Discovery of symptom clusters that do not exist. Why factor analysis is useless for understanding PTSD.

CATEGORY: CONTROL OF LANGUAGE AND IDEAS

Robert H. Pietrzak, PhD

Source: Journal of Psychiatric Research
Read time: 2.5 minutes
 
This Happened
In March 2024, a study was published using an eight-factor model of posttraumatic stress disorder (PTSD) symptoms. Previously, seven had been the highest number of factors considered to be the best model. In contrast, for more than three decades, PTSD had been diagnosed as just three clusters of symptoms.
 
Who Did This?
Psychologist Robert H. Pietrzak was the senior author on the study. He is a professor of Psychiatry and of Public Health at Yale University, and works at the U.S. Department of Veterans Affairs, National Center for PTSD. He has over 400 research publications, including at least twelve using factor analysis of PTSD symptoms.
 
The Premise
In this study, researchers collected data from 3,847 veterans on the twenty symptoms in the PTSD diagnostic criteria [1]. They created eight groupings of those symptoms to replicate an 8-factor model that recently had been created by the same research group [2]. Researchers found that three of the eight factors correlated with depression, four of the eight factors correlated with anxiety, and four of the eight factors correlated with suicidal ideation.
The authors concluded that these findings demonstrated validity of the 8-factor model. They suggested that this model may lead to better assessment, treatment, and prevention of PTSD.
The details of this study are summarized very briefly here because they are inconsequential. As the analysis below reveals, this study is emblematic of a larger problem in trauma research.
 
Analysis
This study is one of approximately six hundred studies over the past four decades using the results of a statistical technique to discover or confirm factors of PTSD. The statistical technique is usually some variation of factor analysis.
Briefly, factor analysis works by calculating how frequently each symptom is present with every other symptom. The mathematics then reduces a large number of items into a handful of factors by grouping items that tend to co-occur with each other into a factor, and separating those that co-occur relatively less frequently into other factors.
The premise of factor analysis is that there exists a latent model of a construct that cannot be directly measured. This type of premise is unprovable with current technology. Whether the latent model exists is a theoretical question.
 
Why is this important? When criteria for psychiatric disorders were revised for the 2013 publication of DSM-5, a tragic decision was made that has gone barely noticed. The old three cluster algorithm of symptoms used to make a diagnosis was thrown out and replaced by a four-cluster algorithm. The sole reason for this momentous change was because many factor analysis studies had shown four factors was the best mathematical solution. For the first time in history, factor analysis research was used to change diagnostic criteria.
 
There have been at least seven important reviews of the massive factor analysis of PTSD literature. In the most recent review, I analyzed 206 studies on different metrics of whether factor analysis is a useful technique for designing diagnostic criteria [3]. There were too many interesting findings to summarize here, so, I’ll mention only two highlights.
1. In sixty-six confirmatory factor analysis studies with adults using DSM-5 criteria, researchers found twelve different best-fitting models that ranged from one to seven factors. A four-factor model was best-fitting most often, however, that was in only 59% of studies. With this variety of outcomes, factor analysis is closer to stamp collecting than a valid method of discovering human nature.
2. Results were contingent on which models researchers opted to study. Whenever four-factor models were tested against models with more than four factors (five, six, or seven factors), the four-factor model was best fitting in only 12.8%! Because factor analysis tries to find the model that accounts for the most mathematical variance, the model with a higher number of factors is almost always the best fitting. That is probably why the new 8-factor model was best-fitting in its inaugural test. When someone creates a 9-factor model, that will be best-fitting.
 
So, how many underlying factors really exist in a model of PTSD? Schmitt and colleagues probably have the answer. They conducted perhaps the most rigorous factor analysis possible, trying to avoid methodology pitfalls that have snared less careful researchers [4]. They concluded overall the “results provide greater evidence for a one-factor model.”
 

References 
[1] Stiltner B, Fischer IC, Duek O, Polimanti R, Harpaz-Rotem I, Pietrzak RH (2024). Functional correlates of a novel 8-factor model of PTSD in U.S. military veterans: Results from the National Health and Resilience in Veterans Study, Journal of Psychiatric Research 171:69-74. DOI: 10.1016/j.jpsychires.2024.01.017.
[2] Gross GM, Spiller TR, Duek O, Pietrzak RH, Harpaz-Rotem I (2023). Clinical significance of novel 8-factor model of DSM-5 PTSD in national VA PTSD residential treatment data: Internally- v. externally-cued intrusions, Journal of Affective Disorders, 328:255-260. DOI: 10.1016/j.jad.2023.02.046.
[3] Scheeringa MS (2024). Is factor analysis useful for revising diagnostic criteria for PTSD? A systematic review of five issues ten years after DSM-5. Journal of Psychiatric Research 176:98-107. DOI: 10.1016/j.jpsychires.2024.05.057.
[4] Schmitt T A, Sass DA, Chappelle W, Thompson W (2018). Selecting the "best" factor structure and moving measurement validation forward: An illustration. Journal of Personality Assessment, 100(4), 345-362. DOI: 10.1080/00223891.2018.1449116.