Moderation, mediation, and moderated mediation

P. Cuijpers1 links a focus on symptom reduction in psychotherapy research to the omnipresence of diagnostic systems such as DSM and ICD that are based on symptom clusters. However, diagnostic systems come and go, usually in an upward spiral. I trained under DSM‐II that was largely guided by dynamic theory. Symptoms represent a form of psychic distress or behavioral dysfunction that is worthy of change on its own merit, whether we lump them together under a unifying theory (DSM‐II) or split them apart (DSM‐III on).

The DSM‐5 has at least made some effort to introduce the notion of dimensionality into the discussion, and the Research Domain Criteria (RDoC) project takes that a step further by focusing on the presumed underlying mechanisms that drive the various disorders, although perhaps from an overly reductionistic perspective, given that the vast majority of people suffer from diagnosable disorders with lower heritability than political preference. Kidneys do not learn, but brains evolved to interact with and be modified by the environment, and we ignore the influence of learning and culture at our peril.

I wholly endorse the call for broadening the targets of treatment. Patients often come to treatment looking for change in their capacity to function or their quality of life, and anything we can do to address those concerns is laudable. I remind my patients that “I work for you, you do not work for me” and I mean that literally. Nonetheless, we often work on multiple goals in treatment; not only was cognitive behavior therapy (CBT) as efficacious as antidepressant medication (ADM) in one of our earlier trials, but it also got more patients back to work2. CBT also cuts risk for relapse by more than half relative to ADM following treatment termination3.

P. Cuijpers is in the vanguard of one of the most interesting developments in recent clinical science. What he has been doing is collecting individual patient data from controlled trials in the treatment of depression and using the aggregated data to test for moderation in samples that are exponentially larger than can be collected in any given trial. He has shown that severity does not moderate differential response to CBT versus ADM4, despite the fact that ADMs separate from placebo only among patients who are more severe5. His individual patient data meta‐analyses can inform the use of machine learning to generate treatment selection algorithms that identify the optimal treatment for a given patient6. This is the essence of precision medicine.

Nonspecific processes account for the lion's share of change in depression, and this is likely true to a lesser extent for most other nonpsychotic disorders. Cuijpers accurately points out that most of the supporting evidence is purely correlational in nature and thus a weak basis for drawing a causal inference, but he himself has provided some of the most compelling evidence for a causal role for such processes. What he did was to conduct a meta‐analysis7 in which he used within‐group change in minimal treatment controls to establish the proportion of variance in change that could be attributed to spontaneous remission, and comparisons both within and between such controls to nonspecific and specific interventions to carve out the rest. He found that about one third of the change in depression was a consequence of spontaneous remission, about half was due to nonspecific factors that would occur in any given treatment, and only about a sixth was due to the specific effects of presumably “active” treatments.

Because all of the studies that he included in his meta‐analysis were randomized controlled trials, he could legitimately draw a causal inference with respect to the nonspecific factors. After decades of process research that sought to determine how treatments work, but could not answer the question as to whether they actually do work (have a causal effect), Cuijpers has provided a most compelling answer and a very clever roadmap for others to follow.

I do think, however, that it is premature for Cuijpers to conclude that there is no evidence that CBT works through cognitive change to produce change in depression. As he points out, the problem is that it is easier to detect an effect than it is to explain it, largely because we can use powerful experimental methods to test for causal effects of treatment on both the purported mediator and the outcome, but are left to rely on purely correlational methods to try to draw a causal inference regarding the link between mediator and outcome. That being said, I think he is wrong when he asserts that the absence of specificity denotes an absence of causal effect. If cognition did not change over the course of CBT then it could not be a mediator, but the fact that it shows comparable change in ADM does not rule such a causal process out.

The problem is that a given process can be both a cause and a consequence of change8. In an earlier trial we found that change in depression‐relevant cognition predicted subsequent change in depressive symptoms with CBT but not with ADM, which likely worked through other causal mechanisms. The issue is one of moderated mediation in which the treatment affects the nature of the relation between the purported mechanisms and the outcome9. While CBT produces change in cognition that leads to (mediated) subsequent change in depression, ADM produces change in depression through other mechanisms that lead to subsequent change in cognition. Absolute change in cognition was comparable between the two treatment modalities, but the causal paths that led to that change were likely quite distinct.

Whereas moderated mediation as a consequence of differential treatment tends to obscure mediational effects that might be present, because it alters the apparent relation between the mediator and the outcome, moderated mediation as a function of individual differences among patients can be used to amplify that signal. As Kazdin10 first pointed out, any instance of moderation suggests that different causal mechanisms may be at work in different patients. This means that tests of mediation can be made more precise (and therefore more powerful) if we include patient by treatment interactions in those analyses.

I agree with Cuijpers that mediation is difficult to detect, but a more sophisticated approach that takes moderated mediation into account may help to clarify the process.