causality bites, epidemiological edition

From CNN.com:

"A study of nearly 2,800 British adults and children bolsters the evidence that patients are prone to suicidal impulses when they are first put on antidepressants... Study bolsters antidepressant-suicide link ... The study looked at four drugs and found that suicidal thoughts or attempts were four times more likely during the first 10 days of treatment than they were after three months. Suicide was almost 40 times more common early on than later in treatment, though there were only 17 suicides, all in patients older than 19."

Yikes, if this study is as it's described, then it really doesn't provide anything anywhere close to the strength of evidence touted by this article. The retort from two psychiatrists, that the result "simply means that antidepressants are being prescribed for the right indication, and that they do not immediately eliminate suicide risk," is buried ten paragraphs down and not explained, even though it would seem exactly on point.

Just think about it in the abstract. Say somebody is severely depressed, worse than they have ever been before. If you had to bet, would you bet they are going to be have a higher level of depression (a) tomorrow or (b) three months from now?

You could say this about every chronic illness that exhibits considerable fluctuation: if you had to guess, people are going to be more like they were yesterday than they were three months ago. When someone initiates a treatment for depression, or when the decision is made to initiate a medication regimen, they are probably near the bottom of their overall distribution of affective fluctuation. So we have two general principles: (1) that people tend to be bad off when they seek treatment or new treatments are tried and (2) that people who are bad off have a greater probability of getting better than getting worse (there are certainly illnesses for which this is not true. Depression is not one of them.)

You really can't know without a clinical trial. The study is not a clinical trial. It's hard to imagine an ethically palatable clinical trial. But this isn't even a strong non-clinical trial. At least as reported. What the study really needs to do is, in the absence of a placebo condition, really do what it can to try to figure out how one might estimate how many participants one would have committed suicide in the absence of a placebo condition. The actual study might try to do this; it's a very hard inferential problem.

The kicker here is to imagine what would happen if the antidepressant really did have a therapeutic effect within the time window that the medications claims (that is, 4 or so weeks, rather than within 2 weeks). Then the above stated results would look even worse! There would be an even stronger call not to give antidepressants to teenagers! In other words, if one adhered the above reasoning, the more effective an antidepressant was, the more dangerous it would seem!