Latent Factor Structure of DSM-5 Posttraumatic Stress Disorder: Evaluation of Method Variance and Construct Validity of Novel Symptom Clusters

Abstract

The Diagnostic and Statistical Manual of Mental Disorders, fifth edition (DSM-5) four-factor model of posttraumatic stress disorder (PTSD) has demonstrated adequate fit in several confirmatory factor analysis (CFA) studies. Although several alternative measurement models have demonstrated better fit, there is no consensus yet on the best model, and newly proposed models lack sufficient construct validation. Notably, these studies have relied exclusively on questionnaire data, and thus their findings may be attributable to a method effect. This study examined the factor structure of DSM-5 PTSD symptoms using both questionnaire and interview data to determine the impact of assessment method on factor structure and construct validity of alternative model symptom clusters. Participants (N = 380) were veterans who completed the PTSD Checklist for DSM-5 (PCL-5; Weathers et al., 2013) and Clinician-Administered PTSD Scale for DSM-5 (CAPS-5; Weathers et al., 2013). Fit was similar across models. However, the seven-factor Hybrid model (Armour et al., 2015) fit best. Limited evidence of a method effect was observed. Results of construct validity analyses were mixed; some of the newly proposed symptom clusters demonstrated hypothesized differential associations with external correlates, but others did not. These findings suggest that results of previous DSM-5 PTSD CFAs supporting the Hybrid model are not attributable to a method effect. However, observed limited difference in model fit and mixed construct validity evidence raise concerns regarding the value of parsing DSM-5 symptom clusters. Constructs implied by the new factors in the more complex measurement models of PTSD require greater explication and construct validation.

Over more than 20 years, a substantial number of studies have aimed to identify the latent factor structure of posttraumatic stress disorder (PTSD; see Armour, Műllerová, & Elhai, 2016 for a review). Although often conceptualized as a univariate construct, factor analytic evidence indicates PTSD symptoms represent a constellation of at least four underling constructs. Determining the factor structure of PTSD has critical implications for informing diagnostic criteria, establishing clinically relevant subtypes, identifying etiological factors, and predicting differential response to treatment.

The fifth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5; American Psychiatric Association [APA], 2013) implies a four-factor structure of PTSD by organizing the 20 core symptoms into four symptom clusters: intrusions (INT), avoidance (AVD), negative alterations in cognition and mood (NCM), and alterations in arousal and reactivity (AAR). Confirmatory factor analytic (CFA) studies have found that this four-factor model provides an adequate representation of symptom covariance across a variety of samples and trauma types (see Armour et al., 2016 for a review). However, several alternative models have been proposed (see Table 1).

Table 1.

Item Mapping for Measurement Models of DSM-5 PTSD Symptoms

PTSD Symptom	DSM-5	Dysphoria	Dysphoric Arousal	Externalizing	Anhedonia	Hybrid
(B1) Intrusive memories	INT	INT	INT	INT	INT	INT
(B2) Nightmares	INT	INT	INT	INT	INT	INT
(B3) Flashbacks	INT	INT	INT	INT	INT	INT
(B4) Emotional reactivity	INT	INT	INT	INT	INT	INT
(B5) Physiological reactivity	INT	INT	INT	INT	INT	INT
(C1) Avoidance of thoughts	AVD	AVD	AVD	AVD	AVD	AVD
(C2) Avoidance of reminders	AVD	AVD	AVD	AVD	AVD	AVD
(D1) Trauma-related amnesia	NCM	DYS	NCM	NCM	NAF	NAF
(D2) Negative beliefs	NCM	DYS	NCM	NCM	NAF	NAF
(D3) Blame	NCM	DYS	NCM	NCM	NAF	NAF
(D4) Negative feelings	NCM	DYS	NCM	NCM	NAF	NAF
(D5) Loss of interest	NCM	DYS	NCM	NCM	ANH	ANH
(D6) Feeling detached	NCM	DYS	NCM	NCM	ANH	ANH
(D7) Difficulty experiencing positive emotions	NCM	DYS	NCM	NCM	ANH	ANH
(E1) Irritability	AAR	DYS	DAR	EXT	DAR	EXT
(E2) Risk taking	AAR	AAR	DAR	EXT	DAR	EXT
(E3) Hypervigilance	AAR	AAR	AXA	AXA	AXA	AXA
(E4) Startle	AAR	AAR	AXA	AXA	AXA	AXA
(E5) Difficulty concentrating	AAR	DYS	DAR	DAR	DAR	DAR
(E6) Sleep disturbance	AAR	DYS	DAR	DAR	DAR	DAR

Note. AAR = alterations in arousal and reactivity; ANH = anhedonia; AVD = avoidance; AXA = anxious arousal; DAR = dysphoric arousal; DYS = dysphoria; EXT = externalizing; INT = intrusions; NAF = negative affect; NCM = negative alterations in cognition and mood.

Two of the alternative models were originally proposed in reference to DSM-IV (APA, 2000) PTSD symptoms and were later modified to reflect DSM-5 PTSD symptoms. First, Simms, Watson, and Doebbeling (2002) proposed a measurement model for DSM-IV PTSD symptoms which considered hypervigilance and exaggerated startle response to be indicators of hyperarousal, and considered the remaining nonspecific symptoms common to other anxiety and mood disorders (e.g., sleep disturbance, irritability, concentration disturbance) to be indicators of a novel dysphoria factor. This four-factor Dysphoria model has since been updated to reflect DSM-5 symptoms and compared with other DSM-5 models (see Armour et al., 2016). Second, Elhai and colleagues (2011) proposed dividing the DSM-IV hyperarousal symptom cluster into anxious arousal (AXA) and dysphoric arousal (DAR) clusters to distinguish between PTSD symptoms reflective of fear-based and distress-based disorders, respectively (Watson, 2005). This five-factor Dysphoric Arousal model has since been modified to reflect DSM-5 symptoms (see Armour et al., 2015).

Additionally, several measurement models have been developed since the release of DSM-5 which proposed alternative organization of the 20 DSM-5 PTSD symptoms. Tsai and colleagues (2015) proposed a six-factor Externalizing model which further divides the DSM-5 AAR cluster by separating Elhai’s proposed DAR cluster into non-specific DAR symptoms (i.e., concentration and sleep disturbance) and a new externalizing behavior cluster (EXT; i.e., irritability, risk taking), reflecting difficulty regulating emotion and impulsive behavior. Drawing on Watson’s program of research highlighting differences between positive and negative affect (e.g., Watson, Clark, & Stasik, 2011), Liu and colleagues (2014) proposed a six-factor Anhedonia model which divides the DSM-5 NCM cluster into symptoms reflecting increased negative affect (NAF; i.e., negative trauma-related beliefs, blame, negative trauma-related feelings) and those reflecting anhedonia (ANH; i.e., detachment, loss of interest, difficulty experiencing positive feelings). Finally, Armour and colleagues (2015) proposed a seven-factor Hybrid model that incorporates all of the symptom cluster divisions proposed by the Dysphoric Arousal, Externalizing, and Anhedonia models, and represents the most complex and, thus, least parsimonious model proposed to date.

Multiple CFA studies have found that several alternative models---including the six-factor Externalizing (Tsai et al., 2014), six-factor Anhedonia (Liu et al., 2014), and seven-factor Hybrid (Armour et al., 2015) models---provide superior fit relative to the DSM-5 model (Armour et al, 2016). Studies comparing the six- and seven-factor models have produced mixed findings regarding which of these models provides the best fit (e.g., Armour et al., 2015; Blevins et al., 2015; Bovin et al., 2015).

Although these less parsimonious DSM-5 PTSD models have provided superior fit relative to the DSM-5 model in several studies, this research has relied exclusively on questionnaire data. Beyond the usual limitations inherent in the use of questionnaires (e.g., Miller & Kozak, 1993; Nisbett & Wilson, 1977), this exclusive reliance on a single assessment modality leaves open the possibility that these results are at least partially attributable to a method effect. Palmieri, Weathers, Difede, and King (2007) examined the extent to which CFA findings may be attributable to a method effect by analyzing data from DSM-IV (APA, 2000) versions of the PTSD Checklist (PCL; Weathers, Litz, Herman, Huska, & Keane, 1993) and Clinician-Administered PTSD Scale (CAPS; Blake et al., 1995). Palmieri et al. found that items on the PCL (a questionnaire) and CAPS (a structured interview) represented the same underlying factors and that there was little evidence of a method effect. To date, no study has examined the degree to which PTSD CFA findings may be confounded by methodology using DSM-5 PTSD symptom measures. Beyond replication and extension, however, there is an urgent need to directly examine method variance in DSM-5 PTSD assessment and not simply assume that DSM-IV results would pertain. Although there is clear continuity in the PTSD criteria from DSM-III-R forward, there are substantial qualitative differences between DSM-5 and DSM-IV criteria (see Friedman, 2013; Hoge, Riviere, Wilk, Herrell, & Weathers, 2014).

Conducting CFA with data from multiple assessment modalities permits detection of a possible method effect by examining several sources of evidence (Brown, 2006; Byrne, 2012). The first source of evidence comes from a model in which only method factors are estimated. If method accounts for little variance, this model should provide poor fit to the data. The second source comes from a model in which both symptom cluster factors and method factors are estimated. If items are better indicators of their respective PTSD symptom clusters than they are of their respective assessment modality, then symptom cluster factor loadings should be higher than method factor loadings. The third source comes from a comparison of a model in which only symptom cluster factors are estimated to a model in which both symptom cluster and method factors are estimated. If a method effect is minimal or absent, method factors should not account for significant item variance beyond variance explained by symptom cluster factors.

In addition to exclusive reliance on questionnaire data, existing research on the newly proposed alternative models of DSM-5 PTSD symptoms lacks sufficient construct validity evidence. As others have noted (e.g., Armour et al., 2016; Pietrzak et al., 2015), compared with the split of avoidance and numbing which expanded the implicit three-factor DSM-IV model to the four-factor DSM-5 model, the conceptual basis for additional splitting of symptom clusters in the alternative models of PTSD symptoms is less compelling. The resulting new factors were identified empirically, the conceptual justifications are largely post hoc, and there is little validity evidence apart from model fit to support the claim that these additional factors represent distinct constructs. Thus, although these less parsimonious models provide better model fit, the added conceptual value of distinctions among the new factors remains unclear. To our knowledge, only one study has examined the degree to which the factors emerging from alternative models demonstrate differential associations with relevant external correlates (Pietrzak et al., 2015). Specifically, Pietrzak and colleagues (2015) examined the association between self-report Hybrid model symptom clusters and self-report screening measures of depression, anxiety, suicidal ideation, hostility, functional impairment, and quality of life among a veteran sample. The authors found that EXT symptoms were more strongly associated with hostility than other symptom clusters. However, several other more complicated associations emerged, including ANH and DAR clusters evidencing similar associations with depression, and DAR emerging as more strongly associated with GAD symptoms than AXA. Of note, in addition to reliance on self-report screening measures, only 5.2% of the sample screened positive for PTSD symptoms, potentially limiting generalizability of the findings to more severe clinical samples. Additional construct validity evidence is essential to determine if, beyond providing better fit, these alternative models provide meaningful conceptual representations of PTSD symptoms.

In this study, we examined the latent factor structure of PTSD symptoms among veterans using data from the DSM-5 versions of the PCL (PCL-5; Weathers, Litz, et al., 2013) and CAPS (CAPS-5; Weathers, Blake, et al. 2013). Competing measurement models were examined while estimating both symptom cluster and assessment method factors. We hypothesized that the four-factor DSM-5 model would adequately fit the data, but that previously identified six- and seven-factor alternative models would provide superior fit while covarying for method. Additionally, based on the results of Palmieri et al. (2007) with DSM-IV PTSD criteria, we predicted little evidence of a method effect while examining the latent factor structure of PTSD. Specifically, we hypothesized that (a) a measurement model in which only method factors are estimated would not fit the data well; (b) when both symptom cluster factors and method factors are estimated, items would have significantly stronger loadings to symptom cluster factors than method factors; and (c) method factors would explain little item variance beyond that explained by symptom cluster factors.

Last, we examined the degree to which symptom clusters included in alternative models of PTSD differed in their associations with measures of relevant external correlates: depression, anxiety, and aggressive driving. Regarding the proposed division of the NCM symptom cluster into NAF and ANH clusters, we hypothesized that NAF symptoms would be more strongly related to cognitive depressive symptoms (e.g., pessimism, punishment feelings, worthlessness) and that ANH symptoms would be more strongly related to affective depressive symptoms (e.g., loss of pleasure, loss of interest). Regarding the proposed division of the AAR symptom clusters into EXT, AXA, and DAR clusters, we hypothesized EXT symptoms would be more strongly related to a measure of aggressive driving, AXA would be more strongly related to anxiety symptoms, and DAR would be more strongly related to somatic depressive symptoms (e.g., sleep disturbance, difficulty concentrating).

Method

Participants and Procedure

Participants from two studies were included in analyses. The first sample consisted of 167 veterans recruited for a study designed to validate the PCL-5 and CAPS-5 (Bovin et al., 2016; Weathers et al., in press). Of these participants, 15 (4.01%) were excluded because they completed the PCL-5 and CAPS-5 using different index events and 2 (1.20%) were excluded as they did not complete either the PCL-5 or the CAPS-5. The second sample included 230 veterans who completed the baseline assessment of an ongoing randomized controlled trial (see Sloan, Unger, & Beck, 2016 for a detailed overview of study procedures). All study procedures were approved by the VA Boston Healthcare System Institutional Review Board (protocol numbers 2625 and 2650).

The final sample (N = 380) was predominantly male (n = 359; 94.47%). The majority of participants identified as White (n = 261; 68.68%). The remaining participants identified as African American/Black (n = 83; 21.84%), American Indian/Alaskan Native (n = 5; 1.32%), or other (n = 21; 5.53%). Eight participants (2.11%) did not report their race. Mean age was 54.96 (SD = 11.98) years. Portions of these samples were used to examine the latent factor structure of the PCL-5 (Bovin et al., 2016) and the CAPS-5 (Weathers et al., in press) separately.

Measures

The PCL-5 is a 20-item questionnaire of DSM-5 PTSD symptoms. Respondents rate how much they were bothered by each symptom in the past month on a five-point scale ranging from 0 = Not at all to 4 = Extremely. PCL-5 scores possess strong reliability and validity in veteran and undergraduate samples (Blevins et al., 2015; Bovin et al., 2015).

The CAPS-5 is a structured diagnostic interview for DSM-5 PTSD. For each symptom interviewers first assess the intensity and frequency over the past month and then combine that information into a symptom severity rating using a five-point scale ranging from 0 = Absent to 4 = Extreme/incapacitating. CAPS-5 scores also possess strong reliability and validity in veterans (Weathers et al., in press). The CAPS-5 was administered by masters- and doctoral-level interviewers who had previous experience with PTSD diagnostic interviews. Training for interviewers, conducted by experts affiliated with the National Center for PTSD, consisted of attending a didactic lecture of administration and scoring procedures, rating mock interviews, and role playing CAPS-5 administration. All interviewers were required to demonstrate competency in administration and scoring prior to conducting study interviews. After study administration began, interviewers participated in regular calibration meetings in which standard administration and scoring procedures were discussed. In addition, audio recordings of interviews were intermittently reviewed to ensure administration and scoring procedures were being followed. Strong interrater reliability was observed in both samples (PTSD diagnosis kappa = .78, CAPS-5 total score intraclass correlation = .91; see Weathers et al., in press for additional details).

The Beck Depression Inventory, second edition (BDI-II; Beck, Steer, & Brown, 1996) is a 21-item self-report measure of unipolar depressive symptoms during the past two weeks. Respondents rate each symptom on a four-point Likert scale from 0 to 3; higher scores indicate greater symptom severity. Psychometric properties of the BDI-II have been well-established in several populations (e.g., Beck et al., 1996; Dozois, Dobson, & Ahnberg, 1998). Although several measurement models for the BDI-II have been proposed, models distinguishing between cognitive (e.g., pessimism, punishment feelings, worthlessness), affective (e.g., loss of pleasure, loss of interest), and somatic (e.g., sleep disturbance, difficulty concentrating) symptoms have garnered the strongest empirical support, with conflicting findings regarding the utility of distinguishing between affective and somatic symptoms (see Huang & Chen, 2015, Vanheule, Desmet, Groenvynck, Rosseel, & Fontaine, 2008). We compared the two-factor model Huang and colleagues (2015) identified as fitting best in their meta-analysis to the three-factor model identified by Vanheule and colleagues (2008). Although the two-factor model provided adequate fit to the data (χ² = 369.10, df = 169, p<.001, Root Mean Square Error of Approximation [RMSEA] = .06 [90%CI = .06-.08], Bentler Comparative Fit Index [CFI] = .96, Tucker-Lewis Index [TLI] = .96), the three-factor model provided much better fit to the data (χ² = 127.47, df= 87, p =.003, RMSEA = .03 [90%CI = .03-.06], CFI = .99, TLI = .98). Accordingly, we elected to retain the 3-factor model. The six-item cognitive and three-item affective depression subscales were selected as the external correlates because these symptom groups measure the increased negative affect and diminished positive affect reflected in the NAF and ANH symptom clusters, respectively (Liu et al, 2014). The six-item somatic depression symptom subscale was used as the external correlate of interest for the DAR symptom cluster because this group of symptoms measure the general restless agitation reflected in this symptom cluster (Elhai et al., 2011; Simms et al., 2002).

The Beck Anxiety Inventory (BAI; Beck, Epstein, Brown, & Steer, 1988) is a 21-item self-report measure of cognitive and somatic anxiety symptoms. Respondents rate the degree to which they have been bothered by each symptom four-point Likert scale from Not at all to Severely − I could barely stand it; higher scores indicate greater symptom severity. Similar to the BDI-II, psychometric properties of the BAI are well established (e.g., Fydrich, Dowdall, & Chambless, 1992; Steer, Ranieri, Beck, & Clark, 1993). The nine-item cognitive and twelve-item somatic BAI subscales were used as the external correlates of interest for the AXA subscale because these symptoms measure the general anxious arousal reflected in the AXA cluster (Simms et al., 2002). Internal consistency was adequate for the cognitive and somatic subscales in the current sample (α = .87 for both scales).

The Driving Behavior Survey (DBS; Clapp et al., 2011) is a 21-item self-report measure of anxious driving. Respondents report the degree to which they generally engage in a list of behaviors on a seven-point Likert scale ranging from never to always. The DBS comprises three subscales: hostile and aggressive driving behaviors (HAB), anxiety-based performance deficits, and exaggerated safety and caution. For this study, only the seven HAB subscale items were included as indicators of externalizing behavior; higher scores indicate more aggressive driving behavior. DBS-HAB scores have strong psychometric properties in multiple samples, including strong factor analytic, test-retest, and construct validity (e.g., Clapp et al., 2011; Clapp, Backer, Litwack, Sloan, & Beck, 2014). The DBS-HAB subscale was used as the external correlate of interest for the EXT cluster because this subscale measures aggressive, reckless behavior reflective of the EXT cluster. Internal consistency was strong for the DBS-HAB subscale in the current sample (α = .93).

Data Analytic Strategy

Analyses were conducted using Mplus version 8 (Muthén & Muthén, 1998–2017). First, a method factor model was examined in which only method factors (i.e., PCL-5 and CAPS-5) were estimated. Next, PTSD measurement models were examined and compared. For these models, following established procedures for CFA of a multitrait-multimethod matrix (Brown, 2006; Byrne, 2012; Palmieri et al., 2007), each item loaded to both a PTSD symptom cluster latent variable (e.g., intrusions) and a method latent variable (e.g., CAPS-5). Symptom cluster latent variables were allowed to correlate; method latent variables were not allowed to correlate with each other or with symptom cluster latent variables. This approach estimated the amount of item variance explained by symptom cluster factors, while covarying for method. Additionally, residuals were correlated by symptom to account for within-symptom variance (see Bauer et al., 2013). Correlating the residuals for symptom D1, as measured by both the CAPS-5 and PCL-5, caused strain in the models. Accordingly, residuals for all symptoms except D1 were correlated.

As shown in Table 1, the specific models examined using this approach were the DSM-5 (APA, 2013), Dysphoria (Simms et al., 2002), Dysphoric Arousal (Elhai et al., 2011), Externalizing (Tsai et al., 2014), Anhedonia (Liu et al., 2014), and Hybrid (Armour et al., 2014) models. Additionally, higher-order versions of these models were examined in which symptom clusters (e.g., intrusions, avoidance) loaded on to a single higher-order PTSD latent variable. Next, once the best-fitting measurement model was identified, this model was examined without method factors to estimate the amount of item variance explained by symptom cluster factors alone. Last, the two versions of this best-fitting model---i.e., the version with method factors and the version without method factors--- were compared to estimate the amount of item variance explained by method factors beyond variance explained by symptom cluster factors. As a formal test of the degree to which items load to symptom clusters and method factors, we used Wald tests to constrain these loadings to equality. Used in this way, a significant Wald test indicates the loadings to symptom cluster and method factors are not equal.

To examine associations between newly proposed PTSD symptom clusters and external correlates, an additional CFA was conducted which estimated depression, anxiety, and aggressive driving latent variables as external correlates in addition to PTSD symptom cluster and method factors.

Using procedures developed by Meng, Rosenthal, and Rubin (1992), planned contrasts examined hypotheses regarding associations between specific symptom clusters and external correlates. Additionally, Cohen’s Q (Cohen, 1988) was calculated as an effect size for each pairwise comparison; values < .10, .10-.30, .30-.50, and > .50 indicate no effect, small effect, medium effect, and large effect, respectively. Within the Hybrid model, the proposed divisions of the DSM-5 negative alterations in cognition and mood (i.e., negative affect and anhedonia) and alterations in arousal and reactivity (i.e., externalizing, dysphoric arousal, and anxious arousal) symptom clusters were compared to examine the utility of dividing DSM-5 symptom clusters.

Individual items were used as estimators in all models. Because not all items approximated a normal distribution (see Supplemental Table 1), items were treated as ordinal (Flora & Curran, 2004; Wirth & Edwards, 2007). Parameters were estimated using the mean- and variance-adjusted weighted least squares (WLSMV) estimator which provides a robust χ² (Brown, 2006). Model fit was evaluated using χ², CFI, TLI, and RMSEA. Fit statistics were collectively evaluated for each model and established criteria were used to determine fit with CFI and TLI ≥ .90 and RMSEA ≤ .08 indicating adequate fit (Browne & Bollen, 1993; MacCallum, Browne, & Sugawara, 1996; Meyers, Gamst, & Guarino, 2006), and χ² p values > .05, CFI and TLI ≥ .95, and lower limit of the RMSEA 95% confidence interval < .05 indicating good fit (Brown, 2006; Hu & Bentler, 1999; Kline 2011). Nested models were compared using the DIFFTEST function in Mplus (Muthen & Muthen, 2006), which allows for comparison of nested models using the WLSMV estimator (Brown, 2006). Regarding missing data, the covariance coverage matrix from the measurement model CFAs ranged from .96 – 1.00 and in the construct validity models from .75 – 1.00 (median = .96). Missing data were handled using multiple imputation.

Results

Fit statistics for CFAs are presented in Table 2. Among models in which both symptom cluster and method factors were estimated, all examined models provided generally good fit to the data, with fit statistics in a similar range across models (e.g., RMSEA 90% CI ranged .04-.06, CFI ranged .95-.97, and TLI ranged .94-.96 for all models). The Anhedonia model failed to converge. The Hybrid model provided significantly better fit than the DSM-5 (χ² = 192.56, df = 15, p < .01), Dysphoric Arousal (χ² = 136.79, df = 11, p < .01), and Externalizing models (χ² = 98.92, df = 6, p < .01). Although the Dysphoria and Hybrid models cannot be directly compared, the Hybrid model appeared to provide generally better fit to the data.

Table 2.

Fit Statistics for DSM-5 PTSD Symptom Measurement Models

Model	χ2	df	CFI	TLI	RMSEA (90% CI)
Lower-Order Models
Method Factors Only	5468.15	721	.64	.61	.13 (.13–.14)
DSM-5	1348.70	675	.95	.94	.05 (.05–.06)
Dysphoria	1370.00	675	.95	.94	.05 (.05–.06)
Dysphoric Arousal	1262.66	671	.96	.95	.05 (.04–.05)
Externalizing	1226.47	666	.96	.95	.05 (.04–.05)
Anhedonia	-	-	-	-	-
Hybrid	1089.89	660	.97	.96	.04 (.04–.05)
DSM-5 no method	2463.00	715	.87	.86	.08 (.08–.08)
Hybrid no method	2090.60	700	.90	.88	.07 (.07–.08)
Higher-Order Models
DSM-5	1384.34	677	.95	.94	.05 (.05–.06)
Dysphoria	1378.01	677	.95	.94	.05 (.05–.06)
Dysphoric Arousal	-	-	-	-	-
Externalizing	1277.46	675	.96	.95	.05 (.04–.05)
Anhedonia	-	-	-	-	-
Hybrid	1151.72	674	.96	.96	.04 (.04–.05)

Note. The lower-order Anhedonia and higher-order Dysphoric Arousal and Anhedonia measurement models failed to converge; CFI = Bentler Comparative Fit Index; DSM-5 = fifth edition of the Diagnostic and Statistical Manual of Mental Disorders; TLI = Tucker Lewis Index; RMSEA = Root Mean Square Error of Approximation.

Higher-order measurement models were examined for each of the above-described models. Although each provided adequate to good fit to the data, the higher-order models provided significantly worse fit compared to their lower-order variants for the DSM-5 (χ² = 21.23, df = 2, p < .01), Dysphoria (χ² = 9.52, df = 2, p = .01), Externalizing (χ² = 45.43, df = 9, p < .01), and Hybrid (χ² = 64.22, df = 14, p < .01) models. Higher-order versions of the Dysphoric Arousal and Anhedonia models did not converge.

Correlations among symptom clusters in each lower-order model are presented in Table 3. The strongest association among any two symptom clusters in any model was between the NCM and DAR symptom clusters in the Dysphoric Arousal model. The magnitude of this association (r = .92) suggests these symptoms may not reflect distinct constructs. Associations between symptom clusters r ≥ .75 were observed in every measurement model, suggesting substantial overlap in symptoms. In particular, the AXA symptom cluster in the Hybrid model evidenced strong associations with the INT (r = .80), NAF (r = .79), ANH (r = .84), EXT (r = .69), and DAR (r = .71) symptom clusters.

Table 3.

Correlations among PTSD Symptom Clusters in Each Measurement Model

Model/Symptom Cluster	Correlations
DSM-5 Model	INT	AVD	NCM	AAR
INT	-	-	-	-
AVD	.64*	-	-	-
NCM	.63*	62*	-	-
AAR	.71*	.62*	.85*	-
Dysphoria Model	INT	AVD	DYS	HYP
INT	-	-	-	-
AVD	.63*	-	-	-
DYS	.69*	.62*	-	-
HYP	.60*	.60*	.75*	-
Dysphoric Arousal Model	INT	AVD	NCM	DAR	AXA
INT	-	-	-	-	-
AVD	.63*	-	-	-	-
NCM	.62*	.61*	-	-	-
DAR	.77*	.64*	.92*	-	-
AXA	.57*	.57*	.61*	.81*	-
Externalizing Model	INT	AVD	NCM	EXT	DAR	AXA
INT	-	-	-	-	-	-
AVD	.63*	-	-	-	-	-
NCM	.62*	.61*	-	-	-	-
EXT	.58*	.59*	.78*	-	-	-
DAR	.81*	.56*	.88*	.71*	-	-
AXA	.57*	.57*	.60*	.72*	.72*	-
Hybrid Model	INT	AVD	NAF	ANH	EXT	DAR	AXA
INT	-	-	-	-	-	-	-
AVD	.64*	-	-	-	-	-	-
NAF	.65*	.62*	-	-	-	-	-
ANH	.49*	.50*	.73*	-	-	-	-
EXT	.56*	.60*	.76*	.67*	-	-	-
DAR	.55*	.56*	.60*	.49*	.70*	-	-
AXA	.80*	.54*	.79*	.84*	.69*	.71*	-

Note. AAR = alterations in arousal and reactivity; ANH = anhedonia; AVD = avoidance; AXA = anxious arousal; DAR = dysphoric arousal; DYS = dysphoria; EXT = externalizing behaviors; HYP = hyperarousal; INT = intrusions; NAF = negative affect; NCM = negative alterations in cognition and mood;

^*= p<.05; coefficients in bold are all r ≥ .7.

Sources of method effect were examined in both the DSM-5 and Hybrid models. All symptom clusters were positively and significantly correlated in both models (see Table 3). First, as hypothesized, the model in which only method factors were estimated resulted in poor fit with and without correlated residuals (see Table 2). This suggests that CAPS-5 and PCL-5 items do not solely represent underlying method constructs.

Second, item loadings on symptom cluster factors and method factors in the DSM-5 and Hybrid models were reviewed to examine the degree to which CAPS-5 and PCL-5 items were better indicators of respective symptom clusters than of respective assessment modality. Within the Hybrid model in which both symptom cluster and method factor latent variables were estimated, items collectively evidenced weak loadings on their respective method factors (median standardized loading = .28; see Table 4). However, the CAPS-5 and PCL-5 varied in loadings to symptom cluster and method factors. For the PCL-5, all items evidenced salient loadings (i.e., standardized parameter estimates of .4 or greater; Brown, 2006) to respective symptom clusters although only items B4 (emotional reactivity), C1 (avoidance of thoughts), and C2 (avoidance of reminders) evidenced salient loadings to the questionnaire method factor, and the magnitude of loadings was higher to symptom cluster factors than the method factor for all items. For the CAPS-5, all items except B3 (flashbacks) and D1 (dissociative amnesia) had salient loadings to respective symptom cluster factors. However, several CAPS-5 items (B1, B4, B5, C1, C2, D4, D5, D6, D7, E3, and E6) evidenced salient loadings to the interview method factor. Item B1 (emotional reactivity) loaded more strongly to the interview method factor than the respective symptom cluster factor. Additionally, several items (B3, B4, B5, C2, E1, E3, and E6) had similar magnitude loadings to both symptom cluster and method factors. When loadings to symptom cluster and method factors were constrained to equality, the Wald test indicated loadings to symptom cluster factors and method factors were not equal (Wald test of parameter constraints = 54.79, df = 32, p < .01).

Table 4.

Hybrid Measurement Model Fully Standardized Item Loadings and Explained Variance with (and without) Covarying for Method

Symptom Cluster								Method		R2
Item	INT	AVD	NAF	ANH	EXT	AAR	DAR	CA	SR	SC	Both
CAPSB1	.53 (.73)							.61*		.53	.68
CAPS B2	.40 (.48)							.30*		.23	.25
CAPS B3	.37 (.45)							.31*		.21	.24
CAPS B4	.53 (.71)							.54*		.51	.57
CAPS B5	.57 (.74)							.53*		.55	.61
CAPS C1		.59 (.76)						.49*		.58	.58
CAPS C2		.55 (.75)						.53*		.56	.58
CAPS D1			.35 (.52)					.13		.13	.14
CAPS D2			.48 (.37)					.25*		.27	.29
CAPS D3			.58 (.67)					.37*		.45	.47
CAPS D4			.59 (.74)					.49*		.54	.59
CAPS D5				.61 (.75)				.48*		.56	.60
CAPS D6				.72 (.84)				.46*		.70	.72
CAPS D7				.65 (.79)				.45*		.62	.64
CAPS E1					.45 (.58)			.38*		.34	.35
CAPS E2					.46 (.49)			.16*		.24	.24
CAPS E3						.56 (.75)		.50*		.55	.57
CAPS E4						.54 (.62)		.32*		.38	.39
CAPS E5							.50 (.56)	.30*		.31	.33
CAPS E6							.40 (.52)	.40*		.27	.32
PCL B1	.84 (.80)								.09	.63	.71
PCL B2	.74 (.70)								.09	.49	.56
PCL B3	.67 (.63)								.09	.40	.45
PCL B4	.73 (.79)								.46*	.62	.74
PCL B5	.76 (.75)								.21*	.57	.61
PCL C1		.71 (.72)							.48*	.67	.74
PCL C2		.75 (.85)							.50*	.73	.80
PCL D1			.45 (.46)						.22*	.21	.25
PCL D2			.76 (.72)						.11	.52	.59
PCL D3			.64 (.65)						.25*	.43	.48
PCL D4			.84 (.84)						.25*	.71	.76
PCL D5				.84 (.85)					.22*	.72	.76
PCL D6				.92 (.91)					.13	.82	.87
PCL D7				.84 (.82)					.11	.67	.72
PCL E1					.86 (.83)				.08	.69	.75
PCL E2					.59 (.60)				.22*	.36	.40
PCL E3						.83 (.84)			.25*	.70	.75
PCL E4						.80 (.78)			.19*	.61	.68
PCL E5							.64 (.61)		.10	.38	.42
PCL E6							.62 (.61)		.18*	.38	.41

Note. Significance of item loadings are only reported for method factors as all symptom cluster loadings were significant (p < .05); AAR = anxious arousal; ANH = anhedonia; AVD = avoidance; Both = item variance accounted for by symptom cluster and method factors; CA = clinician-administered; DAR = dysphoric arousal; EXT = externalizing; INT = intrusions; NAF = negative affect; SC = item variance accounted for by symptom cluster; SR = self-report;

^*p < .05.

Within the DSM-5 model in which both symptom cluster and method factor latent variables were estimated, items on average had modest to weak loadings on their respective method factors (median standardized loading = .28; see Table 5). However, the CAPS-5 and PCL-5 again varied in loadings to symptom cluster and method factors. For the PCL-5, all items evidenced salient loadings to respective symptom clusters, only items E3 (hypervigilance) and E4 (startle) evidenced salient loadings to method factors, and all items had higher loadings to respective symptom clusters than the questionnaire method factor. For the CAPS-5, six items (B2, B3, D1, E1, E2, and E6) did not evidence salient loadings to respective symptom cluster factors; 12 items (B1, B4, B5, C1, C2, D3, D4, D5, D6, D7, E3, and E6) evidenced salient loadings to the interview method factor. Six items (B1, B4, C2, E1, E3, and E6) evidenced loadings to the method factor were stronger than loadings to respective symptom clusters and several additional items evidenced similar magnitude loadings to both method and symptom cluster factors. When loadings to symptom cluster and method factors were constrained to equality, the Wald test indicated loadings to symptom cluster factors and method factors were not equal (Wald test of parameter constraints = 50.40, df = 35, p < .05).

Table 5.

DSM-5 Measurement Model Fully Standardized Item Loadings and Explained Variance with (and without) Covarying for Method

Symptom Cluster					Method		R2
Item	INT	AVD	NCM	AAR	CA	SR	SC	Both
CAPSB1	.51 (.73)				.62*		.53	.64
CAPS B2	.38 (.48)				.32*		.23	.24
CAPS B3	.36 (.45)				.32*		.21	.23
CAPS B4	.51 (.71)				.54*		.51	.56
CAPS B5	.55 (.74)				.54*		.55	.60
CAPS C1		.56 (.76)			.51*		.58	.58
CAPS C2		.52 (.75)			.54*		.56	.56
CAPS D1			.33 (.50)		.15*		.13	.13
CAPS D2			.43 (.36)		.28*		.25	.26
CAPS D3			.53 (.64)		.40*		.41	.44
CAPS D4			.53 (.70)		.52*		.49	.55
CAPS D5			.53 (.69)		.51*		.48	.53
CAPS D6			.62 (.77)		.50*		.60	.64
CAPS D7			.57 (.73)		.50*		.53	.57
CAPS E1				.37 (.50)	.39*		.25	.29
CAPS E2				.39 (.43)	.19*		.18	.18
CAPS E3				.42 (.61)	.51*		.37	.44
CAPS E4				.41 (.52)	.34*		.27	.28
CAPS E5				.48 (.57)	.33*		.32	.33
CAPS E6				.38 (.52)	.41*		.27	.32
PCL B1	.83 (.79)					.09	.63	.69
PCL B2	.73 (.70)					.13*	.49	.55
PCL B3	.66 (.63)					.09	.40	.45
PCL B4	.81 (.79)					.24*	.62	.71
PCL B5	.78 (.75)					.10	.57	.62
PCL C1		.82 (.82)				.27*	.67	.74
PCL C2		.86 (.85)				.23*	.73	.80
PCL D1			.46 (.44)			.18*	.19	.24
PCL D2			.72 (.69)			.09	.47	.53
PCL D3			.64 (.77)			.25*	.39	.47
PCL D4			.81 (.73)			.24*	.63	.72
PCL D5			.83 (.80)			−.22*	.64	.73
PCL D6			.87 (.85)			−.31*	.72	.86
PCL D7			.80 (.77)			−.27*	.59	.71
PCL E1				.75 (.72)		.06	.51	.56
PCL E2				.56 (.53)		.07	.28	.32
PCL E3				.66 (.70)		.50*	.49	.68
PCL E4				.64 (.67)		.43*	.45	.59
PCL E5				.67 (.62)		−.15*	.38	.47
PCL E6				.65 (.62)		.09	.39	.43

Note. Significance of item loadings are only reported for method factors as all symptom cluster loadings were significant (p < .05); ANH = anhedonia; AVD = avoidance; AXA = anxious arousal; Both = item variance accounted for by symptom cluster and method factors; CA = clinician-administered; DAR = dysphoric arousal; EXT = externalizing; INT = intrusions; NAF = negative affect; SR = self-report; R² coefficients in parentheses represent explained variance without covarying for method factors; SC = item variance explained by symptom cluster;

^*p < .05.

Third, comparison of models in which only symptom cluster factors were estimated against models in which both symptom cluster and method factors were estimated was used to examine the amount of item variance explained by method factors beyond variance explained by symptom cluster factors alone. The Hybrid model in which only symptom cluster factors were estimated resulted in borderline adequate fit (χ² = 2090.60, df = 700, p<.001, RMSEA = .07 [90%CI = .07-.08], CFI = .90, TLI = .88) and explained 49.60% of the variance in PTSD symptom items, averaged across all items (see Table 3; R² values in parentheses). The addition of method factors significantly improved model fit (χ² = 619.78, df = 40, p<.01), but only accounted for 4.43% of additional variance, averaged across all items. The DSM-5 model in which only symptom cluster factors were estimated resulted in mediocre fit (χ² = 2463.00, df = 715, p<.001, RMSEA = .08 [90%CI = .08-.08], CFI = .87, TLI = .86) and explained 44.90% of the variance in PTSD symptom items, averaged across all items. The addition of method factors to this model also significantly improved model fit (χ² = 776.35, df = 40, p<.01), but only accounted for 5.70% of additional variance, averaged across all items. Therefore, although the addition of method factors to both models improved model fit, method factors ultimately accounted for little variance in PTSD symptoms beyond what was accounted for by symptom cluster factors.

Lastly, construct validity of novel symptom clusters from alternative models of PTSD was examined by contrasting their associations with measures of relevant external correlates. For the BDI-II, we compared the two-factor cognitive and somatic/affective (Huang & Chen, 2015) and three-factor cognitive, somatic, and affective factor (Vanheule et al., 2008) models. Although the 2-factor model provided adequate fit to the data (χ² = 369.10, df = 169, p<.001, RMSEA = .06 [90%CI = .06-.08], CFI = .96, TLI = .96), the 3-factor model provided generally better fit to the data (χ² = 127.47, df = 87, p =.003, RMSEA = .03 [90%CI = .03-.06], CFI = .99, TLI = .98). The two models cannot be directly compared using the DIFFTEST function due to different number of included items. However, the 3-factor model appeared to provide better fit overall and was retained.

The DSM-5 (χ² = 4342.37, df = 3460, p<.01, RMSEA = .03 [90% CI = .03-.04], CFI = .94, TLI = .93), Dysphoria (χ² = 4370.38, df = 3460, p<.01, RMSEA = .04 [90% CI = .03-.04], CFI = .94, TLI = .93), Dysphoric Arousal (χ² = 4290.61, df = 3450, p<.01, RMSEA = .03 [90% CI = .03-.04], CFI = .94, TLI = .94), and Externalizing (χ² = 4232.96, df = 3439, p<.01, RMSEA = .03 [90% CI = .03-.04], CFI = .94, TLI = .94) models all provided adequate fit to the data. The model estimating Hybrid model symptom clusters and external correlates resulted in good fit (χ² = 4323.95, df = 3427, p<.01, RMSEA = .03 [90% CI = .03-.03], CFI = .94, TLI = .94). With the exception of method factors, items had significant loadings to respective factors (see Supplemental Table 2). All symptom clusters were positively and significantly associated with all of the examined correlates (see Table 6).

Table 6.

Correlations between PTSD Symptom Clusters and External Correlates

	External Correlates
Model/Symptom Cluster	DBS-AD	BDI-Cog	BDI-Aff	BDI-Som	BAI-Som	BAI-Cog
DSM-5 Model
INT	.34*	.37*	.39*	.47*	.60*	.58*
AVD	.23*	.26*	.33*	.38*	.33*	.38*
NCM	.41*	.70*	.78*	.71*	.40*	.60*
AAR	.60*	.54*	.60*	.74*	.61*	.77*
Dysphoria Model
INT	.36*	.39*	.41*	.49*	.63*
AVD	.22*	.26*	.33*	.38*	.33*	.38*
DYS	.45*	.70*	.77*	.75*	.45*	.66*
HYP	.53*	.39*	.42*	.50*	.49*	.60*
Dysphoric Arousal Model
INT	.36*	40*	.42*	.50*	.64*	.61*
AVD	.23*	.26*	.33*	.38*	.34*	.39*
NCM	.41*	.70*	.78*	.71*	.40*	.60*
AXA	.42*	.22*	.32*	.40*	.44*	.51*
DAR	.62*	.64*	.67*	.82*	.61*	.80*
Externalizing Model
INT	.36*	.39*	.42*	.49*	.64*	.61*
AVD	.23*	.26*	.34*	.39*	.34*	.39*
NCM	.41*	.70*	.78*	.71*	.40*	.60*
EXT	.67*	.58*	.57*	.72*	.43*	.65*
AXA	.42*	.22*	.32*	.40*	.44*	.51*
DAR	.32*	.61*	.68*	.80*	.76*	.88*
Hybrid Model
INT	.34*	.37*	.42*	.49*	.64*	.62*
AVD	.18*	.26*	.36*	.41*	.35*	.40*
NAF	.41*	.63*	.54*	.60*	.46*	.66*
ANH	.30*	.66*	.81*	.72*	.30*	.47*
EXT	.66*	.58*	.58*	.72*	.43*	.66*
AXA	.38*	.22*	.32*	.39*	.44*	.50*
DAR	.35*	.60*	.70*	.81*	.76*	.88*

Note. AAR = alterations in arousal and reactivity; ANH = anhedonia; AVD = avoidance; AXA = anxious arousal; BAI-Cog = Beck Anxiety Inventory cognitive subscale; BAI-Som = Beck Anxiety Inventory somatic subscale; BDI-Aff = Beck Depression Inventory, second edition, affective subscale; BDI-Cog = Beck Depression Inventory, second edition, cognitive subscale; BDI-Som = Beck Depression Inventory, second edition, somatic subscale; DAR = dysphoric arousal; DBS-AD = Driving Behavior Survey aggressive driving subscale; DYS = dysphoria; EXT = externalizing behaviors; HYP = hyperarousal; INT = intrusions; NAF = negative affect; NCM = negative alterations in cognition and mood;

^*= p<.05; coefficients in bold are all r ≥ .7.

Within the DSM-5 model, the NCM cluster showed strong associations with all three clusters of depressive symptoms, whereas the AAR cluster showed strong associations with cognitive anxiety symptoms and somatic depressive symptoms (e.g., sleep disturbance, difficulty concentrating). Similar patterns were observed for the Dysphoria model. Within the Hybrid model, the NAF cluster had similar magnitude associations with cognitive and somatic depressive symptoms, as well as cognitive anxiety symptoms. The ANH cluster had similar magnitude associations with affective and somatic depressive symptoms. The EXT cluster was most strongly associated with somatic depressive symptoms (r = .72), an association somewhat stronger than the observed association with aggressive driving (r = .66). The AXA cluster was most strongly associated with cognitive anxiety symptoms, and the DAR showed similar magnitude associations with somatic depressive symptoms and cognitive anxiety symptoms.

Regarding the proposed division of the NCM symptom cluster into NAF and ANH clusters, ANH was significantly more strongly associated with depressive affect than NAF and the magnitude of this effect was large (see Table 7). However, NAF was not significantly more strongly associated with depressive cognition features than ANH. Regarding the proposed division of the AAR symptom cluster into EXT, AXA, and DAR clusters, EXT was more strongly associated with aggressive driving than the AXA and DAR symptom clusters; the effect size for these differences was medium. Likewise, DAR was more strongly associated with somatic depressive symptoms than the EXT and AXA clusters; the effect size for these differences was large. However, the opposite of the hypothesized associations were observed for the AXA symptom cluster. Specifically, both the EXT and DAR symptom clusters were more strongly associated with both cognitive and somatic anxiety symptoms than the AXA cluster. In the largest of these differences, the DAR cluster was more strongly associated with cognitive anxiety symptoms than the AXA cluster (large effect size).

Table 7.

Planned Contrasts among PTSD Symptom Clusters in the Hybrid Model and External Correlates

Contrast	Criterion	Z	Q
NAF vs. ANH	Cognitive Depression	−0.85	0.05
ANH vs. NAF	Affective Depression	8.30*	0.52
EXT vs. DAR	Aggressive Driving	6.64*	0.43
EXT vs. AXA	Aggressive Driving	7.16*	0.39
DAR vs. EXT	Somatic Depression	8.49*	0.54
DAR vs. AXA	Somatic Depression	8.52*	0.52
AXA vs. EXT	Cognitive Anxiety	−4.00*	0.24
AXA vs. DAR	Cognitive Anxiety	−12.40*	0.83
AXA vs. EXT	Somatic Anxiety	−7.10*	0.44
AXA vs. DAR	Somatic Anxiety	−8.52*	0.52

Note. ANH = anhedonia; AVD = avoidance; AXA = anxious arousal; DAR = dysphoric arousal; EXT = externalizing; INT = intrusions; NAF = negative affect;

^*= p < .05.

Discussion

To date, research examining the latent factor structure of DSM-5 PTSD symptoms has relied exclusively on questionnaires, leaving open the possibility of a method effect confound in the results. This study used data from both a questionnaire and structured diagnostic interview to compare competing measurement models of DSM-5 PTSD symptoms and examine the degree to which the measure used to assess symptoms affected the latent factor structure in a veteran sample. The seven-factor Hybrid model provided the best fit to the data and provided significantly better fit than the other models examined. Accordingly, previous findings suggesting that DSM-5 PTSD symptoms are best represented by the seven factors in the Hybrid model do not appear to be an artifact of exclusive reliance on questionnaire data. However, all considered models provided generally good fit to the data; fit statistics varied only slightly across all considered models. These results suggest that, although constraining parameter estimates from the seven-factor Hybrid model to the more parsimonious models provides a significant worsening of fit, the more parsimonious models provide similar quality fit to the data.

This study examined several sources of evidence for a method effect in the latent factor structure of DSM-5 PTSD symptoms. As anticipated, the model estimating only method factors resulted in poor fit, suggesting that PCL-5 and CAPS-5 items do not solely represent underlying method variables. Likewise, item loadings indicated that, on average, PCL-5 and CAPS-5 items are better indicators of the symptom clusters they represent than they are of the assessment modality used. Finally, comparison of a Hybrid model in which only symptom cluster factors were estimated, to a Hybrid model in which both symptom cluster and method factors were estimated, demonstrated that method factors accounted for little variance in PTSD symptoms beyond what was accounted for by symptom cluster factors. Taken together, these results suggest that measurement modality used to assess symptoms had little effect on the latent factor structure of DSM-5 PTSD symptoms.

However, CAPS-5 and PCL-5 items were not uniformly stronger indicators of underlying symptom cluster factors than method factors. In particular, CAPS-5 and PCL-5 items corresponding to symptoms B4 (emotional reactivity), C1 (avoidance of thoughts), and C2 (avoidance of reminders) evidenced salient loadings to method factors for both measures. These items highlight the greatest observed area of strain in the CAPS-5 and PCL-5 assessing the same underlying constructs. The distinction in avoidance symptoms may be partly attributable to differences in the measure instructions; CAPS-5 item severity scores reflect frequency and intensity of each symptom whereas PCL-5 instructions guide respondents to rate the degree to which they were bothered by each symptom. Respondents may not have been bothered by avoiding internal and external reminders of traumatic events; individuals with PTSD typically describe these strategies as main coping mechanisms. Likewise, differential rating of emotional reactivity may be a function of under-reporting distress, a well-established phenomenon among veterans. However, additional research is needed to understand the causes of measure-specific differences in evaluating these symptoms and how these symptoms differ from the majority of symptoms which were consistently evaluated across measures.

Although the primary focus of this study was on the possible effect of assessment method on the factor structure of DSM-5 PTSD, it is important to consider the nature of the measurement models investigated. Consistent with several factor analytic studies (Armour et al., 2016), the seven-factor Hybrid model provided the best fit to the data while covarying for method. At face value, this result is consistent with a growing literature suggesting that these less parsimonious conceptualizations of PTSD provide a better representation of symptom covariance than does the DSM-5 model. However, all considered models provided good fit to the data and differences in fit between models was minimal. Accordingly, although less parsimonious models appear to provide slightly improved model fit, observed limited difference in model fit raise concerns regarding the incremental value in parsing symptom clusters.

Correlations among symptom clusters and results of construct validity analyses raise additional concerns regarding the utility of less parsimonious measurement models. Strong associations between symptom clusters were observed between symptom clusters in all models. Of note, the strongest association among any two symptom clusters in any model was between the NCM and DAR symptom clusters in the Dysphoric Arousal model. The magnitude of this association (r = .92) is reminiscent of the critique raised by Simms and colleagues (2002) that NCM (e.g., loss of interest, detachment, difficulty experiencing positive feelings) and DAR (difficulty concentrating and sleep disturbance) are symptoms of a common general distress construct. Arguably the greatest overlap in any model was observed in the Hybrid model, in which the AXA symptom cluster demonstrated a strong association with all other symptom clusters. Collectively, these results raise concerns regarding the degree to which such symptoms represent distinct underlying constructs.

Results of construct validity analyses were mixed. Whereas some of the newly proposed symptom clusters demonstrated hypothesized differential associations with external correlates, others did not. Specifically, the proposed ANH, EXT, and DAR clusters were associated with theoretically similar external correlates as hypothesized. However, NAF and AXA did not demonstrate unique associations with theoretically similar external correlates. Collectively, these results indicate that proposed divisions of symptom clusters in the Hybrid model do not all demonstrate theoretically driven differential associations with external correlates. These results raise additional concerns regarding the conceptual value and functional utility of dividing the DSM-5 symptom clusters in this manner.

Strong overlap between PTSD symptom clusters and depressive and anxiety symptoms were not limited to the Hybrid model. Observed associations between the DSM-5 model NCM cluster and cognitive (r = .70), affective (r = .78, and somatic (r = .71) depressive symptoms, and between the Externalizing model and somatic depressive symptoms (r = .82), somatic anxiety symptoms (r = .76), and cognitive anxiety symptoms (r = .88) is consistent with an extensive literature critiquing these non-specific symptoms as common to several mood and anxiety disorders (see Simms et al., 2002 for a discussion).

It is worth noting that all research examining the construct validity of the alternative model symptom clusters, including this study, is based solely on self-report measures of imperfect validity constructs (e.g., anxiety, depression, quality of life). Stronger sources of construct validity evidence are needed to evaluate the degree to which these hypothesized constructs are associated with an array of related features (e.g., externalizing symptoms with criminal activity, aggressive driving, interpersonal violence). Additionally, alternative assessment modalities (e.g., corroborative reports, behavioral task data, ecological momentary assessment) are needed to provide more rigorous tests of construct validity. Accordingly, although these findings provide valuable information about the differential associations of the newly proposed symptom clusters with external correlates, these findings only represent a first step in what needs to be a thorough process of explication and construct validation.

The less parsimonious measurement models of DSM-5 PTSD have two major limitations. First, as noted, the justification for splitting of DSM-5 symptom clusters is empirical rather than conceptual. Thus, although these less parsimonious models provide better model fit, the added conceptual value of distinctions among the new factors remains unclear. Unlike the one other existing study examining this question to date (Pietrzak et al., 2015), results from the current study suggest the proposed divisions may not have functional utility. This discrepancy highlights the importance of further replication in examining the value, beyond model fit, of dividing DSM-5 PTSD symptom clusters. Of particular value to this line of inquiry will be further examination of the degree to which these newly proposed symptom clusters demonstrate differential associations with other external correlates of interest (e.g., functional impairment), patterns of change over time, and response to intervention. The relative lack of a method effect observed in our study indicates that such research may proceed utilizing either self-report or clinician-administered measures.

A second major limitation of these less parsimonious DSM-5 PTSD models is content coverage. As the number of proposed symptom clusters accounting for variance in the 20 DSM-5 PTSD symptoms has increased, the number of symptoms per symptom cluster has decreased. For DSM-5 correspondent measures (e.g., CAPS-5, PCL-5), four symptom clusters in the Hybrid model (avoidance, externalizing behaviors, anxious arousal, and dysphoric arousal) are represented by only two items each. As others have noted (e.g., Palmieri, Marshall, & Schell, 2007), it is unlikely that two items are sufficient to assess the full range of these constructs. Further research that uses measures with more than one item per symptom is needed to examine the fit of less parsimonious models.

One practical implication of alternative PTSD measurement models is their impact on diagnostic criteria and associated clinical impact. Shevlin, Hyland, Karatzias, Bisson, and Roberts (2017) expressed concern that none of the authors of the alternative PTSD measurement models proposed diagnostic algorithms for these new models. Shevlin et al. proposed diagnostic algorithms for each of the alternative models and found that PTSD prevalence rates were much lower using the proposed diagnostic algorithms relative to the DSM-5 algorithm as well as important differences between trauma types and probability of meeting diagnostic criteria across models. CFA fit indices inherently favor less parsimonious models and the PTSD factor analytic literature has adopted divisions of theoretically derived symptom clusters in an effort to find models that satisfy this analytic approach. Unfortunately, this literature has dedicated substantially less effort to recognizing the clinical implications of these less parsimonious models. Factor analytic evidence is an essential tool to understanding clinical phenomena, but should be used as one of many sources of construct validity evidence. Work is needed to reconcile the gap between the factor analytic literature and diagnostic practices. We strongly recommend such work rigorously compare several different diagnostic algorithms using a diversity of samples (e.g., veteran and non-veteran clinical and community samples), range of index traumas, and both self-report and clinician-administered measures. Such work should be based on a conceptual understanding of the disorder and input from content experts.

In conclusion, results from this study indicate that all considered measurement models of DSM-5 PTSD symptoms provided similar fit to the data while covarying for method. Little evidence of a method effect was observed, indicating that most symptoms are better indicators of the symptom clusters they represent than the assessment modality used to measure them. Additionally, although the seven-factor Hybrid model (Armour et al., 2015) provided the best fit to the data, observed correlations among symptoms clusters and construct validation evidence raise concerns regarding the functional utility of dividing DSM-5 symptom clusters. Accordingly, the constructs implied by the new factors in the more complex PTSD measurement models require further explication and construct validation efforts.