A study published in the New England Journal of Medicine last month showed that widely-prescribed antidepressants may not be as effective as the published research indicates. After reading about the study in the NYT, I recently read the article and was struck by how well the authors were able to document the somewhat elusive phenomenon of publication bias.
Researchers in most fields can document publication bias only by pointing out patterns in published results. A jump in the density of t-stats around 2 is one strong sign that null reports are not being published; an inverse relationship between average reported effect size and sample size in studies of the same phenomenon is another strong sign (because the only small studies that could be published are the ones with large estimated effects). These meta-analysis procedures are clever because they infer something about unpublished studies from what we see in published studies.
As the NEJM article makes clear, publication bias is more directly observable in drug trials because we have very good information about unpublished trials. When a pharmaceutical company initiates clinical trials for a new drug, the studies are registered with the FDA; in order to get FDA approval to bring the drug to market, the company must submit the results of all of those trials (including the raw data) for FDA review. All trials conducted on a particular drug are therefore reviewed by the FDA, but a subset of those trials are published in medical journals.
The NEJM article uses this information to determine which antidepressant trials made it into the journals:
Among 74 FDA-registered studies, 31%, accounting for 3449 study participants, were not published. Whether and how the studies were published were associated with the study outcome. A total of 37 studies viewed by the FDA as having positive results were published; 1 study viewed as positive was not published. Studies viewed by the FDA as having negative or questionable results were, with 3 exceptions, either not published (22 studies) or published in a way that, in our opinion, conveyed a positive outcome (11 studies). According to the published literature, it appeared that 94% of the trials conducted were positive. By contrast, the FDA analysis showed that 51% were positive. Separate meta-analyses of the FDA and journal data sets showed that the increase in effect size ranged from 11 to 69% for individual drugs and was 32% overall.
One complaint -- I thought it was too bad that the authors did not determine whether the 22 studies that were "negative or questionable" and went unpublished were not submitted ("the file drawer problem") or rejected by the journals. But otherwise very thorough and interesting.
Posted by Andy Eggers at February 23, 2008 2:05 AM