Should low-quality "link-only" answers be flagged for conversion into comments?

I'm sure there must be some general consensus on this but I don't know what it is...

(I don't like the link-rot that comes with them, and I think that very little would be lost in most circumstances by conversion into comments, even if the link actually does answer the question. Obviously I'm not referring to answers which attempt to summarise the link.)

I think they generally should; the 5th bullet point here implies they should be deleted, and they're generally put into the review queue (I assume by a bot) anyway. It makes sense to keep the useful information in them by making it a comment. I usually haven't actually flagged (figuring a moderator would likely see them anyway when they were voted for deletion) but it's probably better to do so.

@gung thanks for your gentle but helpful prodding; I needed to think more carefully.

@Silverfish hi

@glen_b Thanks for that.

No worries.

If you recommend deletion in the review queue, does that mean moderators look over it, or is the deletion automated?

I'd feel more comfortable pressing the button if I knew that they were all reviewed by a mod later along the line

I think some things at least can be automatically deleted, I don't know that it isn't the case here. But the chances of at least one mod not seeing it anyway would be slim.

Could always ask on meta.

I'm sure it should be documented somewhere but I haven't happened across it yet!

I'm certain it's documented somewhere -- but it's not always easy to find. It's a pity all that stuff isn't also collected into a "how SE works" document (even if that was just a collection of links)

I haven't found anything on Meta so far so I think I'll ask there

Well, a collection of headings and links, anyway...

That seems to cover what I wanted, but took a bit of trawling to find it

Though it's from 2011 which is pretty old in CV terms, I assume it's still relevant

(only tracked it down by noticing the "hyperlinks" tag!)

@Silverfish, @Glen_b, there's a little bit of information about deletions in stats.stackexchange.com/help/privileges/trusted-user, I'm not sure where the rest is.

The gist of it, I believe, is that something will be automatically deleted w/o a moderator's action if it accrues 3 delete votes or 5 recommend deletion votes w/i a certain amount of time. (I'm not sure how long.)

@gung thanks. I seem to recall it's a good deal longer than typical times things stay in queues here.

@Glen_b, I'm still not sure if I'm on the same page. I might be off somewhere in left field here.

The OP asked for a test of means. I don't know how strictly that was meant, of course.

The MW is equivalent to a test of means only if the shapes are identical, I believe.

I thought that was what you were getting at.

@gung Yes, I know the OP wants a test of means. I establish in comments the connection between means and stochastic dominance.

I thought it was a rather subtle point, actually.

But I wonder if it will be missed by readers if it isn't made a little more explicit.

But maybe I'm misunderstanding something about MW or your answer or something.

The point I was making with all that rabbiting on was to establish that the MW is a test of means in a broader range of cases than location-shift with identical shapes.

Oh, it may be, but it isn't guaranteed unless the shapes are identical, or am I wrong here?

That was the point of my R example.

bear with me while I try to make it a little less diffuse.

The means were identical, but it's highly significant.

So the OP shouldn't use the MW to test for a difference of means if they aren't willing to assume identical shapes.

No, there's a situation in between; you overgeneralize from your example.

The MW is sensitive to P(X>Y)>1/2

But stochastic dominance is more restrictive

location shift with identical shapes is more restrictive again

Do you mean you can assume less than identical shapes & still have MW be strictly a test of means?

identical under the null, not under the alternative

Consider the null "F_X=F_Y" and the alternative "F_X<F_Y"

i.e. a stochastic dominance type alternative

Under the alternative (and some additional conditions), expectations differ.

Consider rephrasing it as "S_X=S_Y" vs "S_X>S_Y" where S=1-F (assume continuous distributions!)

Now integrate both sides. ... you get E(X)>E(Y)

A lot of books say you have to assume equal shape to get means (and I've said it myself before), but it looks like I'm wrong.

I still wonder if I'm not on the right page. Here's my thinking: the OP has my data & wants to test if the means are the same, but isn't willing to assume that the distributions are normal. The OP wants to try to use the MW to test if the means are identical. So they go ahead, & get a highly significant result. They would then conclude the means are certainly different. But they are identical.

To prevent this, we need to tell them only to use the MW to test for equal means under some conditions, right? We need to specify the assumptions they must be willing to make for the test to be a valid comparison for the means.

Because the assumption of stochastic dominance was wrong. If your assumptions are false your conclusions made on that basis don't hold.

True, so we need to tell them to assume stochastic dominance?

If the restriction under the alternative that S_X>S_Y is true, it implies the means are different.

The difference in means is the difference in integrals of S's

Which (here's where some of those conditions come in) is the integral of the difference.

I'm not sure if I'm following. We should tell the OP to use the MW as a test of means if "the restriction under the alternative that S_X>S_Y is true"?

If they're prepared to make that assumption, it's effectively a test of means, yes. What part am I explaining badly?

I'm sure you're explaining it fine ;-)

I'm not sure that presently my feeble brain is up to introducing all the restrictions correctly.

Why did this take me so long to see before?

I need to write a new section of that answer explaining it more fully so that people like cardinal and whuber can pick at the conditions.

I'm not sure what you're seeing. It must be late, your time, or early? Frankly, I'm never sure if it's today or tomorrow my time...

It

It's nearly 10am. I need to go work now.

Hey there, is anyone on?

Have a good day, @Glen_b.

What's going on, @Howcan?

@gung Hey there. Long story short is that I will need some help with multivariable regression. Do you think you can help?

Why not ask on the main site? What is your question?

@gung Check the room I made with you, if you don't mind.