# How to deal with this interesting but convoluted question

In this question we encounter two problems:

Differing p-values car::Anova, anova & post hoc in logistic regression

• One problem is about differences between the single p-value for an anova test in relation to the multiple p-values for a Tukey's range test, for each single contrast. The single contrasts had much lower p-values (regarded insignificant) than the p-value for the single anova test.

• One other problem is about a "bug" (which is debatable just abuse of a non-supported feature) in the R-package car::Anova regarding the implementation of the method=brglmFit option (which it doesn't use for the reduced models). This problem seems to me sufficiently on-topic for this website. It is a bit much about bugs in software but in order to tackle it, it does help to have some statistics knowledge.

The question was initially written rather confusingly, but currently it is in a state that you can reproduce the problem(s).

Now the problem from the point of view of a person on stackexchange that wants to answer/solve/help this problem (these questions) is how to deal with the multiple issues in a single problem? or how to deal with complex problems?

• On the one hand: answering such questions can be simple.

(Which, for the specific example, I currently did in comments. But I was left wondering how to make those comments in an answer or what to do with the question.)

• But on the other hand: it may be important to correct such questions, cut it down into smaller portions, such that they also remain valuable in the future as part of an easy search-able database of questions.

• This is very meta meta: You are asking a long, complicated multiple-part question on Meta about a long, complicated multiple-part question on the main site, and no one has yet answered either, possibly for loosely similar reasons. There are some simple answers, notably that it is fine for someone to peel off one question or one facet and answer it. I think there is some danger of as much effort being given to asking the OP to rewrite the question as might be given just to answer at least some of it. (It's not my kind of question.) – Nick Cox May 21 '18 at 18:50
• @NickCox Does that make your comment meta very meta meta? ;) – Alexis May 21 '18 at 22:59
• I never meta regress I didn't want to go on forever. – Nick Cox May 21 '18 at 23:20
• @NickCox I've made my question a bit more brief (not sure whether this makes it less complex but I think at least the key-point of the issue/question is more clear). – Sextus Empiricus May 22 '18 at 7:30
• I don't see that the original question on CV raises new issues for Meta. Consensus: it's on-topic. You can't push or pull or even easily persuade people into providing answers they don't want to give. The community has done a good job in commenting. If you (@MartijnWeterings) want to provide an answer, that's good, and it doesn't have to be a complete answer. Like scientific or scholarly papers, all threads here are candidates for greatness, but most don't achieve it, and sometimes a thread doesn't deserve closure but it won't go far. The moral is to ask focused, specific questions. – Nick Cox May 22 '18 at 8:31
• I intended this meta question to be more like about 'how to provide better answers' or 'how to answer, or what to do, when the question is not focused'. I do not want to discuss whether the question is on-topic or not (I believe it is on-topic), but more like whether and how we can improve quality of question+answer. – Sextus Empiricus May 22 '18 at 9:41
• If the answer is that it is best to just ignore this issue (or that it is not necessary to spend much time on improving such questions) and that it is best to just simply provide the solution to the question without trying to improve it on a meta-level (splitting up the questions, make new questions that handle certain topics separately, make the question optimized for search engines, etc.) then this is a fine enough answer to this meta-question (at least for me) when this is the consensus here. – Sextus Empiricus May 22 '18 at 9:49
• I think most of the issues have been covered on Meta. Nothing stops anyone asking a new question inspired by an existing one and even answering it too. What does seem out of bounds is drastically rewriting somebody's question. Edits should be conservative (in terms of not changing the statistical question) and cosmetic (tweaks and fixes to presentation). – Nick Cox May 22 '18 at 11:20
• There is always some caprice over what is answered. I will answer a question specific to someone's data that has little general interest if I can think of something to say. But long term, the value of CV hinges on our high-quality threads with definitive answers, not on it being a helpline in public. – Nick Cox May 22 '18 at 11:22
• "the value of CV hinges on our high-quality threads with definitive answers, not on it being a helpline in public" that is exactly why I posed this question. I had one occasion where I tried to (zealously) change a question (by a too large extend, I can't find the question anymore but the critics were correct) which was frowned upon (which might be a reason that I started to become more careful in these issues). Now (after this post) I believe that (at least for me personally) one way to resolve this issue is to simply post ones own questions that are "better" versions (let the votes decide) – Sextus Empiricus May 22 '18 at 11:31
• To me the point is to get these 'high-quality threads with definitive answers' and sometimes I find the question less suitable to start these threads because the question is very convoluted. So, I guesss, it is best practice to start new threads that deal with the questions/issues individually. – Sextus Empiricus May 22 '18 at 11:37
• I don't take @NickCox's comments here as criticizing your asking of this meta Q, Martijn. This is a perfectly reasonable question to ask on meta.CV (+1 from days ago). – gung - Reinstate Monica May 23 '18 at 14:26

You can ask them to set up a simpler example. For instance this person is clearly overwhelmed by both separation and multiple testing. You can jitter/permute the $Y$ or use example data from another problem to illustrate the problem.