.52) <0.001a , b

Subjective subscale, mean (SD) 13.57 (3.29) 7.45 (2.77) 5.85 (4.32) <0.001a , b , c

ASI, mean (SD) 41.75 (10.66) 37.20 (9.08) 33.96 (9.99) <0.001a , b , c

FEAR, mean (SD) 45.54 (22.01) 39.25 (20.24) 32.68 (20.46) <0.001a , b , c

BAI, Beck anxiety inventory; ASI, anxiety sensitivity index; SD, standard deviation. aPost hoc analysis with Bonferroni correction at 0.05: Cognitive-autonomic is significantly different from aspecific. bPost hoc analysis with Bonferroni correction at 0.05: Autonomic is significantly different from aspecific. cPost hoc analysis with Bonferroni correction at 0.05: Cognitive-autonomic is significantly different from autonomic.

DISCUSSION The present study sought to identify subtypes of

panic disorder in a large patient sample using fac- tor mixture techniques allowing the identification of categorical classes while incorporating dimensional- ity in an integrated model with BAI scores as input.

Based on our data, we propose three symptom-based PD subtypes: a “cognitive-autonomic” subtype with a 29.8% prevalence, an “autonomic” subtype with a 29.9% prevalence, and an “aspecific” subtype with a 40.3% prevalence. We found additional quantitative differ- ences among the subtypes mainly in the severity di- mension, with the cognitive-autonomic subtype being the most and the aspecific subtype the least severe. The

TABLE 3. Clinical characteristics

Class 1 (n = 196) cognitive-autonomic

Class 2 (n = 197) autonomic

Class 3 (n = 265) aspecific

Overall P-value

Psychiatric comorbidity Current comorbid anxiety disorder 67.3% 64.0% 47.2% <0.001a , b

Social phobia 57.1% 48.7% 35.8% <0.001a , b

Generalized anxiety disorder 33.2% 32.5% 23.4% 0.033a , b

PD with agoraphobia 70.9% 59.9% 60.8% 0.037a , b

Current comorbid depressive disorder 66.8% 63.5% 50.6% 0.001a , b

Clinical characteristics Mastery scale, mean (SD) 13.87 (4.32) 14.63 (4.10) 16.42 (3.81) <0.001a , b

WHODAS, mean (SD) 47.07 (21.03) 45.83 (20.33) 30.49 (22.81) <0.001a , b

MASQ Somatic arousal, mean (SD) 22.89 (6.67) 21.63 (6.46) 17.86 (6.78) <0.001a , b

Negative affect, mean (SD) 27.56 (8.95) 25.55 (8.11) 20.93 (7.97) <0.001a , b

Positive affect, mean (SD) 39.45 (7.99) 39.25 (7.55) 35.09 (9.27) <0.001a , b

Number of serious suicide attempts, mean (SD) 0.47 (1.41) 0.46 (0.96) 0.27 (0.87) 0.052a

Current use of benzodiazepines 37.8% 37.1% 23.8% 0.001a , b

Current use of antidepressants 48.5% 49.7% 41.9% 0.184 Physical health indicators

Current smoker 51.5% 54.3% 47.2% 0.750 Sleep (insomnia rating scale >8) 67.5% 61.7% 51.5% 0.005a , b

Number of chronic somatic diseases, mean (SD) 0.95 (1.08) 1.13 (1.17) 1.02 (1.08) 0.295 Physical activity (MET/week), mean (SD) 3,407 (2,781) 3,800 (3,139) 3,674 (3,348) 0.466

PD, panic disorder; MET, metabolic equivalent minutes; MASQ, Mood and Anxiety Symptom Questionnaire; WHODAS, World Health Orga- nization Disability Assessment Schedule II; SD, standard deviation. aPost hoc analysis with Bonferroni correction at 0.05: Cognitive-autonomic is significantly different from aspecific. bPost hoc analysis with Bonferroni correction at 0.05: Autonomic is significantly different from aspecific. cPost hoc analysis with Bonferroni correction at 0.05: Cognitive-autonomic is significantly different from autonomic.

Depression and Anxiety

Research Article: Identifying Panic Disorder Subtypes 515

cognitive-autonomic and autonomic subtypes mostly showed significant differences in anxiety sensitivity and phobic avoidance (as assessed with ASI and FEAR).

Our results are in line with some, but not all, earlier efforts to identify PD subtypes. Adopting a factorial ap- proach, Meuret et al.[34] concluded that panic symptom dimensions have an impact on the severity and mastery of anxiety symptoms and proposed a three-factor solu- tion consisting of a cardiorespiratory, a mixed somatic, and a cognitive dimension. These dimensions coincide with two of our subtypes, where the cardiorespiratory and cognitive dimensions are most strongly represented in the cognitive-autonomic subtype and the mixed so- matic dimension in the autonomic subtype. However, the third category we identified was characterized by the absence of any clear dimensions and hence labeled the aspecific subtype.

Using a similar approach but with different results, Roberson-Nay and Kendler (2011) concluded that in all but one of their samples, a two-class one-factor model was preferable. This discrepancy may be explained by differences among cohorts: a treatment-seeking sam- ple, a twin sample, and three epidemiological sam- ples in the Robertson-Nay and Kendler study[14] and a patient sample recruited from the general population, family practices, and outpatient clinics in ours. Also, earlier diagnoses were based on DSM-III/DSM-III-R criteria rather than the DSM- IV criteria in our study, with definitions differing mainly as to the “required” number and frequency of panic attacks. Theoretically, these disparities may have led to the exclusion of patients with a particular PD type, which would then explain Robertson-Nay and Kendler’s two-class solution.[14] Ad- ditionally, applying DSM-IV criteria, the authors also found a one-factor three-class model with a severe respi- ratory, a mild respiratory, and a nonrespiratory subtype.

Respiratory aspects of PD have attracted ample atten- tion from researchers largely owing to Klein’s false suf- focation alarm theory and the widespread use of carbon dioxide inhalation to induce panic attacks in laboratory settings.[11, 35] Arguably, this field of research tends to view cognitive symptoms as secondary phenomena,[36] potentially creating a bias toward respiratory subtyping. Roberson-Nay and Kendler[14] based their definition of respiratory subtypes on the high loadings on cardiores- piratory symptoms and high endorsements of all DSM items, whereas the more cognitive subtype was charac- terized by high levels of anxiety but low endorsement on all DSM items. In contrast and showing the high- est endorsement on all DSM items, our first class was characterized by prominent cardiorespiratory symptoms but equally so by high anxiety levels. Our findings thus indicate that the extent to which respiratory and cog- nitive subtypes differ may be limited, whereas our dis- tinction between autonomic and aspecific subtypes may have more clinical relevance. Nevertheless, additional longitudinal research on the course and stability of the proposed subtypes is clearly necessary to establish their clinical usefulness.

Evaluating whether the three classes showed substan- tial differences on demographic features, anxiety scores, clinical variables, and physical health indices, we mostly found differences between classes 1 and 2 relative to class 3, with the exception of the ASI and FEAR to- tal scores that also differed between class 1 and class 2, where ASI indicates anxiety sensitivity and FEAR total scores phobic avoidance. This is an interesting finding considering Telch et al.’s context-sensitivity vulnerabil- ity model,[12, 37] where anxiety sensitivity is an important and possibly predictive measure of panic attacks induced by internal or external contextual cues. Class 3 was asso- ciated with lower psychiatric comorbidity, fewer suicide attempts, lower insomnia and WHODAS ratings, and less use of benzodiazepines than classes 1 and 2. Besides qualitative differences on symptomatology, the subtypes were also significantly different on anxiety indicators and clinical characteristics, differences that could not be ex- plained by any demographic dissimilarities.

Following the proposed cut-off values,[18, 36] the BAI total scores for each subtype were also clinically sig- nificant, with class 1 classifying as severe (>25), class 2 as moderate (16–25), and class 3 as mild (8– 15).Comorbidity profiles also characterized classes 1 and 2, while showing that class 3 may represent a less severe clinical entity. Interestingly, the cognitive-autonomic subtype had the highest depressive comorbidity but the lowest use of antidepressants, for which we have no obvious explanation. The above mainly illustrates a severity dimension, where the first class “the cognitive- autonomic subtype” is the most severe, the second class “the autonomic subtype” a more moderate, and the third class “the aspecific subtype” a mild subtype.

Additionally, classes 1 and 2 had differential scores on the BAI subjective subscale, the ASI, and FEAR total scores but not on the BAI somatic subscale or the Mood and Anxiety Symptom Questionnaire subscales. This is consistent with the qualitative differences we obtained for the two subtypes on the cognitive items. Item proba- bilities for all cognitive items were also highest in class 1 and lowest in class 3, in accordance with differential ASI and FEAR total scores, signifying that the level of anxiety sensitivity and/or phobic avoidance may be an impor- tant factor in the development of cognitive symptoms and more severe panic symptoms. Also, there were no between-subtype differences on the mastery scale, sug- gesting that the ASI might be a better predictor than the mastery scale. However, as we computed the highest mastery scale score for class 3, it is plausible to suggest that a higher sense of mastery influences the clinical pre- sentation, which would be consistent with class 3 having the lowest scores for the BAI subjective subscale, ASI, and FEAR.

This paper provides clear evidence that FMM can be effectively used to explore possible subtypes in psy- chiatric research. While FMM principles are generally known, advantages and disadvantages of the technique need to be considered for a proper interpretation of the results. Besides using a robust statistical method to

Depression and Anxiety

516 Pattyn et al.

evaluate the computed FMM models as to classification criteria and probability scores, the decision on the fi- nal best-fit model should take the interpretability of the model into account. While this allows for current theo- ries to be included, thereby enhancing the clinical use- fulness of the results, it also implies that choosing the final model is to some extent subjective. However, in our comprehensive analyses, both the statistical results and interpretability clearly pointed to the same three-class one-factor model.

Our study does have some other limitations that war- rant discussion. While trained interviewers made the PD diagnoses, the BAI—from which we derived the symp- tom ratings—is a self-report questionnaire, which may partly explain differences with earlier studies. Moreover, the BAI was developed by a cognitive psychiatrist,[18] which may have introduced a bias in the nature and num- ber of (cognitive) items. Also, the BAI was not originally designed to measure panic but rather anxiety in gen- eral. However, the literature shows that the scale is more closely aligned to PD than to any other anxiety disor- der, with higher BAI scores in panic populations than in nonpanic anxiety populations,[38–41] while its items cap- ture the DSM-criteria for PD best. Previous studies us- ing DSM panic symptom criteria for statistical analysis were limited by their retrospective assessment of specific symptoms that only briefly occur during a panic attack (seconds to minutes) that may have gone largely unno- ticed at the time of the attack because other symptoms were predominant. Also, in view of the current discus- sions about the boundaries between psychiatric disor- ders, it can be argued that DSM-IV panic symptoms may be too rigid or arbitrary, which is supported by the newly formulated DSM-5 cultural-specific panic attack criteria such as headache, tinnitus, etc. A more extended ques- tionnaire like the BAI may overcome these limitations. Unfortunately, the BAI does not gauge certain panic fea- tures such as anticipated panic, perceived panic conse- quences, or perceived inability to effectively cope with panic attacks. Future research should also consider these features.

In order to validate the proposed PD subtypes and substantiate the clinical relevance of subtyping in gen- eral, future research should investigate their longitu- dinal stability, that is, whether the subtypes are con- sistent throughout the course of the disorder and its treatment. Also, the symptomatically different subtypes should be tested for their potentially different neurobio- logical pathways, etiologies, and treatment approaches. Additional genetic information, neuroanatomical, neu- rophysiological, and cognitive variables on possible sub- types are crucial and should, when becoming available, be included in the subtyping scheme to validate the subtypes further and to improve the clinical benefits of the subtypes. Furthermore, a comparable methodol- ogy should be applied to other nonpanic anxiety groups (e.g., general anxiety disorder and social phobia), to explore whether similar subtyping is viable in these groups.

In conclusion, using factor mixture analysis, a data- driven case-centered approach that permits dimension- ality, we found a three-class one-factor model to be su- perior to all other models, enabling us to identify three qualitatively distinct panic disorder subtypes that showed significant quantitative differences on anxiety sensitiv- ity, phobic avoidance, psychiatric comorbidity, suicide attempts, and other clinical characteristics. Based on their quantitative and qualitative differences, we pro- pose the following three PD subtypes: a severe cognitive- autonomic subtype (29.8%), a moderate autonomic sub- type (29.9%), and a mild aspecific subtype (40.3%).

Acknowledgments. We would like to thank the re- viewers for their much valued time and comments, and Mrs. Meulenbroek for her linguistic suggestions.

Conflict of interest. The authors report no finan- cial or other relationships relevant to the subject of the article.

REFERENCES 1. American Psychiatric Association. Diagnostic and Statistical Man-

ual of Mental Health Disorders: DSM IV. 4th ed. Washington, DC: American Psychiatric Association; 1994.

2. American Psychiatric Association. Diagnostic and Statistical Man- ual of Mental Health Disorders: DSM 5. 5th ed. Washington, DC: American Psychiatric Association; 2013.

3. Kessler RC, Chiu WT, Demler O, Walters EE. Prevalence, severity and comorbidity of 12-month DSM-IV disorders in the National Comorbidity Survey Replication. Arch Gen Psychiatry 2005;62:617–627.

4. Abelson JL, Khan S, Liberzon I, Young EA. HPA axis activity in patients with panic disorder: review and synthesis of four studies. Depress Anxiety 2007;24:66–76.

5. Clark RC, Galletly CA, Ash DJ, et al. Evidence-based medicine evaluation of electrophysiological studies of the anxiety disorders. Clinical EEG Neurosci 2009;40:84–112.

6. Roy-Burne PR, Craske MG, Stein MB. Panic disorder. Lancet 2006;368:1023–1032.

7. Beck AT, Emery G, Greenberg RL. Anxiety Disorders and Pho- bias: A Cognitive Perspective. New York , NY: Basic books; 1985.

8. Clark DM. A cognitive model of panic. Behav Res Ther 1986;24:461–470.

9. Bandura A. Self-efficacy conception of anxiety. Anxiety Res 1988;1:77–98.

10. Fava L, Morton J. Causal modeling of panic disorder theories. Clin Psychol Rev 2009;29:623–637.

11. Klein DF. False suffocation alarms, spontaneous panics and re- lated conditions. An integrative hypothesis. Arch Gen Psychiatry 1993;50:306–317.

12. Telch JM, Smits JAJ, Brown M, et al. Effect of threat context and cardiac sensitivity on fear responding to a 35% CO2 challenge: a test of the context-sensitivity panic vulnerability model. J Behav Ther Exp Psychiatry 2010;41:365–372.

13. Briggs AC, Stretch DD, Brandon S. Subtyping the panic disorder by symptom profile. Br J Psychiatry 1993;163:201–209.

14. Roberson-Nay R, Kendler KS. Panic disorder and its subtypes: a comprehensive analysis of panic symptom heterogeneity us- ing epidemiological and treatment seeking samples. Psychol Med 2011;41:2411–2421.

Depression and Anxiety

Research Article: Identifying Panic Disorder Subtypes 517

15. Penninx B, Beekman A, Smit J, et al. The Netherlands Study of Depression and Anxiety (NESDA): rationale, objectives and meth- ods. Int J Methods Psychiatr Res 2008;17(3):121–140.

16. World Health Organization (1997). Composite International Diagnostic Interview, Core Version 2.1: Interviewer’s Manual. Sydney, Australia.

17. Beck AT, Epstein N, Brown G, Steer RA. An inventory for mea- suring clinical anxiety: psychometric properties. J Consult Clin Psychol 1988;56:893–897.

18. Beck AT, Steer AR. Manual for the Beck Anxiety Inventory. San Antonio, TX: Psychological Corporation; 1993.

19. Osman A, Hoffman J, Kopper FXBBA, et al. Factor structure, reliability and validity of the Beck anxiety inventory in adolescent psychiatric inpatients. J Clin Psychol 2002;58:443–456.

20. Reiss S, Peterson RA, Gursky DM, McNally RJ. Anxiety sensitiv- ity, anxiety frequency and the prediction of fearfulness. Behav Res Ther 1986;24:1–8.

21. Marks IM, Mathews AM. Brief standard self-rating for phobic patients. Behav Res Ther 1979;17:263–267.

22. Kabacoff RI, Segal DL, Hersen M, van Hasselt VB. Psychomet- ric properties and diagnostic utility of the Beck anxiety inventory and the state-trait anxiety inventory with older adult psychiatric outpatients. J Anxiety Disord 1997;11:33–47.

23. Garin O, Ayuso-Mateos JL, Almansa J, et al. Validation of the “World Health Organization Disability Assessment Schedule, WHODAS-2” in patients with chronic diseases. Health Qual Life Outcomes 2010;8:51:1–15.

24. Pearlin LI, Schooler C. The structure of coping. J Health Social Behav 1978;19:2–21.

25. Wardenaar KJ, van Veen T, Giltay EJ, et al. Development and validation of a short version of the mood and anxiety symp- toms questionnaire: the MASQ-30. Psychiatry Res 2010;179: 101–106.

26. Levine DW, Lewis MA, Bowen DJ, et al. Reliability and validity of the Women’s Health Initiative Insomnia Rating Scale. Psychol Assess 2003;15:137–148.

27. Ainsworth BE, Bassett DR, Strath SJ, et al. Comparison of three methods for measuring the time spent in physical activity. Med Sci Sports Exerc 2000;32:457–464.

28. Lubke GH, Muthen B. Investigating population heterogeneity with factor mixture models. Psychol Methods 2005;10(1):21–39.

29. Muthen LK, Muthen BO. Mplus User’s Guide. 5th ed. Los Angeles, CA: Muthen & Muthen; 2007.

30. Schwarz G. Estimating the dimensions of a model. Ann Statist 1978;6:461–464.

31. Nylund KL, Asparouhov T, Muthen BO. Deciding on the number of classes in latent class analysis and growth mixture modeling: a monte carlo simulation study. Struct Equ Modeling 2007;14:535– 569.

32. Lo Y, Mendell NR, Rubin DB. Testing the number of components in a normal mixture. Biometrika 2001;88:77–778.

33. McLachlan GJ. On bootstrapping the likelihood ratio test statistic for the number of components in a normal mixture. Appl Stat 1978;23:318–324.

34. Meuret AE, White KS, Ritz T, et al. Panic attack symptom di- mensions and their relationship to illness characteristics in panic disorder. J Psychiatr Res 2006;40:520–527.

35. Rassovsky Y, Kusher MG. Carbon dioxide in the study of panic disorder: issues of definition, methodology and outcome. J Anxiety Disord 2003;17:1–32.

36. Rush AJ, First MB, Burns B. Handbook of Psychiatric Measures, 2nd ed., Washington, DC: American Psychiatric Publishing; 2008.

37. Telch JM, Harrington PJ, Smits JAJ, Powers MB. Unexpected arousal, anxiety sensitivity, and their interaction on CO2-induced panic: further evidence for the context-sensitivity vulnerability model. J Anxiety Disord 2011;25:6454–653.

38. De Beurs E, Wilson KA, Chambless DL, Goldstein AJ, Feske U. Convergent and divergent validity of the Beck anxiety inventory for patients with panic disorder and agoraphobia. Depress Anxiety 1997;6(4):140–146.

39. Cox BJ, Cohen E, Direnfeld DM, Swinson RP. Does the Beck Anxiety Inventory measure anything beyong panic attack symp- toms? Behav Res Ther 1996;34(11–12):949–954.

40. Leyger OT, Ruberg HL, Woodruff-Borden J. Examination of the utility of the Beck anxiety inventory and its factors as a screener for anxiety disorders. J Anxiety Disord 2006;20(4): 444–58.

41. Muntigh AD, Van der Feltz-Cornelis CM, Van Marwijk HW, Spinhoven P, Penninx BW, Van Balkom AJ. Is the Beck Anxiety Inventory a good tool to assess the severity of anxiety? A primary care study in the Netherlands Study of Depression and Anxiety (NESDA). BMC Fam Pract 2011;4:12.

Depression and Anxiety

Copyright of Depression & Anxiety (1091-4269) is the property of John Wiley & Sons, Inc. and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.