Sex Differences in PTSD: The Response Paper Where I Confused Prevalence with Vulnerability

You have seen a statistic broken down by group, and the number was higher for one of them, and a sentence formed in your head about what that group is like.

That sentence is almost always wrong, and not because the number is wrong. It’s wrong because a rate is a ratio with two moving parts and a measuring device attached, and the sentence in your head quietly assumed all three were fixed.

I wrote that sentence into a first draft, in a course where the readings were sitting right there telling me not to.

The finding

The review’s abstract states two things without apology:

Female participants were more likely than male participants to meet criteria for PTSD — although they were less likely to experience potentially traumatic events.

Then the breakdown by event type. Women were more likely to experience sexual assault and child sexual abuse. Men were more likely to experience accidents, non-sexual assaults, witnessing death or injury, disaster or fire, and combat.

And then the result that closes the easy exits: among victims of specific traumatic events, excluding sexual assault and abuse, female participants still exhibited greater PTSD. The authors conclude that differences in exposure type can only partially account for the differential risk.

The three doors, and why two of them close

When a rate differs between groups, there are exactly three places the difference can come from. My draft went through the first door and stopped.

What my draft did

I wrote, in effect: women have higher PTSD rates at lower exposure, therefore women are more vulnerable to trauma.

Read that as a syllogism and it works. Read it as a measurement claim and it contains a substitution I never declared. The data are about a diagnosis being met. My conclusion is about a person’s capacity to withstand harm. Those are different constructs, and the second is not observed anywhere in the study.

Between them sits a whole chain, and every link is a place the difference could be introduced:

The detail that reframed it

The review notes something about the diagnosis’s own history that I had skimmed on first read and now think is the most important sentence in the introduction.

The codification of post-trauma symptoms into the diagnostic manual was originally prompted by the high prevalence of male Vietnam veterans seeking treatment. The criteria were developed with one population in view. Later epidemiology found the diagnosis was more prevalent among women and girls.

That is not a scandal, and I don’t want to overstate it — criteria get revised, and this one has been, repeatedly. But it is a real measurement fact with a direction: an instrument developed by observing one group’s presentation of a condition will be best calibrated for that group’s presentation.

Applied to another, it may register their distress more readily, less readily, or differently — and any of those shows up in a prevalence comparison looking exactly like a difference in the people.

This is the same defect I’ve written about on the assessment side: a score that reports something about the interaction between a person and an instrument, while being read as a fact about the person.

Where I have to stop myself in the other direction

Having caught one overreach, the tempting move is a second: the difference is all measurement artifact. That is equally unsupported and I want to say so plainly.

The gap survives excluding sexual assault. It appears across twenty-five years and many samples. There are biological and social hypotheses with real evidence behind them, and this course covered some of them — the epigenetics reading in the same week is one line of that work. I’m not in a position to adjudicate. Nor is the review, which is why it says exposure differences account for the gap only partially rather than proposing what accounts for the rest.

So the honest statement is narrower than either of my instincts:

Women meet PTSD criteria at higher rates than men despite lower overall exposure, and differences in the type of event explain only part of it. The remainder is unexplained. Some of it may be differential response to trauma; some may be differential social response after it; some may be how the criteria were built. The data as presented cannot separate these.

That is less quotable than “women are more vulnerable” and it is the sentence the evidence supports.

Why the error is worth publishing

Because of what the wrong version does when it leaves the page.

“More vulnerable” is a trait claim. Trait claims attach to people and travel — into hiring, into who gets assigned the hard account, into whether a person’s distress reads as a reasonable response or a constitutional weakness. And the version I nearly published was sympathetic. I was not trying to disparage anyone. I was trying to take the finding seriously, and the well-meaning version of a trait claim is still a trait claim.

The correction isn’t a matter of tone. It’s that I had inferred something about persons from a measurement of a diagnostic process, and skipped the three social steps in between without noticing they were there.

What I took from it

A rule I now apply to every disaggregated statistic that crosses my desk, and most of them are at work rather than in coursework:

Before explaining why a group’s number is higher, write down what the number counts. Not what it’s about — what it literally counts. Tickets closed. Diagnoses met. Escalations filed. Then ask how many steps sit between that count and the thing you actually care about, and who or what controls each step.

If the answer is “several, and mostly other people,” then the difference in the number is not yet a difference in the group. It’s a research question with a suggestive first result — which is exactly where Tolin and Foa left it, and exactly where I should have.

Thanks to a set of response-paper guidelines that asked for critical thinking rather than agreement, which is the only reason I reread my own first paragraph hard enough to catch it.

Next time a breakdown by group crosses your desk: how many steps sit between the count and the thing you actually care about? I’d be interested in an example where the answer is none.