I am guilty of writing (and I am revising) a meta analysis of poor quality papers. The reason the editors don’t like or peer reviewers, for that matter, is that the data is old and there is no guidance. That is the point. And no, the paper is not out there: it is in Cochrane revision, which is a form of purgatory.

But this got through Cochrane. It is about as good as you can get, and if you look at the main forest plot you would consider that psychotherapies help borderline personality disorder.

The flow chart is well done, and the criteria well organised. I cannot criticize the authors on the development of this: the meta analysis is done and reported correctly.

But the risk of bias in these papers is high. Particularly not reporting all outcomes in the protocol.

The authors report it this way.

The 33 trials included 1169 participants in the investigated treatment group and 1087 participants in the control group. Seventeen trials targeted patients with BPD diagnosed using a structured clinical interview, 10 trials targeted patients with BPD with recent documented self-harm, and the rest targeted special BPD populations (eg, veterans and individuals with addiction). Twenty-two trials had a stand-alone design, and 11 trials had an add-on design. Twelve trials had women-only samples, and this percentage ranged from 0% to 95% in the remainder. The best-represented approaches were DBT (12 trials), psychodynamic therapies (8 trials), and CBT (5 trials). Twenty-one trials had TAU as the control treatment, and the control treatment was manualized in 10 trials. The treatment developer was an author in 20 trials. In 15 trials, the treatment developer was a therapist or supervised therapists directly. Treatment duration ranged from 2.5 to 24 months, and the number of sessions (for individual and group therapy taken together) ranged from 6 to 312.

The ? statistics indicated high interrater agreement for RoB estimations (eMethods in the Supplement), which were variable (eFigure 2 in the Supplement). Sixteen trials reported adequate sequence generation, 12 trials properly concealed allocation, and 20 trials implemented masking of outcome assessors (2 used self-report measures only). However, for incomplete outcome data, only 13 trials were rated as low RoB, and more than half had high RoB. Eleven trials could be rated as low risk on 3 or 4 domains.

I do not like TAU (treatment as usual) as a comparator. It has the virtue of being cheap, but it leasts to variations in clinician time. Particularly as in some areas the plan for people with borderline is to minimise admissions, emergency reviews, or even therapist time. One cannot control for therapeutic nihilism.

The other problem in the paper is that there is publication bias, which can be corrected by using network analysis, but again, this is in part, on papers where the risk of bias is uncertain.

Finally, the types of therapy are pooled. It would be good to do an analysis by type of therapy. It might matter.