3 comments

[ 4.9 ms ] story [ 20.1 ms ] thread
Interesting: "After FDR adjustment, significant differences in 13 of 19 RVPs were found between the AD and HC groups (Table 2 and Figure 2). Logistical models combining parameters were created for combined AD classification. A logistic model combining these 13 RVP’s provided good classification performance (81.2% sensitivity, 75.7% specificity and 87.7% AUC), compared with the logistic model including only age and APOE E4 carrier status (68.0% sensitivity, 61.8% specificity and 63.7% AUC)."

Despite the impressive improvement in sensitivity and AUC, this is a small and presumably preliminary study. Additionally, the training and testing sets appear to be the same.

Despite being on nature.com, the journal this is in ("Translational Psychiatry") is not a prestigious one; in fact, it's an online-only journal and you have to pay to get your article published in it. It is suspicious that they couldn't find a better journal to publish such an amazing result in, so the science is probably bad. The statistical significance of p=0.01 is unimpressive.

Honestly, as non-scientists we should just ignore new papers like these. If it's really important, first it'll be published in an excellent journal, then it'll be replicated successfully, then there'll be meta-analyses that confirm how reliable the technique is, and it'll get cited and built upon a bunch of times, and then it'll actually be something that we should consider as anything other than "probably wrong". Which is what it is right now.

I'm disappointed no one has come up with a BAD SCIENCE browser extension.