SEM Studies on PTSD

by David Baldwin, PhD

It's been over a month since I posted a question here about what sorts of structural equation modeling (SEM) studies have been done to confirm the three PTSD symptom clusters (intrusion, avoidance, hyperarousal) in the DSM's diagnostic criteria.

This post summarizes what I have learned about this since that question, and is contributed by way of thanks to several list-members who suggested various articles and leads. In summarizing some of the relevant studies below, I'll also try to make clear what structural equation modeling is and why this analytic method could be useful in learning more about PTSD, for those who have no idea what I'm talking about here.

Basically, structural (or causal) modeling is descended from, and combines, factor analysis and path analysis. It lets an investigator "fit" several theories (models) to the data, and so can be used to confirm ideas about the definition of factors and relationships among them. This approach does *not* test a hypothesis against the null hypothesis; if a model is right, then it will fit the data (i.e., there will not be a significant difference between observations [data] and that model's predictions). The two major computer programs (for PCs & Macs) that "do" SEM are LISREL and EQS, but probably there are others as well. Peter Bentler (at UCLA), an early proponent, wrote an influential article on this in Annual Review of Psychology (1980) [31, 419-456]. Also, a special section in Journal of Consulting and Clinical Psychology in 1993 [61(3)] on using SEM in clinical research gives background on this analytic procedure without discussing trauma specifically. And Bob Basham mentioned a mail list ("semnet") about structural models, on the net.

First off, apparently nothing's been published yet quite like what I was looking for (a SEM study of the three PTSD clusters, including the causal relationships among them and how these change over time). What I was thinking was maybe somebody had used the DSM-IV criteria as the "theory" and tested it against a PTSD dataset somewhere. No such luck.

Although several reports (summarized briefly below) have addressed parts of this question, all use just self-report data from victims/patients. From what I remember of this stuff when I learned it at OSLC (a decade ago, on non-trauma problems), SEM models work better if they are fed data that includes the perspectives from different agents (e.g., patient, spouse, therapist) and also from different methods (e.g., self-report, observations of patient behavior, and psychophysical measures). Using diverse data such as these allows a more true (well, less biased) measurement of each construct (or factor; e.g., a PTSD symptom cluster in DSM), since each agent and each method share a certain amount of error variance with other data from that source. [Graphically, I suppose, this is like illuminating an object from multiple light sources to minimize shadows (bias).] Anyway, using data from multiple agents and multiple methods should minimize problems related to the unreliability of (all) measures, since constructs are measured by overlapping variance from diverse sources. But I didn't find any reports that included psychophysical measures -- or observational data or collateral reports -- along with self-reports.

Not clear why such data are not being combined in studies of PTSD. It is certainly possible that this old notion of mine that multiple method/agent data are preferable is no longer considered important, but I haven't heard that. More likely there are practical reasons, like the need for relatively large samples so it costs too much, not many clinicians know about this, or its more work or something.

Still, several SEM studies (some are recent ones) seem able to contribute valuable information about the diagnosis. I'll summarize some of these reports below, since they can illustrate the method's usefullness.

Rosemary Webster and Justin Kenardy each mentioned an early study by Mark Creamer et al (1992). In a paper in Journal of Abnormal Psychology [101(3), 452-459], Creamer's group studied people who worked in a downtown Melbourne office building where some guy walked in and fatally shot people on three floors. Office workers in a different Melbourne office building served as a contrast group. Subjects were all followed at 4, 8, and 14 months; measures were the IES' Intrusion and Avoidance scales, SCL-90-R's GSI scale, and did they fear for their safety (subjective; Y/N). People were also categorized for objective trauma exposure: high if they worked on the same floor as where others were shot, and low(er) exposure if they worked on other floors; N = 158 (with data at all points).

The SEM model that fit showed a path from Trauma/Fear to Intrusion and then to Distress and Avoidance (and not directly from Trauma/Fear to Avoidance or to Distress), with Intrusion predicting Distress at all phases (and not the reverse), suggesting that intrusive symptoms "drive" avoidance, at least initially. People's subjective fear reaction was more predictive of their later symptoms than the objective trauma exposure measure. Later on (like at 14 months), avoidance stopped predicting distress -- there was evidence that people with high late avoidance may have used that as a coping strategy: initially with poor results, but gradually they could 'button it up'. Also, Creamer et al thought higher levels of intrusion early on predicted fewer symptoms later (I believe most of it, except for this last part, BTW).

Unfortunately, nobody got diagnosed as PTSD or not, there was no measure of hyperarousal, and of course there is no info about these people from before the shootings (e.g., pre-existing differences in avoidance as a general coping skill, etc.). And then Beth Stamm, who is currently doing a lot on PTSD variables with SEM, sent some stuff about how the IES really measures Cognition and Affect rather than Intrusion and Avoidance. Grrr...

A few other structural model studies are relevant. In particular, some papers by Dan and Lynda King are very good (even if they don't address the specific question I was interested in).

King and King (1994) in Assessment (a new journal published by PAR in Florida; [1(3), 275-291]) looked at the latent structure of the Mississippi PTSD scale. They drew three random samples from a dataset of 2200 Vietnam vets, and found evidence for four factors within a broader PTSD factor: 1) re-experience & situational avoidance; 2) withdrawal & numbing; 3) arousal & lack of control; and guilt & suicidality. Although this suggests that perhaps intrusion and avoidance belong (in a psychometric sense) within the same construct, and that numbing might (similarly) be distinguished from more active avoidance, the factor structure obtained here may stem from the Mississippi scale items used in this analysis, and so may say little about the structure of broader DSM diagnostic criteria.

Finally, in the same journal, Koslowsky, Solomon, Bleich & Laor (1994) [Assessment, 1(2), 143-149] used structural equations to test Horowitz' idea that intrusion and avoidance alternate (or, have a bidirectional influence); they used cross-sectional self-report data from 120 Israeli civilians evacuated from their homes during Scud missile attacks in the Gulf war. Their preferred model had threat of attack driving intrusion and avoidance, but only intrusion drove anxiety and outcome functioning. Later on (a year later, in a different group not evacuated from home), the avoidance-anxiety link was better but still not as strong as their alternative intrusion-anxiety model. Longitudinal data might strengthen this conclusion, of course, but these results seem contrary to Horowitz' bidirectional prediction.

More recently, King, King, Gudanowski & Vreven (1995) in Journal of Abnormal Psychology [104(1), 184-196] examined different perceptions and types of Vietnam war zone traumatic stressors in male and female combat vets. Unfortunately, I can't find this paper now and did a lousy summary, but I remember this as an interesting look into trauma stressor characteristics and differing male/female responses to objectively similar wartime traumatic events.

Clinically, the three symptom clusters of PTSD from the DSM diagnostic criteria make lots of sense to me -- but that doesn't mean that they hold up psychometrically. An implication of SEM studies, potentially anyway, is that if we know psychometrically that some PTSD symptoms exacerbate others, we could focus clinical efforts on those aspects -- since without them the others may wither away. Also, the symptoms most in need of our attention may vary with time since the traumatic experience(s), if more avoidant coping skills develop through time or repeated traumas. Currently, we don't have the empirical data to really resolve these questions, but SEM seems like a good way to uncover this clinically important information.

Many thanks to all those on this list who have helped me learn more about trauma.