I want to recommend this analysis (at the Mind the Brain blog) of a new paper in PLOS Medicine. The paper is an analysis from a very large Swedish data set of possible links between SSRI antidepressant use and violent crime, which is a contentious topic. As the post by James Coyne notes, there are plenty of groups on both sides of this issue with axes to grind, and very quickly you find yourself knee-deep in conflicting claims from trial lawyers, drug companies, Scientologists (and other opponents of psychiatry), gun-control advocates (and opponents), public health administrators, and more.
The study itself looks to be the best of its kind that we’re likely to get, in both size and quality of data: individual medical and criminal records on more than 850,000 patients (out of Sweden’s total population of about 8 million, 15 and over). Coyne does an excellent job of digging into this paper, and illustrating how it tries to deal with the many confounding variables. For example, right up front, the medical conditions that lead to prescriptions for SSRI drugs are likely to affect an individual’s likelihood of being arrested for violent acts. And there are plenty of more subtle effects: one analysis of this same data set suggests that individuals with a sibling diagnosed with depression are themselves statistically slightly more likely to have committed a violent crime, even after controlling for sex, age, income level and so on.
The paper itself, when everything has been worked through, finds a small association between SSRI use and violent crime, but only in one age group (15 to 24 years old), and only in the lowest-dose group (i.e., no dose-response). This could be an actual causal relationship, but looks equally likely to be the result of something else:
In the age bracket for which this association between antidepressant use and conviction of a violent crime was significant, antidepressant use was also associated with an increased risk of violent crime arrests, non-violent crime convictions, and non-violent crime arrests, using emergency inpatient and or outpatient treatment for alcohol intoxication or misuse. . .
. . .Perhaps one of the greatest weaknesses of this data set is its limited assessment of alcohol and substance use and abuse. For alcohol, we are limited to emergency inpatient or outpatient treatment for alcohol intoxication or misuse. For substance abuse, we have only convictions designated as substance-related. These are poor proxies for more common actual alcohol and substance use, which for a variety of reasons may not show up in these administrative data.
There’s also the problem that SSRIs are prescribed for a variety of heterogeneous conditions: unipolar depressive disorders, certainly, but also the depressive phase of bipolar disorder and others. On the behavioral end, we have only arrests as a proxy for violent acts – useful, but imperfect as well. So the answer to the question “Is there an association between SSRI use and violent crime?” could be (at different levels of detail): “Yes” “Possibly” “Nothing that stands out much” and “Maybe just in one particular age group, but maybe not, too”.
The authors of the PLOS Medicine paper have done an excellent job laying out all these issues, but I’m willing to bet that this doesn’t translate very well to a short headline. The various axe-grinders involved with this issue will find no reason to stop grinding, and no doubt all sides will cite this study as proving their points. And I bring this up on this site as an example of how to approach medical statistics in general, such as data from large clinical studies. Those are certainly more controlled than population-level statistics, even ones as comprehensive as those collected by the Swedish authorities. But they have plenty of confounding variables of their own, still. We set up these trials, from the beginning, to try to provide definitive answers (a big difference from observational studies such as this one), but even tying things down in this fashion still leaves plenty of room for the numbers to thrash around.
And this work also illustrates the trickiness of observational studies themselves. This is as large and granular a data set as you could possibly wish to see here in the real world, and there are still all sorts of things that aren’t accounted for. Keep that in mind next time you see a headline (especially a headline) about an association between wearing plaid and being diagnosed with psoriasis or whatever. Observational studies are very blunt tools with which to carve fine distinctions, and the great majority of them are probably reporting either noise or the effect of confounding variables.