Intended for healthcare professionals

CCBYNC Open access

Rapid response to:

Research

Response to acute monotherapy for major depressive disorder in randomized, placebo controlled trials submitted to the US Food and Drug Administration: individual participant data analysis

BMJ 2022; 378 doi: https://doi.org/10.1136/bmj-2021-067606 (Published 02 August 2022) Cite this as: BMJ 2022;378:e067606

Rapid Response:

Data modelling in search of meaning

Dear Editor

Like previous meta-analyses, the FDA individual patient data analysis of antidepressant trials reveals a small average difference between antidepressants and placebo of 1.75 points [95% CI: 1.63 to 1.86] on the HAM-D, a difference likely not clinically important.[1–4]

The authors used finite mixture modelling with three normal sub-distributions of participant responses from the data without a pre-registered protocol. Their Figure 2 shows that distributions of change scores for antidepressants and placebo differ somewhat, but substantially overlap. It is not clear that this pattern demands post-hoc de-composition into new, normalised distributions and the model selected highlights the areas of the distributions that show the greatest differences between the groups.

Mixture modelling was also applied to antidepressant research by Thase et al. who obtained [5] similar results with a two group solution. However, this model was obviously flawed because the “benefiters” and “non-benefiters” distributions were largely overlapping.[6] In this new meta-analysis, the fit of a third distribution (labelled as “non-specific”) artificially gives the impression that this problem has been addressed as there is less overlap between the categories labelled as “large” or “minimal”. But this is misleading as these two distributions largely overlap with the “non-specific distribution”.

It is not clear that focusing on change from baseline scores in undefinable sub-populations of patients is informative. As these are theoretical distributions, each participant has a probability of belonging to each distribution so that the technique does not identify a group of people who show a “large” response or one that benefits more from antidepressants (meaning the study is unable to inform an RCT to test antidepressants in any sub-group). This is also why it is inappropriate to generate a NNT, a calculation that requires sharp categorisation, but is not applicable to probabilistic ‘clouds’ of patients.

Indeed, the categories defined by the authors (i.e.“non-specific”, “large”, and “minimal”) are semantic labels, which may be misleading as they imply a causal relationship to the medication – but establishing this requires a direct comparison between drug and placebo in randomised groups. Indeed, the fact that the trimodal pattern exists in the placebo group as well as the drug group could imply that it is not a drug effect but may indicate different susceptibility to the placebo effect (among those taking the drug and placebo). Additionally, unblinding by side effects (or lack of such) might amplify this effect for the drug group.[1] The authors argue that unblinding is unlikely but the medications that show the strongest effects in this analysis are known to have pronounced side effects. We know unblinding can occur in antidepressants trials [7,8] and that treatment guesses are associated with outcomes independent of specific pharmacological effects.[9]

Other limitations of antidepressant trials remain unaddressed [1,2]: including withdrawal effects from pre-trial medication, short duration (4-12 weeks), and the possibility that antidepressants produce changes in depression scores via numbing or other super-imposed psychoactive effects, without actually producing a change in the underlying disorder or symptom mechanism.[10] The authors argue that a “large” response may represent a true effect on depression rather than mere symptoms attenuation but the present study is unable to evidence this.

In conclusion, the exploratory post-hoc modelling of pre-post differences should not be confused with the comparison of drug-placebo differences on pre-specified outcomes in RCTs, appropriate for causal inferences. Following this logic, the strongest result of this meta-analysis is the very small average differences observed between groups. Even if we accept the premise that 15% of people benefit, 85% of people do not and given the adverse effect burden (and considerable uncertainty about this 15%) it is not clear that this justifies current levels of antidepressant prescribing.

References

1 Munkholm K, Paludan-Müller AS, Boesen K. Considering the methodological limitations in the evidence base of antidepressants for depression: a reanalysis of a network meta-analysis. BMJ Open 2019; 9: e024886.
2 Horowitz M, Wilcock M. Newer generation antidepressants and withdrawal effects: reconsidering the role of antidepressants and helping patients to stop. Drug Ther Bull 2022; 60: 7–12.
3 Leucht S, Fennema H, Engel R, Kaspers-Janssen M, Lepping P, Szegedi A. What does the HAMD mean? J Affect Disord 2013; 148: 243–8.
4 Moncrieff J. Against the stream: Antidepressants are not antidepressants – an alternative approach to drug action and implications for the use of antidepressants. Psychiatrist 2018; 42: 42–4.
5 Thase ME, Larsen KG, Kennedy SH. Assessing the “true” effect of active antidepressant therapy v. placebo in major depressive disorder: use of a mixture model. Br J Psychiatry 2011; 199: 501–7.
6 Naudet F. Can a ‘true’ effect be built on a ‘wrong’ model? Br J Psychiatry 2012; 200: 512–512.
7 Lin Y-H, Sahker E, Shinohara K, et al. Assessment of blinding in randomized controlled trials of antidepressants for depressive disorders 2000-2020: A systematic review and meta-analysis. EClinicalMedicine 2022; 50: 101505.
8 Scott AJ, Sharpe L, Colagiuri B. A systematic review and meta-analysis of the success of blinding in antidepressant RCTs. Psychiatry Res 2022; 307: 114297.
9 Chen JA, Papakostas GI, Youn SJ, et al. Association between patient beliefs regarding assigned treatment and clinical response: reanalysis of data from the Hypericum Depression Trial Study Group. J Clin Psychiatry 2011; 72: 1669–76.
10 Moncrieff J, Cohen D. Do antidepressants cure or create abnormal brain states? PLoS Med 2006; 3: e240.

Competing interests: MH was a reviewer on the original paper and some of the points above were included in the original review but not incorporated into the final paper. MH is a co-founder of Outro Care, a Canadian service aimed to help people stop unnecessary antidepressants. JM receives royalties from 5 books about psychiatric medications. She is co-applicant on the NIHR-funded REDUCE grant examining supported reduction of antidepressants. FN, MP, and JJ report no conflicts of interest.

08 August 2022
Mark A Horowitz
Clinical Research Fellow in Psychiatry
Florian Naudet, Janus Jakobsen, Martin Plöderl, Joanna Moncrieff
North East London NHS Foundation Trust and UCL (honorary)
Goodmayes Hospital, Barley Lane, Ilford, IG3 8XJ.