Author: Jim Young
Position: Biostatistician
Institution: Basel Institute of Clinical Epidemiology, Switzerland
E-mail: jyoung@uhbs.ch
Submitted Date: February 28, 2008
Published Date: February 28, 2008
This comment was originally posted as a “Reader Response” on the publication date indicated above. All Reader Responses are now available as comments.

Kirsch and co-authors conclude that “The relationship between initial severity and antidepressant efficacy is attributable to decreased responsiveness to placebo among very severely depressed patients, rather than to increased responsiveness to medication.”

I don’t think this conclusion is warranted because of the possibility of regression to the mean. This occurs when patients with a high level of some attribute (here depression) are recruited for a trial but there is a degree of random error in the measurement of that attribute. Those recruited will tend to have a higher random error component. If they are measured again some time later, this random component will be lower on average and level of the attribute in these patients will then approach the average in the wider population (hence the expression ‘regression to the mean’) [1].

I think the authors misinterpret Figure 3. Assume the following scenario: (1) that treatment is more effective in patients who are more depressed at the start of a trial; (2) that there is a placebo effect but this effect is no greater in patients who are more depressed at the start of a trial; and (3) that measurement of depression is subject to random error. Figure 3 is entirely consistent with this scenario – a scenario that is very different from the authors’ conclusion. Those patients who are more depressed and receiving a placebo still improve (because of the placebo effect) but on average the random error in their second measurement is lower (compared to less depressed patients) – so for patients on placebo, the line decreases. Those patients who are more depressed and receiving treatment improve more than patients who are less depressed, but do not appear to improve more because on average the random error in their second measurement is lower – so for treated patients, the line remains level.

In fact graphs such as Figure 3 are recommended for detecting regression to the mean [2]. Assume treatment has a constant effect relative to placebo regardless of initial disease severity. In this situation, regression to the mean leads to a plot of improvement versus initial severity with parallel lines (for treatment and placebo) that decrease with increasing initial disease severity (see [2], Figure 3). The distance between the (parallel) lines shows the constant treatment effect.

So all one can conclude from Figure 3 is that with increasing initial depression, the treatment effect increases relative to placebo. That is the nature of randomised controlled trials – they are comparative. It is not possible to say whether treatment is more beneficial or placebo less beneficial with increasing initial depression – only that relative to placebo, the effect of treatment increases as patients become more depressed.

[1] Bland JM, Altman DG. Statistical notes: Some examples of regression towards the mean. BMJ 1994; 309:780 (http://www.bmj.com/cgi/co...).

[2] Barnett AG, van der Pols JC, Dobson AJ. Regression to the mean: what it is and how to deal with it. Int J Epidemiol 2005; 34: 215-220 (http://ije.oxfordjournals...).