12
$\begingroup$

Recently a question came up that I felt had several redundant tags for variants of multilevel or mixed-effects models. As far as I can tell, there are six tags I can find that are related to this topic some of which have tag wikis and some do not (for whatever that's worth):

There are also (at least) two related software tags:

  • (informative tag wiki)
  • (no tag wiki)

This point has been discussed in a previous meta post by Andy W, but it doesn't seem to have resolved everything; or, perhaps, new developments have arisen in the meantime. I realize there are methodological and semantic differences between some of these terms (as chl and Andy W discussed in the comments of that post), but it still seems like an awful lot of tags...!

Suggestions

  1. Questions tagged and should be merged.
  2. Questions tagged and should be together as well.
  3. Questions tagged could be merged into the appropriate one of those categories and then also given a tag? There are probably questions that are not marked as that should be though...
  4. The tag deserves a tag wiki. Unless someone else would contribute something more detailed, I would propose something like

Parameters associated with the particular levels of a covariate are sometimes called the “effects” of the levels. If the levels that are observed represent a random sample from the set of all possible levels we call these effects "random."

which is a paraphrase from Douglas Bates' book. (I think unpublished or in press at this point.)

$\endgroup$
8
  • 1
    $\begingroup$ @Andy, the color scheme of meta hides it a bit (at least to my eyes), but note that smillig does link to your previous thread in this question. It might be worth making a little more prominent. :-) $\endgroup$
    – cardinal
    Aug 28, 2012 at 14:04
  • $\begingroup$ yes I see now, thanks @cardinal, I should have read more carefully. $\endgroup$
    – Andy W
    Aug 28, 2012 at 14:07
  • $\begingroup$ @Andy, I missed it the first time through, myself. I (+1)'d your comment so the link would be more easily seen. $\endgroup$
    – cardinal
    Aug 28, 2012 at 14:08
  • 2
    $\begingroup$ @cardinal: done! Also, I should point out that the first sentence in Suggestion 3 should be read in a kind of valley girl voice sort of way with rising intonation. It's more of a question than a suggestion...! $\endgroup$
    – smillig
    Aug 28, 2012 at 14:22
  • 1
    $\begingroup$ smillig, I will do my best to read it this way, though I am probably not very competent at achieving a faithful reproduction. (+1). $\endgroup$
    – cardinal
    Aug 28, 2012 at 14:37
  • $\begingroup$ We may want to throw fixed-effects-model into the mix for consideration as well. $\endgroup$
    – gung - Reinstate Monica
    Aug 28, 2012 at 20:11
  • 1
    $\begingroup$ Here's an incorrectly placed multilevel-analysis tag: stats.stackexchange.com/questions/22234/…. There should be hierarchical-Bayesian for questions like this, instead. $\endgroup$
    – StasK
    Sep 2, 2012 at 18:13
  • 1
    $\begingroup$ So do we have a closure on this? Otherwise, I will just go pruning and vandalizing some tags, if I understood whuber's proposition correclty :) $\endgroup$
    – StasK
    Sep 19, 2012 at 4:13

3 Answers 3

Reset to default
11
$\begingroup$

While multilevel and multilevel-analysis are probably easily mergeable, some of the others are genuine terminology differences.

  1. "Random effects" to economists is the GLS estimator of the random intercept model in the panel-data models. The "fixed effect" is a conditional estimator that only uses the within-cluster variability. I am sure that whatever is done with the random-effects-model tag will create confusion. Sometimes, economists do say that (what they understand as) the random effect model is valid when you have the random sample of clusters, but generally what they worry about is the correlation between the regressors and the random effects, which rather has to do with how complete your model is. So some of the random-effects-model entries may in fact have to be moved to panel-data if we are to clean this. May be there should be random-effects-model-econometrics and fixed-effects-model-econometrics subtags.
  2. GLMM stands for generalized linear mixed models, and while related to the GLM (generalized linear model, not the terrible "general linear model acronym" coined by SAS, I believe), it is different enough that these tags should live independently of one another.
  3. R packages have somewhat different functionality. Since CV is very heavily R-oriented, it is probably fine that these tags exist as they are, but if you are subscribed to the multilevel list, you will see way many more questions about software packages like HLM, Mplus and Stata (in particular, the GLLAMM package that beats most other ones except for speed). Social scientists tend to go to these lists rather than to some geeky website ;). So to be generalists, should we go ahead and create the tags for all of these other packages???
  4. There is also hierarchical-analysis tag. The phrase is again used in a number of different contexts. Social scientists who use the HLM package would mean the multilevel analysis/mixed models that we are discussing here, while Bayesian statisticians would mean a model with priors and hyperpriors, which may or may not be involve the random effects of the multilevel models.

There was an earlier Pinheiro & Bates book, so may be there is a published version of that quote, if needed.

A good course of action might be to elect a mixed-models officer who'd be cleaning these tags from time to time as they see fit. Otherwise, we'd be just lost in discussing what needs to be done.

$\endgroup$
6
  • $\begingroup$ +1 Thank you for showing us how to navigate the terminology. Re the last paragraph: Are you volunteering? :-) $\endgroup$
    – whuber Mod
    Aug 30, 2012 at 17:34
  • $\begingroup$ @whuber, I am not sure I have enough rep within these tags to do much: I tried editing something, but was not allowed to. (I did not know there were tag-specific reps, at all!) $\endgroup$
    – StasK
    Aug 30, 2012 at 20:00
  • $\begingroup$ I believe you can propose edits (to the wikis) and synonyms, which could then be approved by people with such privileges. Or just provide a few more great answers and bring your rep up to 10K and help us out :-). $\endgroup$
    – whuber Mod
    Aug 30, 2012 at 20:32
  • $\begingroup$ At the current rate, it will probably take me 6-7 months to get that extra 3300 ;). $\endgroup$
    – StasK
    Aug 30, 2012 at 23:27
  • $\begingroup$ A partial follow-up 4 years later: meta.stats.stackexchange.com/questions/4391. I think your input would be very important there. $\endgroup$
    – amoeba
    Oct 11, 2016 at 13:30
  • $\begingroup$ I posted an answer in my follow-up question with very specific suggestion: stats.meta.stackexchange.com/a/4833. Let me know if you have thoughts. Thanks! $\endgroup$
    – amoeba
    May 24, 2017 at 13:45
6
$\begingroup$

I have long believed that organizing tags optimally will require some level of constant maintenance, and is one of the most important aspects of enhancing the usability of the site, and maintaining it more generally. Personally, I tend to be in favor of merging tags more aggressively than seems to be the case, although I certainly agree it could go too far. Basically, I think tags should be merged unless it is easy to think of questions (whether or not anyone has actually asked them yet) that clearly fall under one of the tags, but not the other. Moreover, I don't mean terminological issues in the sense that something is called by one name in one discipline & another name in a different discipline. In such a case, the two tags should exist, but the lower-frequency variant should become a synonym for the higher-frequency variant. That way, when someone searches on the name they know, something will show up, but it will be remapped to the standard version. In the referenced discussion, for example, both @AndyW and @chl seem to agree that there isn't a substantive difference between the tags they are discussing, but that they do often connote different things to some people. In a case like that, I say merge. As a result of these considerations, I would argue that all 4 tags in suggestions 1 & 2 should be merged into mixed-model, which is the most commonly used, and has a wiki excerpt that explains it is used for a bunch of different names for the same thing.

I do think glmm and random-effects-model could stay, because I can imagine questions that fit better under them than mixed / multilevel etc. Some examples could be questions about glmms vs. gee, or the referenced question about the interpretation of the meaning of the random intercept. (For the record, I think random-effect would be better for this than random-effects-model.) I acknowledge that these questions could go without these tags, and just have mixed-model, but I think something would be lost in a way that wouldn't occur if someone asked about students within classes within schools and clicked on the multilevel-analysis tag, but ended up with mixed-model instead.

I don't have a strong sense of what should be done with the R tags (lme, & lmer).

I also strongly believe that all tags should have wiki excerpts at a minimum. (I'm fairly indifferent to whether they have full tag wiki entries--I doubt anyone even reads them.) I think that excerpts would best include a little about the topic, and a little guidance on the appropriate usage of the tag. One thing to note is that the wiki excerpt give you a certain limit of characters, but not all of them will show up when possible tag matches are displayed on a question.

I approve of your suggested excerpt for random-effects-model.

$\endgroup$
4
  • $\begingroup$ +1 Thanks for your answer which is as detailed as always. I may be wrong about this, but I think that "multilevel" often implies hierarchical nesting while "mixed" can mean crossed or non-nested random effects which is why my suggestion was to maintain the distinction. But it's a slippery slope! $\endgroup$
    – smillig
    Aug 28, 2012 at 21:32
  • 2
    $\begingroup$ I like your comment "I think that excerpts would best include ... a little guidance on the appropriate usage of the tag"- I've never thought about this when editing but it would be good to mention, if needed, what distinguishes this from near-duplicate tags (e.g. random-effect vs. random-effects-model). Overall, I think there are way too many near-duplicate tags. For example, I'm having trouble seeing what would be lost if the many tags mentioned in your 2nd paragraph were merged (with synonyms). If anything, this could help novices understand the many ways of referring to such models. $\endgroup$
    – Macro
    Aug 28, 2012 at 22:11
  • $\begingroup$ @smillig, I'm certainly open to the possibility that I'm insufficiently nuanced here, & willing to go along w/ what the larger community thinks best, of course. Using your definitions (which I think are reasonable), multilevel is a special case of mixed where there are no crossed effects, eg. An analogy could be to ANOVA vs. regression, a case where I definitely think the 2 tags should be preserved. The question is whether ML & ME deserve the same deference as those. $\endgroup$
    – gung - Reinstate Monica
    Aug 28, 2012 at 22:18
  • $\begingroup$ @Macro, here's another way of thinking about my point (I may even include this in my answer). Imagine we edited the tag wiki excerpts to have some basic info, & then (eg) "use this tag but not mixed-models when your Q centers on something specific to GLiMMs that is distinct from mixed modeling more generally, otherwise only use MM" (or some such). Then, eg, a Q like: "how are parameters estimated differently in GLiMM from GEE?", would fall under the GLMM tag, but not the mixed-models tag. OTOH, I can't think of an analogous condition / question that differentiates ML from ME. $\endgroup$
    – gung - Reinstate Monica
    Aug 28, 2012 at 22:29
-4
$\begingroup$

I think most of these tags should stay, because people search on all (or nearly all) of them. I don't see how a WIKI about them will help, but maybe I am missing something.

$\endgroup$
1
  • 2
    $\begingroup$ When you make synonyms, if you search for one tag the other "main" tag pops up in its place. Gung makes the point that the wiki excerpt, which comes up when you first put tags into the question, is an immediate guide about how to utilize the tags. We should encourage fidelity with how tags are assigned, as they are the predominate mechanism to find similar questions on the site. It does no one any good to have a bunch of redundant but disconnected tags. $\endgroup$
    – Andy W
    Sep 2, 2012 at 12:26

You must log in to answer this question.

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