Skip to main content
Cross Validated

Timeline for When are ROC curves to compare imaging tests valid? (Focus on the example below)

Current License: CC BY-SA 4.0

42 events
when toggle format what by license comment
Dec 28, 2020 at 14:47 vote accept Antoni Parellada
S Dec 28, 2020 at 14:47 history bounty ended Antoni Parellada
S Dec 28, 2020 at 14:47 history notice removed Antoni Parellada
Dec 27, 2020 at 22:08 answer added EdM timeline score: 2
Dec 21, 2020 at 22:56 comment added Antoni Parellada @Scortchi-ReinstateMonica Towards the end of D. Hand's article, on an example, it becomes clear that the word 'classifier' does not refer to different medical tests, or ways to measure the cervical canal, but rather whether logistic regression or classification trees, for example, are used to score cases... A "tiny" linguistic barrier of entry for outsiders! And, yes, it reframes the point he is making!
S Dec 21, 2020 at 17:36 history bounty started Antoni Parellada
S Dec 21, 2020 at 17:36 history notice added Antoni Parellada Canonical answer required
Dec 21, 2020 at 15:40 comment added Antoni Parellada @Scortchi-ReinstateMonica Thank you for pointing that out! It makes perfect sense - he derives $L=2\pi_0\pi_1(1-AUC),$ concluding that the area under the curve is an average of losses at different thresholds. The threshold $T$ is just the variable he integrates over!
Dec 21, 2020 at 15:33 history edited Antoni Parellada CC BY-SA 4.0
deleted 242 characters in body
Dec 21, 2020 at 15:24 comment added Scortchi♦ NB It's not "the use of ROC curves to determine a threshold" that Hand calls nonsensical, but the use of the area under the curves to compare the performance of different classifiers, which doesn't involve determining a threshold. (And of course he doesn't contest comparison based on the curves' not crossing.)
Dec 21, 2020 at 14:18 history edited Antoni Parellada CC BY-SA 4.0
added 567 characters in body
Dec 20, 2020 at 0:40 comment added Scortchi♦ @AntoniParellada: I've just added a link to a paper making a similar, though simpler, point about Youden's J.
Dec 20, 2020 at 0:17 comment added Scortchi♦ See stats.stackexchange.com/a/58869/17230 & Smits (2010), "A note on Youden's J and its cost ratio", B.M.C. Med. Res. Methodol., 10, 89 for discussion of determining thresholds for classification from the R.O.C. alone.
Dec 20, 2020 at 0:06 history edited Antoni Parellada CC BY-SA 4.0
added 474 characters in body
Dec 19, 2020 at 23:46 comment added Carl Antoni, when I review papers for real, I usually write 3 to 5 pages with 10 to 20 references. This web site isn't really a good platform for that. The paper does not pass the first sniff test. It's kind of stinky.
Dec 19, 2020 at 23:42 history edited Antoni Parellada CC BY-SA 4.0
deleted 92 characters in body
Dec 19, 2020 at 23:40 comment added Carl What I was referring to I added into your question because I cannot do that in a comment. I just wanted you to see it, you can remove it, if you wish. I could do a more formal review, but that takes time. As things stand now, my informal opinion is that the paper does not hang together. That opinion is likely to get worse the more I look at in it.
Dec 19, 2020 at 23:34 history edited Carl CC BY-SA 4.0
added 85 characters in body
Dec 19, 2020 at 22:43 comment added Carl Yes to 3. above. All the ratios below the line of identity should be reciprocated, if those are actually shown correctly. However, I doubt it. I do not see how it is possible to have a correlation that is positive, if not very good (*R=*0.652 for C5 SL/VB) and have a ROC AUC less than 0.5 for C5 SL/VB. That does not seem reasonable. Moreover, I am not flabbergasted by an explained fraction of $R^2=(0.652)^2=0.425$, where is the other 0.575 explained fraction?
Dec 19, 2020 at 22:34 history edited Antoni Parellada CC BY-SA 4.0
added 397 characters in body
Dec 19, 2020 at 20:29 history edited Antoni Parellada CC BY-SA 4.0
added 231 characters in body
Dec 19, 2020 at 20:11 history edited Antoni Parellada CC BY-SA 4.0
added 456 characters in body
Dec 19, 2020 at 18:58 history edited Antoni Parellada CC BY-SA 4.0
added 400 characters in body
Dec 19, 2020 at 14:51 history edited Antoni Parellada CC BY-SA 4.0
deleted 1301 characters in body
Dec 18, 2020 at 2:56 history edited Antoni Parellada CC BY-SA 4.0
added 961 characters in body; edited tags
Dec 17, 2020 at 21:25 history edited Antoni Parellada CC BY-SA 4.0
added 151 characters in body
Dec 17, 2020 at 19:23 history edited Antoni Parellada CC BY-SA 4.0
added 13 characters in body
Dec 17, 2020 at 17:54 history edited Antoni Parellada CC BY-SA 4.0
edited tags
Dec 17, 2020 at 15:00 history tweeted twitter.com/StackStats/status/1339586169076477955
Dec 17, 2020 at 14:49 history edited Antoni Parellada CC BY-SA 4.0
added 15 characters in body
Dec 17, 2020 at 14:37 history edited Antoni Parellada CC BY-SA 4.0
added 41 characters in body
Dec 17, 2020 at 14:14 history edited Antoni Parellada CC BY-SA 4.0
added 4 characters in body
Dec 17, 2020 at 13:36 history edited Antoni Parellada CC BY-SA 4.0
added 257 characters in body
Dec 17, 2020 at 13:25 history edited Antoni Parellada CC BY-SA 4.0
deleted 338 characters in body
Dec 17, 2020 at 13:07 history edited Antoni Parellada CC BY-SA 4.0
deleted 338 characters in body
Dec 17, 2020 at 6:20 history edited Antoni Parellada CC BY-SA 4.0
edited title
Dec 17, 2020 at 5:36 history edited Antoni Parellada CC BY-SA 4.0
added 414 characters in body
Dec 17, 2020 at 5:28 history edited Antoni Parellada CC BY-SA 4.0
added 414 characters in body
Dec 17, 2020 at 5:00 comment added Dave Is the spinal canal classification whether or not the diameter exceeds 12mm? Assessing spinal canal diameter seems like it would not have a categorical response variable.
Dec 17, 2020 at 4:57 history edited Antoni Parellada CC BY-SA 4.0
deleted 4 characters in body
Dec 17, 2020 at 4:50 history edited Antoni Parellada CC BY-SA 4.0
added 171 characters in body
Dec 17, 2020 at 4:41 history asked Antoni Parellada CC BY-SA 4.0
toggle format