Symptom structure of PTSD: support for a hierarchical model separating core PTSD symptoms from dysphoria

Background As of yet, no collective agreement has been reached regarding the precise factor structure of posttraumatic stress disorder (PTSD). Several alternative factor-models have been proposed in the last decades. Objective The current study examined the fit of a hierarchical adaptation of the Simms et al. (2002) dysphoria model and compared it to the fit of the PTSD model as depicted in the Diagnostic and Statistical Manual for Mental Disorders, Fourth Edition (DSM-IV), a correlated four-factor emotional numbing, and a correlated four-factor dysphoria model. Methods Data were collected using the Clinician-Administered PTSD Scale in a mixed-trauma sample of treatment-seeking PTSD patients (N=276). Results All examined models provided superior fit to the three-factor model of DSM-IV. The hierarchical four-factor solution provided a better fit than competing models. Conclusion The present study provides empirical support for a conceptualization of PTSD that includes a higher-order PTSD factor that encompasses re-experiencing, arousal, and effortful avoidance sub-factors and a dysphoria factor.

I ndividuals who are exposed to extreme or prolonged stress are at risk of developing mental disorders such as posttraumatic stress disorder (PTSD). In the latest revision of the Diagnostic and Statistical manual for Mental disorders (DSM-IV-TR; American Psychiatric Association, 2000), PTSD is categorized as an anxiety disorder. Symptoms are organized into three different clusters: re-experiencing (B criteria), avoidance and numbing (C criteria), and hyperarousal (D criteria). However, the nosologic validity of PTSD, including the diagnostic criteria as formulated in DSM, has raised considerable debate (Spitzer, First, & Wakefield, 2007). Issues concerning the diagnosis include (but are not limited to) marked heterogeneity in symptom patterns across individuals diagnosed with PTSD (Foa, Riggs, & Gershuny, 1995), as well as high comorbidity and symptom overlap with other mental disorders (Deering, Glover, Ready, Eddleman, & Alarcon, 1996;Spitzer et al., 2007). Moreover, attempts to provide empirical support for specific pathogenic mechanisms underlying PTSD have yielded equivocal results at best (Rosen & Lilienfeld, 2008).
The marked comorbidity of PTSD with other disorders and the lack of compelling evidence for a specific pathogenesis and etiology of PTSD may imply that there are, in fact, multiple psychopathological dimensions underlying the PTSD construct. Each of these underlying dimensions is potentially related to distinct pathogenic processes. Structural studies may help to identify the underlying (latent) components and to determine which of these are specific to PTSD, versus those that would explain existing comorbidity with other (mood and anxiety) disorders (Simms, Watson, & Doebbeling, 2002). For instance, it has previously been proposed that PTSD shares a broad, non-specific factor with ''distress'' disorders (major depressive disorder [MDD], generalized anxiety disorder [GAD], dysthymic disorder [DD]), as well as with other emotional/internalizing disorders (i.e., fear disorders and bipolar disorders), but that it is distinguishable from these disorders by one or more specific (lower-level) components (Watson, 2005). Identification of shared versus specific components of PTSD would facilitate research into the (neurobiological) basis of PTSD symptoms. Moreover, identification of psychopathological constructs underlying PTSD could aid in diagnosis and effective treatment.
The King et al. ''emotional numbing'' model contains four correlated factors: re-experiencing, (effortful) avoidance, emotional numbing, and arousal. In this model symptoms from the avoidance cluster (C) in DSM-IV are split into two different factors: active avoidance and emotional numbing. The avoidance factor consists of item C1 and C2. The emotional numbing factor includes items C3ÁC7 (see Table 2 for a legend of the symptom codes and for item allocation across models). The Simms et al. correlated four-factor ''dysphoria'' model contains the factors: re-experiencing, avoidance, dysphoria, and arousal. The re-experiencing factor is identical to the reexperiencing cluster in DSM-IV and contains all B cluster items. The avoidance factor is identical to the active avoidance factor in King et al. emotional numbing model, containing items C1 and C2. The dysphoria factor comprises items C3 through C7 as well as D1 through D3. The arousal factor comprises items D4 and D5. Recently, Elhai and colleagues (2011) proposed a model that places items D1 through D3 in a separate factor, and there is some evidence for enhanced model fit of this this five-factor hybrid of the numbing and dysphoria model for PTSD (Pietrzak, Tsai, Harpaz-Rotem, Whealin, & Southwick, 2012;Wang et al., 2011;Wang, Elhai, Dai, & Yao, 2012), as well as for acute stress disorder (Hansen, Armour, & Elklit, 2012).
As noted before, both the emotional numbing and dysphoria models differentiate between active or effortful avoidance and emotional numbing (Asmundson, Stapleton, & Taylor, 2004;Foa et al., 1995). King and colleagues constructed a separate factor for emotional numbing symptoms, whereas Simms and colleagues clustered the numbing symptoms under a broad dysphoria factor. The inclusion of this broad dysphoria factor in the Simms et al. model is theoretically appealing because it differentiates between specific and non-specific symptoms (King et al., 2006). Although Simms and colleagues did not test the fit of a hierarchical model that separates this non-specific factor from a higher-order PTSD factor that incorporates three sub-clusters (i.e., reexperiencing, effortful avoidance, arousal), the authors demonstrated that the dysphoria factor was highly correlated with indices of anxiety and depressive symptoms. This suggests that the dysphoria factor may reflect symptoms of general distress that PTSD has in common with other ''distress'' disorders (Clark, Watson, & Mineka, 1994). Indeed, using CFA Grant and colleagues demonstrated that PTSD could be distinguished from MDD and GAD and that intrusions, avoidance, and hypervigilance reflected lower-level PTSD symptom clusters,whereas dysphoria was best conceptualized as a higher-order factor common in PTSD, MDD, and GAD (Grant, Beck, Marques, Palyo, & Clapp, 2008). It should be noted, however, that a study by Marshall and colleagues demonstrated that this pattern of association is only present at the factor level. At the item level, correlations between the eight dysphoria items and an external measure of general distress are equally strong as the correlations of the PTSD items with measures of general distress (Marshall, Schell, & Miles, 2010).
So far, we have not come across studies that effectively separated PTSD-specific clusters from factors that may be associated with other psychiatric disorders in hierarchical models. Rather, studies that examined the fit of hierarchical models for PTSD either subsumed all factors under one higher-order PTSD factor or clustered avoidance and re-experiencing versus arousal and numbing symptoms in separate factors. On the whole, these studies did not provide much empirical support for the utility of hierarchical models to describe the symptom structure of PTSD (Asmundson et al., 2000;King et al., 1998;Taylor et al., 1998;Yufik & Simms, 2010). Therefore, the present study was aimed at examining the statistical validity of a hierarchical adaptation of the dysphoria model with clusters of re-experiencing, arousal, and effortful avoidance symptoms subsumed under a higher-order PTSD factor versus a dysphoria factor (see Fig. 1). To do so we tested the fit of this PTSD model on Clinician-Administered PTSD Scale (CAPS; Blake et al., 1995) data that were obtained from a heterogeneous trauma sample and compared it to the fit of alternative factor solutions.

Methods
Participants & procedure Data were acquired from treatment-seeking PTSD patients from military (n0115) and civilian (n0161) treatment facilities. Sites included the department of Military Mental Healthcare in Utrecht, the Netherlands, the Overwaal Centre for Anxiety Disorders in Lent, the Netherlands, and the Hendriks & Rooseboom Psychiatric and Psychotherapeutic Centre in Arnhem, the Netherlands. Data were collected from 2002 till 2011. Data were included from all patients who consented to have their data used for research. The CAPS was administered during the initial intake prior to the start of treatment. Participants were included in the study if they met DSM-IV criteria for PTSD, with a CAPS total score of ]45 on the CAPS (Blake et al., 1995), which is indicative of (at least) moderate PTSD symptom severity according to the Dutch manual (Hovens, Luinge, & Van Minnen, 2005).
A total of 276 participants (48% female) were included in the analysis. Trauma exposure was assessed with the Life Event Checklist (LEC; Gray, Litz, Hsu, & Lombardo, 2004), which is routinely administered as part of the CAPS interview. Participants identified their index traumatic event as their ''worst or most upsetting'' traumatic event. Index traumatic events are reported in Table 1. The mean age of the sample was 36.3 years (SD 09.96, range 16Á63). A total of 27% of the participants completed university or college or had at least some years of college/university education, 44.2% completed high school or had at least some year high school education, and 7.2% only completed elementary school. Mean elapsed time (years) since the traumatic experience (index event) was 11.9 (SD 011.04, range 0Á54). The mean CAPS score was 73.3 (SD 016.78, range 45Á125).

Measures
PTSD symptoms were assessed with the Dutch translation of the CAPS (Blake et al., 1995;Hovens et al., 2005). The CAPS is a 30-item structured interview that corresponds to DSM-IV criteria for PTSD. It contains 17 items that correspond to B, C, and D criteria for which frequency and intensity are rated on a five-point scale (0Á4). Total PTSD severity is calculated as the sum of 17 symptom frequency and intensity scores. The CAPS has been shown to have excellent psychometric properties in a variety of populations (Weathers, Keane, & Davidson, 2001). The psychometric properties of the Dutch translation have been reported to be satisfactory (Hovens et al., 1994;Hovens et al., 2005).

Statistical analyses
Data were multivariate non-normal distributed (Mardia's kurtosis coefficient 019.17; c.r. 06.173); therefore, CFA was conducted using maximum likelihood estimation with robust standard errors (MLR) in M-Plus v. 6.11 (Muthén & Muthén, 2011). MLR is robust to nonnormality. It yields factor loadings identical to maximum likelihood estimation but adjusts the Chi-square and associated fit-indices and standard errors (Yuan &   Bentler , 2000). Missing values were present for 3.8% of CAPS items. The maximum number of missing values for any individual was 2 out of 17 items, and the amount of missing values constituted just 0.002% of the total values. MLR estimation assumes missingness at random (MAR). Hereby, individuals with missing data are not excluded list-wise from the analyses, but all available data in the observed information matrix is used for parameter estimation, leading to unbiased parameter estimates.
Overall model fit was measured using the MLR x 2 statistic, which is equivalent to the Yuan-Bentler T*2 statistic, Root-Mean-Square-Error of Approximation (RMSEA), Standardized Root-Mean-Square Residual (SRMR), Comparative Fit Index (CFI), Tucker-Lewis Index (TLI), Akaike's Information Criterion (AIC), and the Bayesian Information Criterion (BIC). Cut-off scores of 0.95 for CFI and TLI, B0.07 for RMSEA, and B0.08 for SRMR suggest good model fit (Byrne, 1989(Byrne, , 1991Hu & Bentler, 1999;Steiger, 2007). The AIC and BIC are used to compare model fit of models with different amounts of parameters. Both indices represent a trade-off between model accuracy and model complexity (the BIC also corrects for the number of parameters). Lower values represent better fit. Additionally, when a model has a BIC value of 10 points less than another model, the odds that the model with the smaller BIC value is a better-fitting model would be 150:1, very strong evidence according to Raftery (1995).
Models that were investigated included: the current DSM-IV three-factor model (Model 1); the King et al. correlated four-factor emotional numbing model (King et al., 1998) Table 2 displays the item allocation across models. Table 3 shows fit-indices of the models that were evaluated. CFI and TLI values were below cut-off in all models. All models provided better fit than the DSM-IV reference model (1). Variations of the dysphoria model slightly outperformed the numbing model. Additionally, although differences in fit-indices across models were small, the 10 point difference in BIC value between Simms et al. correlated (3a) and hierarchical dysphoria model (3b) should be considered as positive evidence that the latter is a better-fitting model (Raftery, 1995). Moreover, a corrected Chi-square difference test for nested models was non-significant [S-B scaled x 2 -diff (2, N0276) 00.64, p00.73], suggesting that the more parsimonious nested hierarchical model (3b) should be retained over the inter-correlated four-factor dysphoria model (3a). As can be seen in Table 4, standardized factor loadings of CAPS items vary modestly across models. Items with poor factor loadings across models included C3, inability to recall aspects of trauma, and D1, sleep disturbances. However, removing these items from the analyses did not substantially improve model fit.

Discussion
The present study examined the latent structure of PTSD symptoms in a mixed-trauma sample. Although the differences between fit-indices were small, the hierarchical dysphoria model fitted the data best. The proposed model that included clusters of arousal, intrusion, and effortful avoidance symptoms subsumed under a higher-order PTSD factor, and a separate dysphoria factor fitted the data (marginally) better than the original dysphoria model. Superior fit was also confirmed by a Chi-square difference test. The focus on a mixed-trauma sample in the present study suggests that the reported factor structure is fairly robust to inter-subject variation in type of trauma.   performed better-albeit modestly-than the King et al. numbing model (King et al., 1998).  Item C3 (trauma-related amnesia) performed poorly across models. Therefore, these findings lend empirical support to McNally's (2009) proposal to remove the item from the diagnostic criteria for PTSD. Although removal of item C3 did not substantially improve model fit in the present study, other studies provide additional support for the removal of item C3 (Grant et al., 2008;McWilliams, Cox, & Asmundson, 2005;Olff et al., 2009). Item D1 (sleep problems) also displayed poor factor loadings across models. Again, removing the item did not improve model fit in the present study. This finding may be taken to imply that this item taps on a separate latent variable representing sleeping problems (King et al., 2009;Morina, et al., 2010), although we did not test the fit of such a model in the present study.
The present study underscores the validity of a demarcation between specific and non-specific (i.e., dysphoria) PTSD symptoms. This view is further supported by the proposed reformulation of PTSD by Brewin and colleagues (Brewin, Lanius, Novac, Schnyder, & Galea, 2009). Separating PTSD from dysphoria symptoms provides the possibility to examine the etiology and pathogenesis of PTSD within an integrative hierarchical model for mood and anxiety disorders (Mineka, Watson, & Clark, 1998;Simms et al., 2002;see also, Watson, 2005), which postulates that all mood and anxiety disorders share a higher-order component of general distress, versus lowerlevel factors underlying specific (clusters of) mood and anxiety disorders. PTSD is proposed to share a broad, non-specific factor with other ''distress'' and emotional/ internalizing disorders but is expected to be distinguishable by one or more specific lower-level components (Watson, 2005). Accordingly, a recent meta-analysis demonstrated that dysphoria is more strongly correlated to indices of depression, anxiety, panic, and substance abuse than the intrusions, avoidance, and hypervigilance factors (Gootzeit & Markon, 2011). Grant and colleagues (2008) observed that PTSD, MDD, and GAD were highly correlated disorders in a sample of MVA victims and that a higher-order dysphoria factor accounted for these correlations. Also, Gros and colleagues (2010) demonstrated that emotional numbing and dysphoria symptoms increased the likelihood of meeting diagnostic criteria for (comorbid) MDD in PTSD patients. Moreover, a recent study by Wolf et al. (2010) confirmed that PTSD shares a common genetic risk factor with MDD, GAD, panic disorder, and DD. It should be noted, however, that the dysphoria factor is not uniquely associated with external measures of general distress. A recent study by Armour and colleagues demonstrated that a significant amount of variance in the dysphoria factor is accounted for by MDD and GAD, but that the variance in intrusive symptoms was more strongly affected by statistically controlling for the presence for GAD (after controlling for MDD) than the variance in dysphoria items (Armour, McBride, Shevlin, & Adamson, 2011). Marshall and colleagues (2010) confirmed that the dysphoria factor is most strongly related to external measures of depressive symptoms and anxiety. More importantly, however, they observed that at the item level, correlations of general distress with external measures were not stronger for the dysphoria items than for the PTSD items in Simms et al. model. To account for their findings, the authors posited that the dysphoria factor might actually tap on impaired functioning (analogous to Criterion F in DSM-IV) as caused by the remaining PTSD items. An alternative explanation is that the items in the dysphoria cluster actually tap on aspects of (trait) neuroticism/negative affectivity (N/NA). N/NA incorporates negative emotional states such as fear, anger, sadness, guilt, and disgust (Clark et al., 1994;Watson, 2005, p. 525). Although N/NA is not specific to mood and anxiety disorders (Mineka et al., 1998), prospective research has demonstrated that N/NA composes a particularly salient feature in the etiology of distress disorders like PTSD (e.g., Parslow, Jorm, & Christensen, 2006;Rademaker, van Zuiden, Vermetten, & Geuze, 2011) and MDD (Christensen & Kessing, 2006).
The relevance of effectively differentiating between PTSD symptoms and general distress becomes all the more apparent when we examine the proposed revisions for PTSD in DSM-V. The proposed revision indicates a four-factor model that includes re-experiencing, active avoidance, negative mood/cognition, and hyperarousal factors. Although the specification of a separate active avoidance cluster corresponds to the extant body of CFA results, the negative emotionality and hyperarousal cluster closely resemble the C and D criteria in DSM-IV. Therefore, it seems likely that the proposed diagnostic criteria for PTSD in DSM-V will continue to be a source for debate. More importantly, however, the inclusion of items in DSM-V to assess ''persistent blame of self and others'' and a ''pervasive negative emotional state'', which appear to tap on (trait) N/NA, can be expected to further increase the overlap of PTSD with other distress disorders. If so, this would inadvertently and somewhat paradoxically strengthen claims for a need to change the nosologic system of DSM (e.g., Maser et al., 2009;Watson, 2005). Marshall et al. (2010) proposed that the dysphoria factor might best be relabeled to reflect that it represents a multifacetted cluster of items associated with impaired functioning (e.g., sleeping problems, difficulty concentrating, reduced affect, loss of interest). Indeed, recent data suggest that the dysphoria factor may comprise multiple aspects of impaired functioning. Studies have demonstrated that the dysphoria factor includes items that can be subsumed under (sub) factors for emotional numbing and sleep disturbances (King et al., 2009;Morina et al., 2010). Additionally, the dysphoria factor contains items pertaining to anhedonia and irritability, and there is evidence that both anhedonia and emotional numbing represent distinct pathogenic dimensions (Kashdan, Elhai, & Frueh, 2006). Moreover, since hostility has been reported to be a key feature in PTSD patients (Orth & Wieland, 2006), which is not specific to PTSD but occurs in various distress disorders (Moscovitch, McCabe, Antony, Rocca, & Swinson, 2008;Perlis et al., 2009), hostility could be a distinct sub-factor underlying the general distress/dysphoria cluster. Therefore, ''lumping together'' different symptom clusters into one dysphoria/general distress factor is inconsistent with evidence supporting the distinctness of these clusters (Palmieri, Weathers, Difede, & King, 2007, p. 339), and additional research into the dimensions underlying the dysphoria construct is recommended.
The results of the present study have to be viewed in light of a number of limitations. First, the observed fit-indices were inconsistent. Whereas the RMSEA suggested good fit (B0.07) in all models (except the DSM-IV model), the TLI and CFI values indicated poor model fit. The CFI values were lower in the present study than in other studies that used the CAPS (Blake et al., 1995) to assess PTSD symptoms (Buckley et al., 1998;Palmieri et al., 2007). The CFI is affected by the average size of correlations in the data (Kenny, 2012), which may have been lower in the present sample as compared to results from previous studies in more homogenous samples. 1 For instance, whereas previous studies examined fairly homogenous trauma samples like female victims of domestic violence , earthquake survivors (Wang et al., 2011;Wang et al., 2012), or military veterans (King et al., 1998), the present study included a range of trauma types including sexual trauma, combat, and interpersonal violence. Nevertheless, the CFI, TLI, and BIC values of the hierarchical adaptation of the Simms et al. dysphoria model were superior to those of competing models.
A second limitation pertains to the fact that the present sample consisted of treatment-seeking individuals diagnosed with PTSD only. Therefore, additional research is needed to determine whether results can be generalized to community-dwelling trauma-exposed individuals. Relatedly, it is possible that selecting a sample of individuals diagnosed with PTSD exclusively, rather than a sample of individuals meeting DSM-IV criterion A (and not necessarily meeting all diagnostic criteria for PTSD), as was the case in most previous studies, may have biased the results. Indeed, a factorial invariance study in US veterans suggests that although model fit may be equal across groups (PTSD vs. criterion A), there may exists (subtle) differences between (treatment-seeking) PTSD patients and trauma-controls in the way they perceive and respond to PTSD questionnaires (McDonald, Beckham, Morey, & Calhoun, 2009).

Conclusion
Uncovering the latent structure of PTSD is important in identifying distinct aspects of PTSD symptomatology and sources of comorbidity with order disorders. The present study demonstrates once more that the threefactor solution in DSM-IV-TR is untenable and that a model that separates core PTSD symptoms from items associated with general distress or impaired functioning, which are shared with other psychiatric disorders, provides better fit to the data. Future studies examining the (neurobiological) correlates of the factors described in this and previous studies, as well as studies into dimensions that may underlie the general distress/dysphoria factor, would help to further our understanding of the pathogenesis of PTSD and other distress disorders.