Which is better for my depressed client – CBT or counselling? Or does it make no difference?
As early as 1936, Rozenzweig argued that common factors were so pervasive across different psychotherapies that any differences in outcomes between them would be small. He likened the conclusion of this research to that reached by the Dodo bird from Alice in Wonderland: “Everybody has won, and all must have prizes” (Carroll, 1865/1962, p.412). This so-called Dodo verdict has seemingly been killed and resuscitated with exhausting regularity ever since, but arguably for good reasons. There are certainly numerous studies that have found that one therapeutic approach has achieved better results than another, at least for some clients, at least in the hands of some therapists. However, some authors have argued that these differences can be explained by confounding factors like level of training, supervision, allegiance to the protocol, or therapist personality or ability, rather than the psychotherapeutic “strategy” per se. To the extent that training in diverse approaches is necessary, training costs increase, training programs must be more selective, and client access to qualified providers reduces. Conversely, if the quantifiable benefits of additional training and supervision are minimal, training could be less complicated, allowing more people to be trained, and the burden on skill maintenance among providers to be reduced. These changes might contribute to lower burnout and greater longevity for providers, all of which would improve client access. So, it’s an important question: does the approach we take matter at all? This week we look at an innovative investigation into this question with major depressive disorder.
There has been an accumulating weight of evidence that at least for depression, all psychotherapy approaches are similarly effective, including non-directive, supportive counselling (Cuijpers et al., 2013; Pybis, Saxon, Hill & Barkham, 2017). On the other hand, most clinicians have no doubt had the experience of helping a depressed client with the very same approach that didn’t seem to help another, or that different clients achieved their improvement via different treatment plans. It has been persuasively argued that previous randomised controlled trials are statistically underpowered to detect these client-treatment interactions. The latest would-be assassins of the Dodo verdict are looking to cutting-edge machine-learning statistical approaches with large, real-world data sets to see whether they can identify a set of client variables that would predict which psychotherapeutic approach would be optimal for any given person.
Delgadillo and Duhne (2020) took a data set of 1435 clients from an IAPT site in Northern England who had a primary diagnosis of major depressive disorder or recurrent depression, who had attended at least two sessions, whose baseline PHQ-9 score was in at least the moderate range, and who received either CBT (n=1,104) or Counselling For Depression (CfD) (n=331), a manualised counselling protocol developed within IAPT. The authors had access to a number of potential prognostic indicators (e.g., age, ethnicity, chronicity, number of previous treatment episodes). Each client was classified as a responder or not, based on whether they had achieved reliable change on the PHQ-9 (> 6 point change) and their final score was below the threshold for the moderate depression range. The goal of the study was to develop an algorithm that would predict – on the basis of client variable values – whether that client would respond best to CBT, counselling or would have no optimal treatment.
Machine learning is an artificial intelligence approach to statistical prediction which has the computer program calculate functional rules that link inputs (in this case, client variables) to outputs (in this case, reliable clinically significant change) from a “training” subset of the data (in this case, n=1,085 cases). The effectiveness of the algorithm is then evaluated in the “testing” subset of the data (in this case, n=350 cases). In this study, each case (client) in the test subset had a prognostic index for CBT and CfD. In other words, based on their age, gender, ethnicity and so on, the machine learning algorithm calculated a prediction of their chance of recovery in CBT and CfD. The authors then subtracted the CfD prognosis from the CBT prognosis to give each case in the test subset a net likelihood of greater odds of recovery from CBT (positive values) or CfD (negative values). These predicted odds of recovery were then compared with actual recovery rates.
Six factors predicted prognosis across both CBT and CfD: age (older better), employment status (unemployed worse), disability (people with disabilities did worse with CBT and better with CfD), baseline PHQ-9 (higher baseline depression did worse), baseline work and social adjustment (more difficulties did worse), socioeconomic status of neighbourhood (those in lower SES neighbourhoods did worse in CBT and better in CfD). One prognostic factor was specific to CBT: people from minority ethnic groups did worse than Caucasians. Four prognostic factors were specific to CfD: people with higher baseline anxiety, longer chronicity, lower outcome expectancies and those not on antidepressant medication did better than those with lower anxiety, briefer chronicity, higher outcome expectations and those on antidepressant medication. For CBT, baseline PHQ9 scores made the biggest difference to prognosis (more than twice as important as the next best predictor, unemployment status). For CfD, baseline work and social adjustment made the biggest difference to prognosis (nearly three times as important as next biggest predictor, age).
Interestingly for the Dodo verdict debate, across the test sample, there was no optimal treatment selected for 68% of the sample. CBT was identified as optimal for 13.7% of the sample and CfD for 18.3%. These results suggest that most of the time when clients aren’t being matched to treatment, this will not likely affect their outcomes. However, there remained a sizeable minority (32%) who did show differential treatment response. Within this subset, when clients were assigned to their optimal treatment, their odds of recovery were twice as high as when assigned to their suboptimal treatment. In practice, only 57% of clients who demonstrated differential treatment response were assigned to their algorithm-predicted optimal treatment. This raises the possibility that outcomes could have been much better for the sample had more people received optimal treatment.
Of course, this study has limitations. Arguably, whether counselling or CBT are really equivalent or different for depression should be decided in a randomized controlled trial - and the authors have such a non-inferiority trial on the way. An arguably better design for testing the significance of prognostic indicators is to randomly assign clients to either the treatment indicated by their values on the prognostic indicators or a suboptimal treatment and demonstrate differential responding. On the other hand, such trials are both expensive and ethically challenging, so we can’t rely on such experiments to regularly inform clinical decision making. Innovative designs such as this with plentiful and meaningful real-world data sets should certainly inform service decisions alongside true experiments.
So, the Dodo debate won’t end here. Both sides have ammunition from this study. Clinicians who have experienced that the treatment plan does matter are not just fooling themselves. It does. Sometimes. However, there are a lot of clients for whom it does not matter. And if you – like many of both my student and experienced clinician supervisees – are worried about whether you are doing “just counselling” when you “should be doing manualised psychotherapy”… odds are that – at least with your depressed clients - counselling will not be selling your clients short.
Take home: Clinical Implications
About 30% of clients with major depressive disorder or recurrent depression showed differential response to CBT or counselling. For ~70%, odds of recovery did not depend on which therapy they received.
Of those who show differential responding, those who received the treatment a machine-learning algorithm predicted was optimal had recovery rates twice as high as those who received the suboptimal treatment
Counselling recommended over CBT for people with depression from ethnic minorities, poor socioeconomic status neighbourhoods and who have disabilities
CBT recommended over counselling for people with depression who have shorter chronicity of depression, higher outcome expectations and who were taking antidepressant medication
For the original article, go to:
https://psycnet.apa.org/fulltext/2019-77500-001.html
REFERENCES
Cuijpers, P., Berking, M., Andersson, G., Quigley, L., Kleiboer, A., & Dobson, K.S. (2013). A meta-analysis of cognitive-behavioural therapy for adult depression, alone and in comparison with other treatments. Canadian Journal of Psychiatry, 58: 376-385.
Delgadillo, J., & Duhne, P.G.S. (2020). Targeted prescription of cognitive-behavioral therapy v person-centered counseling for depression using a machine learning approach. Journal of Consulting and Clinical Psychology, 88:14-24.
Pybis, J., Saxon, D., Hill, A., & Barkham, P. (2017). The comparative effectiveness and efficiency of cognitive behaviour therapy and generic counselling in the treatment of depression: evidence from the 2nd UK national audit of psychological therapies. BMC Psychiatry, 17: 215-228.
Rosenzweig, S. (1936). Some implicit common factors in diverse methods in psychotherapy. American Journal of Orthopsychiatry, 6: 412-415.