11
$\begingroup$

Consider the following tags:

The ordinal-logit, ordered-logit, and proportional-odds are 100% synonyms and refer to https://en.wikipedia.org/wiki/Ordered_logit. The ordinal-probit and ordered-probit are also synonyms and refer to the much less used https://en.wikipedia.org/wiki/Ordered_probit. Wikipedia puts these two terms together as instances of https://en.wikipedia.org/wiki/Ordinal_regression.

The biggest problem, however, is that [ordinal] contains lots of threads about this topic too, and even mentions it in the wiki excerpt ("Special kinds of analysis exist for such data, bearing adjective "ordinal", for example ordinal logistic regression"). My estimate is that there are close to 200 Qs there about ordered logit ([ordinal] [logistic] yields 101, [ordinal] -[logistic] logistic is:question yields another 61, [ordinal] -[logistic] is:question -logistic logit another 37) and around 50 Qs about ordered probit.

The current situation is unacceptable. I can see three possible solutions:

  1. Map everything to [ordinal]. Ordered logit regression should be [ordered]+[logistic], ordered probit should be [ordinal]+[probit], and no special tags.

  2. Use [ordered-logit] and [ordered-probit] as our main tags for ordered regressions. Move relevant threads from [ordinal] in there. Rename [ordinal] into [ordinal-data] (similar to our tags [categorical-data], [binary-data], etc.). This would require a lot of retagging, but in fact we don't need to retag all ~200+50 threads, it might be enough to retag some most prominent ones (e.g. with 5+ upvotes) and leave the rest as is.

  3. Create new tag [ordinal-regression] for both ordered logit and ordered probit, map other tags there, and proceed further as in solution #2.

I don't like #1. I don't have a strong preference between #2 and #3, perhaps a slight tendency towards #2. Note that [ordered-logit] is a relatively well-established tag, with 60+ threads.

$\endgroup$
  • 2
    $\begingroup$ Don't like #1 either - someone could be asking about, say, an ordinal predictor in an ordinary (dichotomous response) logistic regression when they use [ordered] + [logistic]. $\endgroup$ – Scortchi - Reinstate Monica♦ Apr 13 '17 at 12:56
  • 1
    $\begingroup$ Good point @Scortchi (this also means that my estimate of the number of [ordinal] threads about [ordered-logit] might be an overestimation). Do you have a preference between #2 and #3? $\endgroup$ – amoeba says Reinstate Monica Apr 13 '17 at 12:57
  • 1
    $\begingroup$ I like choice 3. In books and other places, I have seen ordinal logistic used much more often than ordered logistic. $\endgroup$ – Peter Flom - Reinstate Monica♦ Apr 13 '17 at 12:58
  • 2
    $\begingroup$ @PeterFlom When I search on Google Scholar, I get 21700 hits for "ordinal logistic regression" and 10400 for "ordered logistic regression", supporting your impression. However, "ordered logit" yields 24600. $\endgroup$ – amoeba says Reinstate Monica Apr 13 '17 at 13:01
  • 1
    $\begingroup$ Hmm. I just tried and got 135,000 for "ordinal logistic regression", 56,000 for "ordered logistic regression", 123,000 for "ordered logit" $\endgroup$ – Peter Flom - Reinstate Monica♦ Apr 13 '17 at 13:30
  • 2
    $\begingroup$ @PeterFlom Interesting. That's what I get on Google too, the numbers reported in my comment above are in Google Scholar. $\endgroup$ – amoeba says Reinstate Monica Apr 13 '17 at 13:55
  • 3
    $\begingroup$ Also proportional odds models are a special case of ordinal/ordered logit models (cf. e.g. continuation ratio models) $\endgroup$ – Scortchi - Reinstate Monica♦ Apr 13 '17 at 14:29
  • $\begingroup$ Are ordered probit and logistic really distinct enough to warrant separate tags? I like #3. $\endgroup$ – Dougal Apr 15 '17 at 11:11
  • $\begingroup$ @Dougal They are clearly conceptually very similar. But e.g. for normal (not ordered) logit/probit models we have [logistic] and [probit] tags, and nobody is suggesting to merge them together... $\endgroup$ – amoeba says Reinstate Monica Apr 15 '17 at 19:07
  • $\begingroup$ I have a prejudice against ordinal beca,use of its role in Steven's typology which I regard as having been a malign influence so I would prefer ordered over ordinal. I think re-tagging 200 is not an insuperable problem. $\endgroup$ – mdewey Apr 16 '17 at 11:28
  • 1
    $\begingroup$ @mdewey But do you prefer to have one single master tag [ordered-regression] or two distinct master tags [ordered-logit] and [ordered-probit]? $\endgroup$ – amoeba says Reinstate Monica Apr 16 '17 at 11:34
  • 1
    $\begingroup$ I pinged Franck Harrel hoping that he would comment here; he replied but cryptically and I don't quite understand what options he prefers. See here: stats.stackexchange.com/questions/111010/…. $\endgroup$ – amoeba says Reinstate Monica Apr 16 '17 at 11:35
  • $\begingroup$ I think your response to Dougal sums it up perfectly. Separate for me. $\endgroup$ – mdewey Apr 16 '17 at 11:35
  • $\begingroup$ This was up-voted (9 times as I type) so I assume there is consensus that something needs to be done but perhaps less of a consensus on what that might be. Do you want to post some alternative answers so people can vote for them separately? Happy to do some re-tagging if we agree on a programme. $\endgroup$ – mdewey May 1 '17 at 11:17
  • 1
    $\begingroup$ @amoeba: I've created those synonyms $\endgroup$ – Scortchi - Reinstate Monica♦ May 4 '17 at 7:32
6
$\begingroup$

Let me make a concrete proposal here along the lines of @amoeba's option 2.

  1. Use [ordered-logit] and [ordered-probit] as our main tags for ordered regressions. Move relevant threads from [ordinal] in there. Also retag [ordinal-logit] and [ordinal-probit].
  2. Rename [ordinal] into [ordinal-data] (similar to our tags [categorical-data], [binary-data], etc.).

Retagging (item 1) would involve ~200+50 threads so would take some time but is doable, some other recent retagging efforts have involved similar numbers.

If this goes ahead I will start on the retagging process.

Update on May 16

I have now re-tagged all the ones which straightforward searches found and were about ordered-logit or ordered-probit. I did not re-tag ones which were already closed, duplicate, and so on. If the OP had already used both [ordinal] and [ordered-logit] I decided to respect that decision. There are quite a few which turn out to be about ordered categorical variables on the right hand side in binary logistic regression but I cannot think of a good tag for them and the issue is not restricted to logistic regression so I have left them. As far as I can see [ordinal] could now be moved to [ordinal-data].

$\endgroup$
  • 1
    $\begingroup$ +1. In addition, I would make [ordinal-logit] and [ordinal-probit] synonyms of [ordered-logit] and [ordered-probit]. And [proportional-odds] a synonym of [ordered-logit]. $\endgroup$ – amoeba says Reinstate Monica May 1 '17 at 14:15
  • 1
    $\begingroup$ Also, I don't think that we necessarily need to move all [ordinal] threads about ordered logit into [ordered-logit]. After all, ordered-logit is also about ordinal-data, so if some threads end up in [ordinal-data], it's no big deal. To be honest, I would only move several dozens (?) of most important/upvoted threads and leave the rest as is. $\endgroup$ – amoeba says Reinstate Monica May 1 '17 at 14:16
  • $\begingroup$ @amoeba for me proportional odds is a proper subset of ordered logit models as there also exist continuation ratio, adjacent category, and probably some others of which I am ignorant so I would not see them as synonyms. $\endgroup$ – mdewey May 1 '17 at 15:40
  • $\begingroup$ This might very well be (honestly, I am probably more ignorant than you here), but it's not always useful to have tag B if it's a proper subset of A. Currently, [proportional-odds] has 2 threads and [ordered-logit] has 62 threads, and I guess that most ordered-logit questions are about proportional-odds. This makes the proportional-odds on its own a pretty useless tag. Hence my suggestion to make it a synonym (even though, technically, it's not a "synonym"). $\endgroup$ – amoeba says Reinstate Monica May 1 '17 at 19:38
  • $\begingroup$ @amoeba I see your point, and we do not have tags for the other subsets. $\endgroup$ – mdewey May 1 '17 at 20:30
  • $\begingroup$ Update: Scortchi has created the synonyms. $\endgroup$ – amoeba says Reinstate Monica May 4 '17 at 9:09
  • $\begingroup$ Please ping me somehow when you think that [ordinal] is sufficiently cleaned from ordered-logit / ordered-probit, and then we can ask one of the mods to rename it into [ordinal-data]. As I said, I really don't think you need to sift through all 700 Qs for that. $\endgroup$ – amoeba says Reinstate Monica May 4 '17 at 9:11
  • $\begingroup$ @amoeba will ping you, probably later this month. $\endgroup$ – mdewey May 4 '17 at 10:47
  • $\begingroup$ Re your update: I edited the tag wiki and wiki excerpt for [ordinal]. $\endgroup$ – amoeba says Reinstate Monica May 16 '17 at 6:18
  • $\begingroup$ I pinged Scortchi above and he has renamed the tag. I accept this answer now, case closed. Thanks a lot for your help! $\endgroup$ – amoeba says Reinstate Monica May 16 '17 at 8:34

You must log in to answer this question.

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