Here is a contradiction worth sitting with.
One of the field's most-cited early studies on prolonged exposure looked hard for reasons to withhold the treatment from difficult patients—and found none. Depression, guilt, shame, anger, personality pathology: none of it reliably predicted who would fail or drop out. The authors concluded, reasonably for their data, that excluding patients from exposure based on their profile "is not justified."
That paper is still used to wave away caution. Give everyone the gold-standard therapy; the pre-treatment red flags don't hold up.
Now put a civilian trial next to it. In a randomized comparison of prolonged exposure and interpersonal psychotherapy, half the civilians with comorbid major depression assigned to PE dropped out, per Markowitz et al. Among those without depression: 6 percent.
Fifty versus six. That is not statistical noise. That is a treatment shedding the sickest patients before it can help them.
Fifty versus six. That is not statistical noise. That is a treatment shedding the sickest patients before it can help them.
So which is it—does comorbid depression matter or not? The answer is that it depends entirely on who is in the room, and the reassuring "gold-standard" language papers over exactly that difference.
Start with the arithmetic the guidelines don't lead with. A 2024 meta-analysis pooled 86 studies and 7,894 patients and found a weighted average nonresponse rate to first-line psychological treatments of 39.23 percent. Nearly four in ten patients finish the recommended therapy still carrying clinically significant PTSD.
In military samples the numbers are worse—Steenkamp et al reportedly put nonresponse between 50 and 72 percent. A residential UK sample of 960 veterans found more than a quarter fell into a "treatment-resistant" class showing essentially no change.
We call this gold-standard because the group-level effect sizes are genuinely large. But an effect size is an average, and an average can be impressive while a big minority gets nothing. The confidence rests on a statistical framing that reports the mean and stays quiet about the tail.
Here is the important part: that tail is not random. The same predictors keep surfacing across independent datasets, which means failure is often visible before treatment begins.
Take the depressed patient again. In that PE-versus-IPT trial, comorbid depression didn't just drive dropout—among those who stayed, only 30 percent responded to prolonged exposure, per Markowitz et al. In IPT, the response rate for depressed patients was 50 percent, and dropout was 20 percent rather than 50, per Markowitz et al.
Read that carefully. For the depressed patient, the trauma-focused therapy was both harder to tolerate and less likely to work than the non-trauma-focused alternative built originally for depression.
The meta-analytic picture converges. Across those 7,894 patients, PTSD symptom severity, comorbid depression, and baseline depression scores all emerged as significant predictors of nonresponse. The UK residential study reportedly found depression, anxiety, and having held a combat role all raised the odds of landing in the treatment-resistant class—and higher pre-treatment reexperiencing symptoms raised those odds more than fivefold.
A national VA sample of 2,715 veterans in residential care adds texture: worse outcomes tracked with comorbid personality disorder, greater pain severity, and—tellingly—a pending disability-compensation claim, while more social and psychological resources predicted better response.
Low social support keeps recurring. So does chronic and early-life trauma, severe baseline symptoms, and comorbid substance use. These are not exotic markers requiring special instruments. They are things you already know about a patient by the end of intake.
There is even a plausible mechanism for why some patients can't do the core work of cognitive therapy—and it's not motivation.
Cognitive restructuring asks a patient to take new information ("you didn't decide to be attacked") and update an old belief ("I'm to blame"). That updating is a specific neural operation, and in some brains it appears broken.
Researchers reportedly scanned 136 people—70 with PTSD, 66 trauma-exposed without it—while they challenged negative self-beliefs inside the scanner. In the PTSD group, the prefrontal cortex was worse at regulating the thalamus, the relay hub that lets brain regions represent and revise abstract information like self-concept.
And within the PTSD group, the patients with the most severe negative beliefs showed the weakest connectivity in exactly that pathway during restructuring. The people whose beliefs most need to shift are the ones whose belief-updating machinery is least able to shift them.
This is the clinical face of that finding: patients who understand, intellectually, that they weren't to blame, while the understanding "never quite shifts the part of them that still feels responsible." We hand these patients a therapy whose central engine is belief revision and then record their failure to revise as a poor response.
There is a parallel deficit on the social side. A case-control study of nearly 200 people found some evidence that hypomentalising—difficulty reading mental states, as measured via the Reflective Functioning Questionnaire—was associated with more than fourfold higher odds of PTSD symptoms. If a therapy depends on reflecting on your own and others' minds, and that capacity is impaired, the therapy is building on sand.
Even medication history carries signal. An early PE study reportedly found indications that benzodiazepine use was linked to both worse outcome and dropout—one reason a patient's prescription list belongs in the triage conversation, a point developed in the companion piece on benzodiazepines. The medication in someone's history is itself a predictor of how they'll fare in psychotherapy.
Now, honesty requires naming the friction. That same early PE study, and later work, found many pre-treatment variables to be weak, unreplicated predictors at the individual level. You cannot look at one depressed patient and say with confidence that PE will fail them.
But "we can't predict any individual's outcome with certainty" is not the same claim as "we should route everyone into the same pipeline." The first is true. The second doesn't follow.
What the data support is probabilistic triage, and probabilistic triage is what medicine does everywhere else. When a marker reliably shifts response rates—when depressed patients respond to IPT at 50 percent and to PE at 30, and drop out of PE at 50 percent versus 20—that is enough to change the default.
So the honest response to a 39 percent nonresponse rate is not to defend prolonged exposure harder. It is to sort before treating.
Route the depressed patient toward IPT, where depression stops predicting dropout and response rates hold up. Don't send them first into the therapy their comorbidity most reliably breaks.
For patients whose beliefs are "stuck"—severe negative self-cognitions, the neural signature of poor belief-updating—consider emotion-regulation or mentalising-focused preparation before, not instead of, trauma-focused work. Build the capacity the therapy assumes, rather than testing the patient against a capacity they don't yet have.
And honor preference. A doubly randomized preference trial found that higher avoidance and being medication-naïve reportedly predicted poorer outcomes specifically for patients who didn't get their preferred treatment. Preference is not a courtesy; it is a predictor.
None of this is a case against prolonged exposure. For the patient without comorbid depression, PE dropout ran at 6 percent per Markowitz et al, and the trauma-focused therapies remain the most effective tools we have for a large share of patients.
The case is against the reflex of treating the average as the mandate. The gold-standard framing lets a field congratulate itself on a strong mean while a predictable minority—disproportionately the depressed, the severely symptomatic, the socially isolated veteran—cycles through a treatment we could have guessed would fail them.
We usually only find out who they are after we've lost them. The data say we could often know sooner. The question is whether we're willing to sort at the front door instead of counting casualties at the back.