A multi-sample confirmatory factor analysis of PTSD symptoms: What exactly is wrong with the DSM-IV structure?

https://doi.org/10.1016/j.cpr.2012.10.004Get rights and content

Abstract

Within the DSM-IV, PTSD symptoms are rationally classified as assessing one of three symptom domains: reexperiencing, avoidance/numbing, or hyperarousal. However, two alternative four-factor models have been advocated as superior to the DSM-IV framework, based on confirmatory factor analysis. In the Numbing model, symptoms of emotional numbing are differentiated from avoidance. In the Dysphoria model, several symptoms of numbing and hyperarousal are combined to form a factor purported to assess general psychological distress. Examination of these models, within 29 separate data sets, supports two conclusions. First, contrary to its conceptual underpinnings, the Dysphoria model differs empirically from the Numbing model solely in the correlation predicted between two hyperarousal symptoms; all other predicted correlations made by the two models are substantively identical. Second, when the factor analytic presumption of simple structure is relaxed to allow for potential presentation order effects, other plausible symptom structures emerge. In particular, the fit of the DSM-IV model improved dramatically and was a better fit to the data than either four-factor model. The ostensible inferiority of the DSM-IV model may be due to a methodological artifact stemming from the order in which symptoms are typically assessed. The provisional decision to revise the structure of PTSD symptoms in the DSM-5 in light of confirmatory factor analytic results may be misguided.

Highlights

► The symptom structure of posttraumatic stress disorder (PTSD) is controversial. ► The difference between the Numbing and Dysphoria models has been mischaracterized. ► Non-clinical factors may underlie the seeming misfit of the DSM-IV model of PTSD. ► Proposed revisions to the forthcoming DSM-5 may be misguided.

Introduction

Posttraumatic stress disorder (PTSD) has been the subject of considerable controversy, predating its formal codification in the Diagnostic and Statistical Manual of Mental Disorders-III (DSM-III; American Psychiatric Association, 1980). Areas of disagreement range from concerns about whether PTSD is a legitimate disorder as opposed to a sociopolitical construct (Scott, 1990, Summerfield, 2001, Young, 1995) to differences regarding the defining features of PTSD (Bovin and Marx, 2011, Spitzer et al., 2007). One of the more challenging issues is the multiplicity of views regarding the optimal framework for conceptualizing the core symptoms of PTSD. Continued progression of knowledge regarding PTSD symptom structure is essential inasmuch as improvements in diagnosis, assessment, prevention, and treatment are contingent on a solid understanding the nature of PTSD. Yet, there is reason for concern that the pace of advancement in this area may have slowed.

As currently conceptualized, PTSD is an anxiety disorder exemplified by three clusters of symptoms that may follow from exposure to a traumatic life event (DSM-IV; American Psychiatric Association, 1994). As described in the DSM-IV, these three symptom clusters are composed of seventeen symptoms and reflect the phenomena of reexperiencing (Criterion B), e.g., intrusive thoughts of the trauma, avoidance and emotional numbing (Criterion C), e.g., avoiding reminders of the trauma, and hyperarousal (Criterion D), e.g., hypervigilance.

Whereas the organizing framework represented in the DSM-IV was derived on the basis of expert clinical consensus, confirmatory factor analytic techniques have been used to examine the structure of PTSD symptoms (for reviews, see Elhai and Palmieri, 2011, Yufik and Simms, 2010). Based on model fit, two distinct four-factor models of PTSD symptom structure are now widely viewed as superior to the three-factor representation embodied in the DSM-IV. The first of these alternative models, proposed by King and colleagues (King, Leskin, King, & Weathers, 1998), posits the existence of four correlated factors. Two of these symptom clusters, i.e., reexperiencing and hyperarousal, map directly onto clusters proposed in the DSM-IV. The model departs from the DSM-IV framework, however, by hypothesizing that the Criterion C symptom cluster is composed of two separable symptom sets assessing avoidance and emotional numbing. Numerous confirmatory factor analytic studies have revealed this model—hereafter referred to as the Numbing model—to provide a very good fit to the data (e.g., Asmundson et al., 2000, Asmundson et al., 2003, ⁎DuHamel et al., 2004, King et al., 1998, ⁎Marshall, 2004, ⁎Palmieri and Fitzgerald, 2005, ⁎Palmieri, Marshall and Schell, 2007, ⁎Palmieri, Weathers, Difede and King, 2007.

A second four-factor model, proposed by Simms, Watson, & Doebbeling (2002), holds that PTSD symptoms can be differentiated into those assessing reexperiencing, avoidance, dysphoria, and hyperarousal. The first two of these dimensions, i.e., reexperiencing and avoidance, are identical in item content to those found in the Numbing model. This model differs from the Numbing model, however, in positing a Dysphoria factor which is thought to assess non-specific psychological distress characteristic of various disorders. The dysphoria dimension contains all symptoms construed in the Numbing model as measuring emotional numbing as well as three of the five symptoms regarded in the DSM-IV as assessing hyperarousal. The hyperarousal dimension is retained but is indicated by only the last two DSM-IV hyperarousal symptoms, i.e., hypervigilance (D4) and exaggerated startle response (D5). This conceptual framework, which we refer to as the Dysphoria model, has also been well-supported in confirmatory factor analytic investigations (e.g., ⁎Elhai et al., 2007, Elklit and Shevlin, 2007, ⁎Krause et al., 2007, ⁎Palmieri, Weathers, Difede and King, 2007). Abbreviated symptom content and symptom-dimension alignments for the DSM-IV, Numbing, and Dysphoria models are depicted in Table 1.

At this point, neither four-factor alternative to the DSM-IV framework of PTSD symptom structure has emerged as clearly preferable, although a recent meta-analytic comparison of the two models determined that the Dysphoria model may provide a marginally better fit (Yufik & Simms, 2010). Despite the absence of consensus regarding an optimal representation of PTSD symptom structure, there is widespread agreement, based on factor analytic results, that the DSM-IV model provides an inadequate depiction of PTSD symptom structure. Moreover, this commonly-held conviction has prompted the frequent observation that the DSM should be reformulated to align the conceptualization of PTSD with confirmatory factor analytic findings (e.g., ⁎DuHamel et al., 2004, ⁎Hetzel-Riggin, 2009, King et al., 1998, ⁎Palmieri, Weathers, Difede and King, 2007, ⁎Simms et al., 2002).

The calls for revision appear to have been heeded. The provisional draft of the DSM-5 embraces certain aspects of the findings derived from confirmatory factor analysis (American Psychiatric Association, 2010). At first glance, the integration of confirmatory factor analytic research results might be regarded as an unmitigated advance in the development of a more empirically-based psychiatric taxonomy. On closer inspection, however, it is not clear that the body of confirmatory factor analytic research findings is sufficiently compelling to have warranted revision of the forthcoming DSM-5.

The current study is a reanalysis of this literature, designed to investigate two limitations of confirmatory factor analytic research examining the structure of PTSD. First, factor analytic examinations of PTSD symptoms, just as structural equation modeling more generally (MacCallum, Wegener, Uchino, & Fabrigar, 1993), have tended to consider as plausible only a small set of potential models. Specifically, PTSD researchers have demonstrated a pronounced preference for models that conform to simple structure. The assumption that any pair of PTSD symptoms is correlated because the items share one, and only one, underlying cause simplifies model interpretation. However, simple structure is, at best, a heuristic aid, and may not adequately characterize psychological phenomena (Marsh, Hau, & Grayson, 2005). Near exclusive reliance on simple structure models, in the absence of clear and compelling justification, artificially constrains the scope of models that have come under examination and may have led to a biased understanding of the nature of PTSD.

In considering what additional influences on PTSD symptom structure might warrant departure from simple structure, it is noteworthy that PTSD researchers have focused almost exclusively on semantic content as the sole determinant of symptom covariation. Yet, reported PTSD symptoms may covary for reasons other than item content. In particular, the order in which a series of items is presented can create a context that influences covariation between items. Various theoretical accounts bear on the widespread empirical finding that responses to contiguous items may be more strongly correlated relative to responses to the same questions when placed further apart in an instrument (e.g., Schwarz, 1999, Schwarz and Sudman, 1992, Tourangeau and Rasinski, 1988). For example, a considerable body of theory holds that people are motivated to maintain consistency (Bem, 1967, Festinger, 1957, Heider, 1958, Osgood et al., 1957). From this perspective, responses to successive questions might be answered in a manner that increases consistency with earlier responses. Other theories suggest cognitive mechanisms that could result in order effects. For example, an initial response may serve as a cognitive anchor that constrains the range of subsequent responses (Tversky & Kahneman, 1974). In addition, the process of answering an initial question may make information available that is primed for use in subsequent questions (e.g., Collins and Quillian, 1970, Schell et al., 1996).

To the degree that responses to queries about PTSD symptoms may vary as a function of the order in which items are presented, the failure to consider possible order effects in PTSD models may constitute a significant oversight. This omission is compounded inasmuch as PTSD symptoms are almost always presented to respondents in the sequence in which they are enumerated in the DSM. To model an order effect requires deviation from simple structure and, to our knowledge, has not been done in any of the published PTSD confirmatory factor analytic studies to date.

A second shortcoming of existing research is that the literature may rely too heavily on global model fit and visual inspection of path diagrams to draw theoretical inferences regarding PTSD symptom structure. Although path diagrams are quite useful for visualizing and communicating model assumptions, they have limited ability to facilitate understanding how and why any given model is superior to another in accounting for actual data. Indeed, path diagrams embodying substantially different causal assumptions can represent mathematically equivalent models (Hershberger, 2006, Lee and Hershberger, 1990, MacCallum et al., 1993, Raykov and Marcoulides, 2001). That is to say, the models make exactly the same predictions about the empirical covariance matrix despite substantially different path diagrams. For example, given a data set in which the Dysphoria model fits better than the Numbing model, it would be common practice to use the Dysphoria model path diagram as a basis for inferring that symptom D1, i.e., sleep disturbance, should be interpreted as measuring dysphoria rather than hyperarousal. Yet, it is entirely possible that the Numbing model might actually be better at explaining correlations involving D1. The actual difference between models cannot be inferred easily from differences in the path diagrams.

Several theorists within the methodological literature have discussed the limitations of the standard practice in which researchers theoretically interpret the path diagram of the model with the best global fit (e.g., Bollen, 1989, Pearl, 1988). Although some methodologists have tried to work out new procedures that rely on local fit indices to derive theoretical conclusions (e.g., Shipley, 2000), these methods have not been adopted in any substantive literature. In addition, these methods do not address the challenge facing the PTSD researchers in which multiple plausible factor models fit the data similarly well. Of particular interest in the current analyses is whether the Dysphoria and Numbing models differ empirically in ways that are consistent with the conceptual distinction that has been made between the two models.

To improve the theoretical interpretation of empirical differences between models, this paper uses a novel descriptive approach to directly assess differences between two non-nested models. Specifically, we compute the difference between predictions made by each model for each element of the correlation matrix. In other words, we compare the correlations implied by the Numbing and Dysphoria models, on an element by element basis, to determine precisely where the two models make discrepant empirical predictions.1 This strategy allows precise identification of differences in hypotheses implied by the models regardless of how those models are represented in path diagrams. In this way, the theoretical meaning of differences in aggregate model fit across two models can be grounded in a description of the specific data elements about which the models make different predictions.

This paper uses data sets from previously-published confirmatory factor analytic examinations of PTSD symptom structure to address two objectives. After first examining the fit of the DSM-IV, Numbing, and Dysphoria models across data sets, we sought: (1) to identify the specific correlations for which the Dysphoria and Numbing models make different predictions to guide the theoretical interpretation of differences in model fit; and (2) to examine the extent to which an alternative factor model, which relaxes the simple structure assumption to allow for possible order effects, provides a plausible account of the data.

Section snippets

Identification of eligible studies

To examine the underlying structure of the DSM-IV PTSD symptoms, we reanalyzed data from studies that used confirmatory factor analysis to evaluate DSM-based PTSD measures. Eligible studies were identified through computerized database searches of MEDLINE and PsycINFO, using posttraumatic stress disorder, confirmatory factor analysis, and symptom structure as keywords. We confined ourselves to confirmatory factor analytic studies of traumatized adult samples that examined 17-item PTSD symptom

The fit of alternative models

As an initial step, the fit of alternative models of PTSD symptom structure were examined. As shown in Table 3, nearly all models across all data sets (5 models × 29 data sets) had acceptable model fit as judged by the two-index criteria advocated by Hu and Bentler (1999). Specifically, most models either had an SRMR below .06 or a CFI above .95. In most cases, this was due to low SRMR values, indicating that the correlations implied by the models were very close to the actual correlations in

Discussion

Extensive attention has been devoted to identifying the structure of PTSD symptoms using confirmatory factor analysis. In recent years, the main focus of research has expanded from documenting the fit of the DSM-IV model to determining which of two four-factor alternative conceptualizations of PTSD symptom structure should be preferred over the current DSM-IV framework (e.g., Yufik & Simms, 2010). This line of investigation has had considerable impact, undermining support for the existing

Conclusions

Despite the considerable energy devoted to examining the factor structure of PTSD symptoms, the scientific path the field has undertaken may not provide firm footing for DSM-5 revisions to the definition of the disorder. The current findings cast doubt on the claim that the DSM-IV model of PTSD must be changed to accommodate the findings from available factor analytic studies. Indeed, the only thing “wrong” with the DSM-IV model of PTSD is that it systematically underestimates the correlations

References13 (89)

  • H. Akaike

    Factor analysis and the AIC

    Psychometrika

    (1987)
  • American Psychiatric Association

    Diagnostic and Statistical Manual of Mental Disorders

    (1980)
  • American Psychiatric Association

    Diagnostic and Statistical Manual of Mental Disorders

    (1987)
  • American Psychiatric Association

    Diagnostic and Statistical Manual of Mental Disorders

    (1994)
  • American Psychiatric Association

    DSM-5 development. Posttraumatic stress disorder

  • G.J.G. Asmundson et al.

    Post-traumatic stress disorder symptoms in United Nations peacekeepers: An examination of factor structure in peacekeepers with and without chronic pain

    Cognitive Behaviour Therapy

    (2003)
  • J.S. *Baschnagel et al.

    Factor structure of posttraumatic stress among Western New York undergraduates following the September 11th terrorist attack on the World Trade Center

    Journal of Traumatic Stress

    (2005)
  • D.J. Bem

    Self-perception: An alternative interpretation of cognitive dissonance phenomena

    Psychological Review

    (1967)
  • P.M. Bentler

    Comparative fit indexes in structural models

    Psychological Bulletin

    (1990)
  • P.M. Bentler

    EQS Structural Equations Program Manual

    (1995)
  • D.D. Blake et al.

    The development of a clinician-administered PTSD scale

    Journal of Traumatic Stress

    (1995)
  • K.A. Bollen

    Structural equations with latent variables

    (1989)
  • M.J. Bovin et al.

    The importance of the peritraumatic experience in defining traumatic stress

    Psychological Bulletin

    (2011)
  • G.E.P. Box

    Science and statistics

    Journal of the American Statistical Association

    (1976)
  • K.P. Burnham et al.

    Multimodel Inference: Understanding AIC and BIC in model selection

    Sociological Methods & Research November

    (2004)
  • M.W. Cheung et al.

    Meta-analytic structural equation modeling: A two-stage approach

    Psychological Methods

    (2005)
  • M.R. Cross et al.

    Validation of a self-report measure of posttraumatic stress disorder in a sample of college-age women

    Journal of Traumatic Stress

    (2001)
  • C.A. *Cuevas et al.

    HIV/AIDS Cost Study: Construct validity and factor structure of the PTSD checklist in dually diagnosed HIV-seropositive adults

    Journal of Trauma Practice

    (2006)
  • M.A. *Cyders et al.

    Disaggregating the relationship between posttraumatic stress disorder symptom clusters and chronic orofacial pain: Implications for the prediction of health outcomes with PTSD symptom clusters

    Annals of Behavioral Medicine

    (2010)
  • J.R. Davidson et al.

    Assessment of a new self-rating scale for post-traumatic stress disorder

    Psychological Medicine

    (1997)
  • K.N. *DuHamel et al.

    Construct validity of the posttraumatic stress disorder checklist in cancer survivors: Analyses based on two samples

    Psychological Assessment

    (2004)
  • J.D. *Elhai et al.

    Assessing posttraumatic stress disorder with or without reference to a single, worst traumatic event: Examining differences in factor structure

    Psychological Assessment

    (2009)
  • J.D. *Elhai et al.

    Structural validity of the posttraumatic stress disorder checklist among college students with a trauma history

    Journal of Interpersonal Violence

    (2007)
  • J.D. *Elhai et al.

    Posttraumatic stress disorder's frequency and intensity ratings are associated with factor structure differences in military veterans

    Psychological Assessment

    (2010)
  • J.D. Elhai et al.

    Evidence for a unique PTSD construct represented by PTSD's D1-D3 symptoms

    Journal of Anxiety Disorders

    (2011)
  • A. Elklit et al.

    The structure of PTSD symptoms: A test of alternative models using confirmatory factor analysis

    British Journal of Clinical Psychology

    (2007)
  • R.M. *Engdahl et al.

    Comparing posttraumatic stress disorder's symptom structure between deployed and nondeployed veterans

    Psychological Assessment

    (2011)
  • L. Festinger

    A theory of cognitive dissonance

    (1957)
  • E.B. Foa et al.

    The validation of a self-report measure of posttraumatic stress disorder: The Posttraumatic Diagnostic Scale

    Psychological Assessment

    (1997)
  • E.B. Foa et al.

    Reliability and validity of a brief instrument for assessing post-traumatic stress disorder

    Journal of Traumatic Stress

    (1993)
  • A. Frances

    A warning sign on the road to DSM-V: Beware of its unintended consequences

    Psychiatric Times

    (2009)
  • A. Gelman et al.

    Avoiding model selection in Bayesian social research

    Sociological Methodology

    (1995)
  • M. Hallquist

    MplusAutomation: Automating Mplus model estimation and interpretation

    R package version 0.4-3

    (2011)
  • F. Heider

    The psychology of interpersonal relations

    (1958)
  • Cited by (34)

    • Locating event centrality in associations of emotion regulation with posttraumatic stress disorder symptoms and posttraumatic growth in emerging adults

      2022, Journal of Migration and Health
      Citation Excerpt :

      Posttraumatic stress disorder (PTSD) and posttraumatic growth (PTG) are typical posttraumatic reactions (Brewin et al., 2000; Elderton et al., 2017; Helgeson et al., 2006). PTSD is characterized by persistent intrusions or recollections of the traumatic event, avoidance of the event reminders and emotional numbing, hyper-arousal and negative cognitions and emotions associated with the traumatic event (Marshall et al., 2013). PTG, on the other hand, is the emergence of positive behavioral tendencies or transformation within the individual after a traumatic experience (Tedeschi and Calhoun, 1996), but such behavioral tendencies are not part of a typical developmental process (Cook, 2017).

    • The symptom structure of posttraumatic stress disorder and co-morbid depression among college students with childhood abuse experience: A network analysis

      2021, Journal of Affective Disorders
      Citation Excerpt :

      Generalization of the study results into populations with clinical diagnoses could be limited. Thirdly, instead of the true construct differences, the clusters of PTSD and depression symptoms could represent scale factors, such as the order of items (Marshall et al., 2013). As a covariance-based method, network analysis could have Berkson's bias (Berkson, 2014; Phillips et al., 2018; de Ron et al., 2019; Lazarov et al., 2019) when using the same scale items to select the study group and build the network, leading to spurious correlations.

    • Elevated C-reactive protein and posttraumatic stress pathology among survivors of the 9/11 World Trade Center attacks

      2017, Journal of Psychiatric Research
      Citation Excerpt :

      Post traumatic stress pathology is heterogeneous (Galatzer-Levy and Bryant, 2013) and symptoms of PTSD have been divided into four clusters reflecting these diverse components including symptoms of re-experiencing, avoidance, negative alterations in cognitions and mood (negative cognitions/mood), and alterations in arousal and reactivity (arousal). These components are supported by factor-analytic studies (Cox et al., 2002; Marshall et al., 2013) and have been adopted for the DSM-V criteria. Elevated CRP levels have been reported in association with re-experiencing symptoms of PTSD (Canetti et al., 2014; Miller et al., 2001; von Känel et al., 2007), and recently, with arousal symptoms (Michopoulos et al., 2015).

    View all citing articles on Scopus

    This research was supported by grants R01MH56122, R01MH071636, and R01MH087657 from the National Institute of Mental Health and grant R01AA014246 from the National Institute on Alcohol Abuse and Alcoholism. The views expressed are ours and do not necessarily reflect those of the funders or RAND. We thank the many researchers who graciously shared their data with us as well as the thousands of trauma survivors who participated in these studies.

    13

    References marked with an asterisk indicate studies included in the data analysis.

    View full text