As usual he's provided a cogent and thorough analysis. Frankly, I agree with him that this is both a case of bad media reportage of science and a less than stellar study.
Aside from the fact that this is an observational study, with unexamined confounders, the correlation is actually quite weak. Quoting Ben Goldacre:
Then if you read the academic paper you find that the associations reported are weak. For the benefit of those who understand “regression” (and it makes anybody’s head hurt), 18% of the variance in the LSHS score is explained by gender, age and stress. When you add in caffeine to those three things, 21% of the variance in the LSHS score is explained: only an extra 3%, so caffeine adds very little. The finding is statistically significant, as the researchers point out, so its unlikely to be due to chance, but that doesn’t affect the fact that it’s still weak, it explains only a tiny amount of the overall variance in scores on the “predisposed-to-hallucinations” scale.
http://www.badscience.net/2009/01/drink-coffee-see-dead-peopleSo while there was a statistically significant correlation between those who reported high caffeine consumption and those who scored high on the Launay-Slade Hallucination Scale (LSHS), it was a relatively weak one. It's also important to remember that the study group was only 219 college students who filled out a survey. It was a small, self-selected cohort and we have no way of knowing if they are representative of the general population.
Also, the researchers apparently drove the media reporting by putting an unfounded figure into their press release:
Lastly, most newspapers reported a rather dramatic claim, that 7 cups of coffee a day is associated with a three times higher prevalence of hallucinations. This figure does not appear anywhere in the paper. It seems to be an ad hoc analysis done afterwards by the researchers, and put into the press release, so you cannot tell you how they did it, or whether they controlled appropriately for problems in the data, like something called “multiple comparisons“.
Here is the problem. Apparently this 3 times greater risk is for the top 10% of caffeine consumers, compared with the bottom 10%. They say that heavy caffeine drinkers were three times more likely to have answered affirmatively to just one LSHS question: “In the past, I have had the experience of hearing a person’s voice and then found that noone was there”.
Ben Goldacre quite rightly points out that this is likely a case of the Texas Sharpshooter Fallacy. We don't know why the researchers chose 10% as their magic number. Perhaps their reasoning is justified, perhaps not. We don't know though because they didn't put it in the paper. Again, Goldacre is correct in judging this to be a subversion of the peer-review process. It's science by press-release.