16
$\begingroup$

(3858) is one of our most popular tags; its wiki excerpt says

Refers generally to statistical procedures that utilize the logistic function, most commonly various forms of logistic regression

It has a merged tag synonym . I looked at the questions in this tag (both newest questions and most upvoted questions), and it seems that overwhelming majority is about logistic regression as opposed to logistic function as such. I would therefore suggest to make [logistic-regression] the master tag. I don't like the current tag name because "logistic" is an adjective, and it's not completely clear from the tag name if it refers to the logistic regression or possibly to something else. I think it makes more sense if tag names are nouns.

More importantly, we have a separate (432) tag. I feel that most of the questions in this tag are also about logistic regression. Wikipedia starts the https://en.wikipedia.org/wiki/Logistic_regression article with "In statistics, logistic regression, or logit regression, or logit model..." We have a [probit] tag too, so anybody asking about probit vs logit is likely to use [logit] tag, but their question is really about logistic regression. Again, some small fraction of the questions might be about logit function as such, but most of the threads are about logistic regression (I checked), and separation of this topic into two tags (logistic and logit) is very unhelpful.

My suggestion is to map both [logistic] and [logit] to [logistic-regression].

| |
$\endgroup$
  • $\begingroup$ In the context of log odds and Bayesian updates it is possible that the "logit" tag could be used. Glancing at the linked query doesn't seem to indicate it currently is ... though I had reason to refer to it this way in a comment recently :) $\endgroup$ – GeoMatt22 Apr 17 '17 at 23:15
  • 4
    $\begingroup$ I'm not totally happy with the current setup but the proposal doesn't fix what I see as the problem, it makes it worse - questions about logistic fits (e.g. via nonlinear least squares) are lumped in with questions they're not closely related to; under this proposal they'd be (incorrectly, I think) tagged with logistic-regression. I realize their numbers are fewer (I've seen a good number over the years I have been on site, though), but I don't think that's not a good reason to make them impossible to find. If we do anything like this proposal I think we'd need a new tag for them $\endgroup$ – Glen_b Apr 17 '17 at 23:50
  • $\begingroup$ @Glen_b I see. That's an argument against making [logistic-regression] the master name. But what about the second part of this proposal, i.e. mapping [logit] and [logistic] together, e.g. under [logistic]? $\endgroup$ – amoeba Apr 18 '17 at 6:00
  • 4
    $\begingroup$ I personally don't have an issue with that; I imagine we might need to edit the logistic tag wiki to specifically mention the logit as well. I'm not sure there's a particularly satisfactory way to bring in logistic curve fits that aren't logistic regression together with logistic regression $\endgroup$ – Glen_b Apr 18 '17 at 8:08
  • 1
    $\begingroup$ @Glen_b One option that comes to mind regarding logistic curve fits is to create logistic-function tag, move a couple of dozens of most prominent threads there (there is no way one go through all 4k [logistic] threads, but finding most prominent ones might be possible), and then rename logistic to logistic-regression. Or do you think it's futile? $\endgroup$ – amoeba Apr 18 '17 at 8:17
  • 1
    $\begingroup$ I don't think it's futile. That might work okay. Perhaps that's a suggestion worth debating/voting on. $\endgroup$ – Glen_b Apr 18 '17 at 8:21
  • 1
    $\begingroup$ @Glen_b What I currently have no sense for, is how many questions are there in this topic. I guess under current tag system, they should be tagged with [logistic] and [curve-fitting], but there are only 6 questions like that. This IMHO would not be worth a separate tag (given that we do have [curve-fitting]); but maybe this search misses a lot of threads. $\endgroup$ – amoeba Apr 18 '17 at 8:33
  • 1
    $\begingroup$ Yes, if we still had logistic it would be covered by those two tags; it might be worth trying to identify a few more posts that could have those tags but presently don't $\endgroup$ – Glen_b Apr 18 '17 at 8:43
  • $\begingroup$ Just a note on the [logistic-function] idea: there is already a (small) sigmoid tag. Also there may be neural-net applications that use a logistic activation "inside", i.e. not just in the output layer? (and tough to call that logistic regression) $\endgroup$ – GeoMatt22 Apr 21 '17 at 21:42
  • 1
    $\begingroup$ @GeoMatt22 Thanks for mentioning the sigmoid tag. It does not look very useful to me (and does not even have a wiki excerpt); we don't even have a tag for relu neuron, or for any other activation function. I would be rather inclined to eliminate the [sigmoid], just using our generic [neural-networks]. $\endgroup$ – amoeba Apr 22 '17 at 21:48
  • $\begingroup$ I am fine with [sigmoid] being eliminated. Also was noting that the NN usage could fit under [logistic-function]/[logistic] (e.g. a tag wiki) in addition to the logit/log-odds usage. (as I think not all NN uses are cleanly [logistic-regression]) $\endgroup$ – GeoMatt22 Apr 23 '17 at 0:02
  • $\begingroup$ @GeoMatt22 "Logistic regression" is GLM with logit link and binomial family, i.e. a linear model. Nonlinear NN can't be called "logistic regression". Regarding [logistic-function] tag, I am still not fully convinced that it's needed at all (if we rename [logistic] and [logit] to [logistic-regression]). But most threads tagged with [logistic] and [neural-networks] (stats.stackexchange.com/questions/tagged/…), that's 35 of them, are not about logistic-regression, you are right. It's a small drop in [logistic] usage though. $\endgroup$ – amoeba Apr 23 '17 at 6:23
  • $\begingroup$ @amoeba your ideas are OK by me. (I was only loosely referring to many NN's that seem ~ input --> nonlinear feature mapping --> logistic-regression = output-layer). In your plan, should tags be removed from the small minority of questions that are not about logistic regression? $\endgroup$ – GeoMatt22 Apr 23 '17 at 6:51
  • $\begingroup$ @amoeba Whatever we end up doing we need to end up with a suitable place for questions about the logistic distribution $\endgroup$ – Glen_b Apr 26 '17 at 0:04
  • $\begingroup$ @Glen_b I find around 30 questions mentioning "logistic distribution", but looking at them probably around half wouldn't deserve the corresponding tag (and some of them currently don't even have [logistic] tag). So if we were to create [logistic-distribution], which would be one obvious solution, I estimate it would get maybe around 20 questions only. What do you think? $\endgroup$ – amoeba Apr 26 '17 at 7:35
7
$\begingroup$

There seems to be a wide support (currently 10 upvotes) in favour of my suggestion to map both [logistic] and [logit] to . This will be a big tag with ~4200 Qs.

However, @Glen_b raised the concern that there are some topics among [logistic] threads that are not about logistic regression, and we should perhaps make separate tags for them. After some discussion, I am thinking of:

@GeoMatt22 mentioned tag with 17 Qs and no wiki excerpt; I went through these threads, re-tagged them as appropriate, and eliminated this tag. Most questions were about neural networks; I don't think we need a separate tag for [logistic-activation-function] (e.g. we don't have a separate tag for ReLUs), and generic [neural-network] should do fine. I have also looked through all [logistic] [neural-networks] -"logistic regression" is:question threads (a couple of dozens), and now this search returns zero results.

Status update (May 17): I would still like to scan for possible [logit] threads that are not about logistic regression. E.g. https://stats.stackexchange.com/search?tab=votes&q=%5blogit%5d%20is%3aquestion%20-%22logistic%20regression%22%20-%22logit%20model%22%20-probit%20-%22logit-regression%22%20-glm%20-glmer%20-mlogit. Will report later.

| |
$\endgroup$
  • $\begingroup$ These stats.stackexchange.com/questions/tagged/… should be dealt with together with [sigmoid]. Some are about logistic regression but some others are about sigmoid neurons; there we should remove the tag. $\endgroup$ – amoeba Apr 28 '17 at 8:10
5
$\begingroup$

Not an answer, but rather a suggestion (plus, bumping the thread).

In my opinion, we should mark as synonyms:

(456) $\rightarrow$ (4150)

While they are not the exact synonyms, a quick glance an the Q&A's reveals that they are used as if they were synonyms.

While the points made in comments and by @amoeba are convincing, this merge should not be controversial.

EDIT: I suggested it as synonym, please cast your votes if you agree.

| |
$\endgroup$
  • $\begingroup$ +1, thanks for doing this. $\endgroup$ – gung - Reinstate Monica♦ Aug 14 '17 at 19:14
  • $\begingroup$ I agree and upvoted the suggestion. I got somewhat stuck sorting everything into logistic-regression, logistic-curve, and logistic-distribution (as per my answer above) because there is some amount of questions about logistic neurons (that I aimed to tag neural-network, but perhaps we need logistic-neuron tag? as a synonym to neural-network or stand-alone?) and also about logit transformation as such and I did not make up my mind about what to do with it. $\endgroup$ – amoeba Aug 16 '17 at 12:12
  • $\begingroup$ @amoeba I'm not the neural-network kind of person, so I cannot comment on that. $\endgroup$ – Tim♦ Aug 16 '17 at 12:16

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .