3 ms·
A 30% error rate is far, far too large for this to be useful as a first line filter. Even error rates of less than 1/3 of that (10%) lead to combinatorial effec
by shawn-furyan 12y ago
A 30% error rate is far, far too large for this to be useful as a first line filter. Even error rates of less than 1/3 of that (10%) lead to combinatorial effects that make for very unreliable tests. At best, this would be a corroborating test, and not a very strong one at that.
- rflrob 12y agoWhat I, if I were a reviewer on this paper, would have liked to have seen is a ROC curve—basically, if you tweak the cut offs, can you get fewer false positives at the expense of more false negatives. Assuming fraud in science is rare (I think even the gloom-and-doom-iest reports put it under 20% for any kind of research misconduct, not just fraud), even a moderately low false positive rate will flag a lot more papers than a perfect true positive rate.
- a3n 12y agoThink of it as early days.