Source
- Salkovskis, Sighvatsson and Sigurdsson (2024), LIT-538, read in full in NOTE-428. The logical point is on p. 599. The standard of evidence is Kazdin’s (2007) six criteria, set out on pp. 608–609.
- Hofmann and Hayes (2018), LIT-544, read in full in NOTE-437, in the section “The Rise of Mediation and Moderation Studies”.
What was actually shown
This is an argument, not a finding, and both sources state it in their own words.
Salkovskis et al. (LIT-538) put it plainly: “The effectiveness of a treatment is not evidence for the validity of the theory on which it is based with one important exception”. The exception is a treatment that demonstrably changes a factor crucial to the theory and does not change the target problem. That disconfirms the theory. Successful, theory-informed treatment is only “consistent with” its theory (NOTE-428, C1). Their example of the error is exposure. It was adopted as the mechanism of anxiety treatment because it was common to the treatments that worked, and then confirmed from the fear reduction it was invoked to explain. They call this circular, and say it “has in our view held the field back”. What would count instead are Kazdin’s criteria: the treatment changes the mediator, and the mediator then changes the outcome; the relation is specific; other mediators are ruled out experimentally; and there is consistency, a gradient and more than one source of data. They also say mechanism research has mostly failed these criteria, above all on temporal precedence (C8).
Hofmann and Hayes (LIT-544) state both directions. “Even if the intervention works well, if the specified process of change cannot be shown to be consistently applicable, the underlying model is incorrect and needs to be set aside.” And: “if a procedure fails to alter putatively critical processes of change … then the model remains untested even if the procedure fails”. Their admission rule asks for both legs: a procedure “reliably produces gains and manipulates a process that mediates these gains”. They grant that “a statistically significant mediator is not synonymous with a generative mechanism because third variables can always exist” (NOTE-437, Assumptions).
Why it is close to logical. Every theory of change on offer predicts that the treatment it informs will work. So the treatment’s working cannot tell those theories apart, and a common factor present in every working treatment cannot either. What can tell them apart is the process each names: whether it moves, whether it moves first, and whether it moves the outcome more than the rival’s process does. The asymmetry follows. A treatment that moves the process and leaves the outcome alone is the case the theory forbids. A treatment that fails to move the process has not tested it.
The two sources come from the two wings of CBT, Beckian and contextual behavioural, and they disagree about the answer: one mechanism for all effective therapy (LIT-538) against many processes in a hierarchy (LIT-544). They agree on this standard. That agreement across a rivalry is why the account is Active.
What this does not say
- It does not say mediation proves a mechanism. Both sources deny it. Mediation is necessary evidence and not sufficient. Third variables, reverse causation and a process measured after the outcome began to move all defeat it. The ACT review (LIT-555, NOTE-431) is a worked example: its mediation studies are the right kind of test, but its authors note that process measures were often taken “after outcomes began to improve significantly”.
- It does not say efficacy is unimportant. Efficacy is evidence that a treatment works. It is not evidence about why.
- Disconfirmation is not automatic. A process that moved without the outcome may have been measured badly, moved too little, or be one of several the theory requires. Hofmann and Hayes allow “short delays while measurement issues are worked out”. This is the ordinary point that a test bears on a theory only together with its auxiliary assumptions. The record holds no philosophy-of-science reading on confirmation that states it, so it is made here without a source.
- It takes no side on which theory of change is right. That is THEORY-058’s and THEORY-050’s business, and each is Proposed because its own mediation evidence is not yet in.
How it bears on causation and mind
- The standard is interventionist. “The treatment changes the mediator, which then changes the outcome” is causation as difference made under intervention, the notion Borsboom uses for symptom networks (LIT-547) and that Pernu (LIT-147) and Hazelwood (LIT-156) test against exclusion and composition. Hazelwood’s condition bears on mediation directly: a mediator that is partly composed of the outcome, as a self-report of distress-related avoidance may be of distress, is not independently fixable, and its “mediation” is not causal evidence. That application is the record’s.
- It is a mental-causation claim in clinical form. A theory of change names a psychological process (a belief, a meaning, a way of relating to thoughts) as what produces change. This account says how such a claim is tested. It does not decide whether such processes are causes in their own right or by identity with their realisers (THEORY-048).
- A parallel elsewhere in the record. THEORY-023 finds that behavioural evidence cannot settle AI consciousness because every rival theory predicts the behaviour. The structure is the same: evidence every candidate predicts discriminates none of them.
Connections
- Macri and Rogge’s 2024 meta-analysis of ACT’s processes as mediators (LIT-543) is the largest test of this kind in the batch. It is unread.
- Both sources open with Gordon Paul’s question as the field’s founding statement. Hofmann and Hayes use its 1969 form, which adds “and how does it come about?”; Salkovskis et al. use the 1967 form without it (NOTE-428). This account is what that clause demands of an answer.
Note of 2026-10-29: RefraMED’s mechanism analyses, a worked case
The RefraMED trial’s NIHR monograph, LIT-1838 (NOTE-1421), reports causal mediation analyses that were meant as the evidence this account asks for, and they meet neither leg of it.
- Not specific. The mediators are sessions attended, therapeutic alliance and DBT skills. Alliance and exposure are common factors, not RO DBT’s theorised process. The processes RO DBT names, flexibility and emotional expression, are reported only as outcomes compared between arms and are never entered as mediators.
- No precedence. Mediator and outcome are both measured at 7 months, and the planned analyses of temporal order are not reported.
- Confounding not handled. The instrumented analysis, the part designed for it, failed on weak instruments.
It is a worked example of this account’s point that mechanism research has mostly failed on temporal precedence, and of Hofmann and Hayes’s case of a model left untested. It does not disconfirm RO DBT’s theory: the processes did move, and whether the outcome moved through them was not estimated.