An orchestra of thirty-six instruments plays almost in unison. The conductor tries to pick out which voice is carrying the melody, but the sounds blur together, and each successive bar only adds doubt. That's how researchers studying the hippocampus try to figure out which of its subfields is responsible for memory or thinking, but neighboring subfields look so much alike that teasing apart their contributions is nearly impossible.
To test whether such a link can be found at all, the scientists took data from 638 people in the Cambridge cohort study. They measured the volumes of 36 hippocampal subfields in each participant and matched them against cognitive test results. At first, not a single subfield showed a clear association with the scores: all the signals were too weak and unstable.
So the researchers resorted to a trick. They added an artificial effect to the data, as if they knew for certain that one particular subfield really mattered, and checked whether the analysis could find it. It turned out that even with a strong effect, the right subfield was identified only in some cases. If the effect was weak, as is most often the case in real studies, then even perfectly clean data couldn't reliably point to the correct subfield.
Curiously, the way the data were processed influenced the result more than the presence of correlations. With one analysis method, the target subfield was found more often; with another, less often, even though the data were identical. And more than half of the subfield pairs turned out to be interchangeable: the data couldn't tell one from the other.
Let's return to the orchestra. Even if one instrument really is carrying the melody, a listener in a dense unison won't be able to name it with confidence. The same goes for the hippocampus: as long as effects are small and subfields are tightly linked, precise localization remains more of a dream than an achievable goal.
The condition number of the design matrix was 1727, and the median variance inflation factor was 7.3, meaning that information about many subfields is almost entirely duplicated by their neighbors. At the same time, 32 of the 36 bivariate correlations with the cognitive factor were positive, so a negative coefficient in a multivariate model would have to be interpreted as suppression rather than an inverse relationship.
Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.