@BabakP If you did want a higher accept rate, I'd suggest avoiding self-study questions (especially the ones not even tagged as self-study) .

@whuber Re Epic badges: I was making a joke about the unattainability of those badges. You're leading that, by the way. On the mean rep per answer, I'm not certain that's the best indicator of likely quality of moderation. Indeed, if someone is only answering questions which are likely to get high upvotes, I'd consider whether they might be shirking in their duty to the people that ask questions which won't get many upvotes.

(Not that I am making any such suggestion about any candidates, all of whom I have no such qualms about.) It is an indicator of quality, but it could also be an indicator that someone has too much of an eye on their reputation.

Or rather reputation-per-question.

@whuber On the 'test' questions -- I've been tested at least once to my recollection.

@Glen_b, have you been 'tested' about reviews on CV? If so, it's odd that I haven't...

@gung That's my belief, yes.

At the end of the review the behavior is somewhat different from reviewing an ordinary spam-edit

And I bid you all a fond farewell and ado :)

@BabakP That'd be "adieu", which is basically "goodbye" in French.

@gung Rick-Blaine-with-a-mug

@Glen_b, it might be bourbon in that mug

I'm not sure what the purpose of starring chat comments is but those made me laugh

Uh, was that farewell from @BabakP him actually leaving CV after all?

His account seems to have gone. I thought he had changed his mind.

Hmm. I have no basis on which to assume a gender though. Him or her.

@Glen_b He/she asked the SE team to delete his account.

A second time? Wow

Or was that still the first one? It would seem odd, because he/she was participating again, even to the extent of answering the questions in the moderator election.

Either way I am mystified.

@Glen_b Yes, but the first one was aborted (in extremis).

It makes no sense to continue to actively participate in the moderator election if you're leaving the request for deletion going, but on the other hand, if it was a new request, I've seen no hint of a reason for it.

@chl ah, well thanks.

I wonder why the new one.

In any case, I'm saddened by the departure, because in a very short time the quality of the answers being generated had become quite high.

It is always difficult to know why people are leaving. However, that is apparently a personal decision: nobody forced him to leave.

Sheesh that kid sure is sensitive isn't he

@Macro Probably, and it would have been preferable he talked more openly. Anyway, it's done now (or for the moment).

This is probably because his moderator questionnaire answers didn't get many upvotes

@Macro That's what this comment of him would suggest.

well what does he expect? He flamed out like a week ago. Did he really think people would vote him in as moderator?

Not to mention, he asks elementary questions about the site on nearly a daily basis. The moderators are supposed to answer questions like that, not ask them

@Macro That's something users can learn quickly :) He probably would have done so as well. I'm more concerned with such unsteady behavior. But that's my personal view.

It does take a while to learn the site. I am still coming to grips with some parts of how things work.

True

In any case, I think this is for the best

He seemed VERY pre-occupied with reputation and badges

more so than anyone I've ever seen in the Stats SE

doesn't seem like that would pan out well after he eventually realizes rep doesn't really matter

Ok, I don't like talking about specific user in public room. I'm sad when I see users leaving the site so abruptly (remember Procrastinator?), but that's community life after all.

@Macro I remember worst cases :(

I have to go back to work. Cheers,

ok @chl, fair enough about not wanting to talk trash, but I think his odd behavior - in the context of him being a moderator candidate - does warrant some public discussion.

I should probably try to do some work today also. Cheers all

For me it's already tomorrow.

And for me its "My god, what day is it?!" Pesky new job and altered sleep schedule.

Is cross validated the right place to ask questions about 'R'? Or would that be better at stackoverflow

Eterm - it very much depends on the question. Does it involve statistical knowledge, or is it primarily an issue of programming?

I guess more programming; I'd like to pick n samples of which of 4 distributions is highest sampling each every time; If there's a analytic/statistical solution that would short-cut it then great but I suspect it's more a case of programming it!

So there isn't a R-specific SE yet? That's a shame!

That sounds like a question for SO then. Generally with R questions, the "Is there statistical knowledge involved" is a decent dividing line, and you should never take it personally if your question gets migrated :). As for an R-specific SE site, there's...plusses and minuses to that approach, but I don't know if there will ever be one.

Ok thanks, my query does sound better suited to SO

@Glen_b (re "someone has too much of an eye on ... reputation-per-question"). That's a great point and I am suitably chastened by it. I would like to add, though, that I think it's a good idea for everyone to keep an eye on their acceptance rate. It should be the rare question that we answer without intending to write something that deserves to be accepted!

I totally agree that such situations should be rare; if one wished to add information rather than write an answer worth being accepted, that's what comments are for.

Though no chastening was intended.

@whuber, I still think there are some caveats to that ("deserves to be accepted"), although, I acknowledge that in general it's true. Consider that there are cases where OPs don't accept b/c they don't seem to know to, or they drop a question & never return. I've answered several questions where I knew it wouldn't be accepted (for instance where there is already an accepted answer), but I thought I could add something positive to the discussion (eg, my logit vs probit answer is a case in point.)

As far as the moderator election is concerned, it is clear that the desertion of one of the candidates was caused by self-inflicted problems. Although we should always regret the loss of active and able community members, I believe there is nothing we could have done to change this outcome (apart from the many accommodations we had already made) and there is nothing we should change about this site as a result.

@Eterm I think the opposite (now, I didn't in the past) as it would create more confusion for users, now there are 2 SE sites where R-related Qs may be appropriate, but with a specific SE site there would be 3. Plus you'd dilute the pool of people who'd look at the question. Given R's focus, I think the current situation is about right.

@gung Yes, there are clear exceptions as you and @Glen_b duly note. But "always write your answers in a way that makes them deserving of acceptance" is an excellent guiding ideal. It encapsulates most of what SE strives for.

I suspect we really don't have disagreement here.

@whuber But you are at the whim of the OP, plus what if there is more than one (good) answer, or the user accepts before you supply your answer? I fully agree with the sentiment (write as though you deserve the accept tick) though.

@Glen_b I'm sure we do not disagree. I'm actually expressing my own experience: looking back through the chronology of my answers, I am reminded that it took me 14 tries to get even one acceptance and a good 100 tries to learn how to write answers that had some reasonable chance of being accepted. Fortunately I had @chl (and a few others) to show the way.

@whuber I don't think you were implying it, but just in case there's any miscommunication - my sadness at the departure wasn't to suggest that too little was done. I don't think that was the case.

No doubt there's stuff I wasn't aware of, as well.

@GavinSimpson Yes, one is always uncomfortable being judged. It's a little like being a politician, whose job depends on the whims of the voters. But some individuals have shown that it is possible consistently to have a high acceptance rate, and my own experience shows that it is possible to increase one's rate substantially: thus, part of that is under one's control.

@Glen_b Most of this debacle was carried out in public. I wrote a private moderator message a couple weeks ago to stave off the initial threat of departure (and keep the SE team from pulling the trigger prematurely). Although that was not seen by anyone but the mods and the SE team, it said the same things the rest of you subsequently said in this forum.

@whuber I know about the difficulties when starting out - though my first acceptance came on only my second question, it took me about another 20 answers to get a second one.

Would my question be better on sciComp? stats.stackexchange.com/q/71154/2426

@GavinSimpson The politician analogy might not be a bad one. For instance, it occurs to me that there are some simple "tricks" to increasing the chance of acceptance. One of them I don't like and try not to stoop to: answer quickly. The ones I do like are (a) include pretty illustrations, (b) provide plenty of useful links and authoritative references, and (c) praise the question itself.

Although it might seem crass to include these elements for their own sake, the fact is that they generally improve the answer anyway.

I put it here because I'm mostly concerned with finding standard errors on my fit parameters, so I figured it's nore statistical in nature

but in a way it's also about numerical accuracy, so...

@ColinK It's your call. I see a statistical element lurking there, but you would need to provide more information about the model. It seems to be a question of how well the errors are described by a Gaussian. If they are not, there will be an error in your second-order calculations. This is something well understood in the context of Maximum Likelihood estimation.

hmm, I'm not really sure

@ColinK I do see it as a question that could be relevant here.

In considering how gaussian my errors are, is the relevant factor how gaussian the residuals are? or how gaussian I'd expect them to be a priori?

@ColinK, it's sounds like a potentially good question. This seems to be closely related to whether or not it's best to approximate the fisher information by the averaged gradient outer product or the (negative) average numerically differentiated gradient. Clearly they are both reasonable but I've only done the former.

@whuber I'm more than happy to be judged by this criterion, or any other one, as long as the "features" of the given metric are acknowledged. (I do quite well out of the suggested metric among the other Moderator candidates, but, with one exception, I wouldn't place myself ahead of any of the other candidates.) In the long-run, the idiosyncrasies of OP accepting, etc should even out though I guess.

@ColinK Actually, I'd like to see a really good answer to that one. I may learn something.

because what I'm doing here is fitting a pulse shape to measured data. The physics that creats the pulse shape is well understood, so I'm very confident my model is a complete model of the data, and that the noise is additive gaussian noise from the electronics. However, fitting pulse-like data is notorious for small fit errors near the peak, so the actual residuals are definitely not gaussian

I agree with Glen

@Colin I am not sure I follow your final sentence above.

Can you explain what you mean by 'small fit errors' and in what way they yields non-Gaussian residuals for known gaussian errors?

Ditto [to @macro]. We do have some peak-fitting threads appearing here from time to time; a few might be worth reading. A recent one that comes to mind (because I was one of the respondents :-) is at stats.stackexchange.com/questions/70870. However, that one does not explicitly address the question of errors of estimates.

it sounds like you're describing some form of bias.

Or lack of fit, perhaps.

in fitting sharp peaks in the presence of noise, it's typical to have a small ripple in the residuals near the center of the peak. If you look at a histogram of the residuals there is a clear non-gaussian shape due to this ripple

Thanks. Clear now.

Still sounds like a lack of fit.

but a priori, I'd expect exactly gaussian errors

yes I think that would be a correct description

If by "sharp" you mean having a cusp, with high derivatives immediately on either side, then the rippling is readily understood as the result of a tiny horizontal shift. That's interesting.

the model is violated by small imperfections in electronics and higher order physics and who knows what in the region of the peak where physical quantities are changing rapidly

it's not a cusp but it can be a relatively shapr peak

If I'm right, the problem might ultimately stem from using data too far from the peak in order to estimate the position of the peak.

strictly it's the convolution of a gaussian and an exponential decay

(It would only have to appear to be a cusp at the level of the horizontal resolution.)

@whuber: that is an interesting statement

about "using data too far from the peak in order to estimate the position of the peak"

I'm selecting a portion of the data containing the pulse and some amount of data before and after. If I allow the whole trace to be fitted, I get slightly different parameter values. The full trace has lots of data points in the pulse baseline

I'd forgotten about that aspect actually

@ColinK Yes: if you're trying to pin down the peak precisely, it makes (intuitive) sense to overweight data near the peak. The ripples in your residuals appear to be one clue about what's going on.

I've never been sure what to make of it

You could add a term in your model that allows for a small amount of (localized, nonlinear) shifting of the horizontal coordinate. Alternatively, you could use (say) an iterative reweighting method to improve the accuracy and robustness of the estimate of the peak's location.

right! I did consider weighting it in some way to do that, but then again my model includes the baseline, so I figured the contribution of those points to the gradient would vanish anyway, and then the weighting just made it more complicated to calculate parameter errors

That's reasonable, but from what I have heard, it sounds like you are looking for relatively high precision in locating the peak, so even a relatively small influence from the distant data might screw that up.

actually locating hte peak is not a major concern. The peak location is a fitted parameter, but only because I can't depend on it being in the same place each time. What I really care about is the pulse integral and the width parameter of the gaussian term in the convolution that produces the pulse shape

Ah...The width parameter is probably quite robust to errors in the peak position.

Indeed, I'm not seeing major errors there

I just want to provide the best error estimates I'm able to

The pulse integral is more problematic because (ordinarily) the estimate of baseline will have a heavy influence.

and understand the mathematicla background of what I'm doing a little better in the process

Any error in baseline is corrected by pre-processing steps

although I'd considered putting a baseline parameter into the fit. I think I tested it and the results weren't significantly improved

Right, but that correction is subject to a little bit of error that gets integrated across the entire pulse: it could add to a lot.

that's true

So although you probably cannot improve appreciably on the accuracy of what you're currently doing, it might help to incorporate baseline estimation in your model so you can estimate its error and its correlations with the other parameter estimates.

but in this case I haven't had any issues. There is so much baseline available in the data that I can subtract it out with high accuracy

hmm, that's not a bad idea

I'd see that in the off-diagonal elements of the covariance matrix, right?

Well, that's up to you: you know your data and your procedures and you have the best sense of what approaches are likely to be most appropriate for your needs.

@ColinK Yes, the covariance matrix yields the correlations.

I probably won't end up implementing it in the final code, but I'dd definitely do some tests, if only to verify that it's not causing any problems

One way to broach this subject on our site is to provide an illustration, some sample data, and a description of what you would like to accomplish, and (gently) challenge your readers. We often like a bit of a contest here :-).

Provided it doesn't take too much analysis, often you can get several people offering creative worked solutions.

If I see significant cross covariance, is there anything I should do with that knowledge? My understanding is that it's not necessarily a bad thing as long as the fit converges

I'd be happy to do that, depending on how much free time I find :) although the analytical gradients for this fit function are hideous

There's usually little to be done about correlation, but you want to measure it. When it's high, that means you probably need to study the joint error structure and not be content with looking at individual error variances.

@gung There are other factors as well - for example, looking purely at a high % of answers accepted discourages "supplemental answers", which is something I tend to do, and am rather alright with.

@whuber How would I go about using that in presenting the result to others, for example? If they use the pulse area to infer other information, but the baseline is physically meaningless, should they be interested in much other than the error bars on the area?

Ah, yes: the original question. That is more of a computing issue, because it seems to come down to how accurate the finite differences will be. Unless the gradient and Hessian are really misbehaved, you can often do just fine with the finite difference approximations.

(One concern about coding analytical solutions lies in the risk of making a tiny unnoticed coding error. Unless you test really thoroughly, that risk is high enough that you might favor finite differences for their simplicity alone.)

I tested the analytical gradients against finite differences

@ColinK I don't know the answer to that, partly because I am ignorant of the application. But if area is the sole estimate of interest, then you don't need the full covariance matrix, just the area error variance. The rest become "nuisance parameters."

the analytical gradients are pretty solid, and much faster than finite differences, obviously

I starred that ... because you're explaining what you care about most. That's worth bringing up.

@ColinK Right: comparing two totally different methods of calculation to each other is good. But the issue is with the thoroughness of the testing. There can be little error-accumulation or numerical-over/underflow bombs lurking that go off only when certain combinations of the arguments appear. How do you test against that possibility?

Oh, and then you deleted it.

@Glen_b: My question? Thank you :)

oh, not my question, nevermind

I starred the comment you made and deleted.

I haven't deleted any coments

@whuber: indeed, I can't test for that exhaustively, but the agreement with FD suggests that I haven't made any gross errors

Right here, just a minute or two ago.

Was it this one? "actually locating hte peak is not a major concern. The peak location is a fitted parameter, but only because I can't depend on it being in the same place each time. What I really care about is the pulse integral and the width parameter of the gaussian term in the convolution that produces the pulse shape" - That one is both starred and fairly far up the page

given the form of the derivatives it's likely that any errors would result in huge inaccuracy in the output, so I feel pretty good about it

not that that's a rigorous claim :)

@EpiGrad Supplemental answers are good. I still maintain that following the ideal I enunciated applies to supplemental answers anyway: even if there is no chance they will be accepted, make the OP wish they could accept them!

also I've already taken great pains to avoid floating point issues/rounding error, etc

@Colink My apologies. Something went odd with my scrolling.

np :)

@ColinK Good. But since that's the case, why are you concerned with one method versus the other? Use the one that is most expedient in terms of your time and energy.

Oh, but I did star your question. I want to follow it.

@whuber I think that's probably a fair point. I often aspire to be "Unaccepted answer with many upvotes" if there's already a solid answer that could use some expanding on, or especially if someone comes along with an analytic answer, because "Look, I brought proofs!" is something I'm rather terrible at.

mainly because I can't find any resources that cover this particular case. There are lots of people saying to use J'J in the context of SSE minimization, and a lot of people saying to use inv(H) in the context of MLE, and people saying to avoid a FD Hessian when you don't have any analytical derivatives

nobody talks about the case of having analytical Jacobian, but NOT an analytical hessian

I'm mainly interested in understanding which estimate of H is the better estimate

@EpiGrad There's even a gold badge for that--and you are one of the four to obtain one.

@whuber Indeed, this is what I've done in practice. I use the J'J method because I get the last jacobian out of my fitting routine for free, while the Hessian is not easily obtained, so I'd have to compute it myself after the optimization

@ColinK That's precisely the kind of point where I believe you are likely to get better-informed answers on compsci.

@whuber I confess to being terribly proud of that one - and speaks rather well of the "Don't go badge hunting" philosophy. That one happened entirely by mistake. As did my only other gold badge on Programmers.

In fact it was quite dificult to convince myself that in the SSE case the covariance matrix is sigma^2 inv(H), because most people that talk about inverting the hessian are working with a MLE estimate, which is equivalent to having that factor of sigma^2 in the error function

(for gaussian errors)

Anyway, do you think it would be best for me to edit my question and have it migrated to sciComp? Or would this be one of the rare cases where it would be best to make a more statistically flavored version for stats and a more computational flavored version for sciComp?

@EpiGrad Yet I am still happy with my answer--which is the one you trumped! Mine was even later than yours. You wound up (although you could not know it at the time) in the happy situation of either having your answer accepted or getting a gold badge if it were not accepted. :-)

@ColinK It depends on what you still need to know. If you think some of the statistical questions remain unresolved, then it sounds like you have a suitable question for our site. I don't think anybody would have problems with posting a related question on sciComp (or elsewhere on SE) that focused on the computing issues you face.

I suppose this would be more statistical in nature if I didn't have reason to believe my errors were gaussian. Since I know that (and correct me if I'm wrong) the MSSE and the MLE are identical, so this becomes more about optimization than statistics

If my erros were something nastier it might require a more heavy duty statustician

wow I can't spell today

@whuber You said: "in the happy situation of either having your answer accepted or getting a gold badge if it were not accepted".... but only if the OP accepted an answer at all. If not, bad luck.

I've got to run for the moment, but when I get back I'll bug the sciComp people to see if they'd like the question. If so, I'll edit it and flag it for migration. If not I'll clarify some of the statistical points and leave it here, since it seems there are some aspects that interest the stats community

Thanks for all your help, you guys are a great bunch as always

@Glen_b In fairness, I also got a bunch of rep :)

We should have moderator elections more often just for the interesting things that play out with the dynamics of the social aspects of this site. Or is it always like this in chat?

Took me 1 year to figure out there is a chat

@Momo chat's usually pretty quiet. I haven't seen more than 3-4 people for a good while

Must go

Bye

I took me more than a year to notice the chat feature, and I only started using it about a month ago. It's kind of addictive, but it's also another distraction...

I'm in good company then :)

Regarding the earlier discussion re BabakP's leaving the site, he (Tony Smith) just emailed me privately to say:

> I didn't leave because my feelings were hurt I left because I was spending way too much time on the site. I need to devote way more time to my own research and finishing my PhD so this site was causing me a much unneeded distraction (I was literately addicted to the site!)

Tony asked me to post this here on his behalf.

@GavinSimpson Thank you to keep us informed, Gavin. Alas, that's not the first time I hear such things... Other users have left for similar reasons. (However, in this particular case, applying for a mod position may just add confusion, IMO).

Given all the talk about accepted answers I (wasted time) and made a query of the data explorer to look at that.

data.stackexchange.com/stats/query/139329/…

If you order by accepted percent there are alot of gem individuals who have very few answers but a high proportion of them accepted.

Comparing percentage accepted to reputation cardinal appears to be a high leverage outlier. The correlation overall is pretty weak, but looks to me like consistent evidence that higher rep people have a higher percent of answers. Of those selected , the mean answer % is .31 (and the median .3).

Ignoring question views, it looks to me like IMO you should aim for 30%+ for your answers to be accepted (by individuals who answer questions regularly).

But that is just my guess at a guideline, Jeromy Anglim only has 22% of his answers accepted, and I general think he gives excellent answers.

@AndyW Answering self-study (as Glen pointed out), low quality or very focused questions may not contribute to increase such statistics. That's my impression, at least. However, I agree that it is a good indicator of your contribution in the long run. Voting activity is also essential to this site.

@AndyW, thanks for sharing this. One comment is that the measurements of "accept percantage" are inherently heteroskedastic, because the percentages based on a smaller numbers of observations are "noisier", and the noise level itself is positively associated with rep. It seems like this would tend to shrink your estimate of the correlation toward zero. Quantifying how much exactly seems complicated.

Also, to clarify, when you say the mean answer % is 31 - do you mean the sample average of the percents, or the overall percent of questions accepted?

@Macro I believe the 31% (not .31%!) value is intended to match the more current 30% figure I documented yesterday at chat.stackexchange.com/transcript/message/11389661#11389661. At any rate, the heteroscedasticity is easily handled with a binomial GLM.

The GLM would still be an approximation due to competition: when two or more individuals answer a question, only one can be accepted. It would be interesting to write down a model for this and fit it. I have noted that frequently when certain pairs of individuals answer a common question, a particular one of them almost always gets the acceptance (if one is offered at all): it's a kind of head-to-head competition :-).

@chl Of the 880 self-study questions, 327 (37.2%) have been accepted answers. Of all 25963 questions that have been asked, 9723 (37.4%) have accepted answers.

The difference is not significant (p = 0.93): chisq.test(matrix(c(25963-880,9723-327,880,327), nrow=2)).

I do, however, have the same impression that low-quality questions tend not to be accepted.

There are 19970 questions with a score of 1 or greater. 8498 (42.6%) of those have accepted answers. Thus there are 5993 low-quality questions with 1225 acceptances (20.4%): the difference is profound.

Although these data do not support the following conclusion, they are consistent with it: working to improve the quality of a question may improve the chances that an/your subsequent answer is accepted. So if you care about your acceptance rate, you have a choice: don't answer low-quality questions or (generously, optimistically) help improve them first.

(I hope my understanding of the search engine syntax is correct!)

hello everybody

@StatMan Hi!

I wonder now if this one should be closed?

I originally took this to have some statistical aspects, but as I was answering I began to doubt myself. Now it seems more like a "what is the right incantation?" or "can I do this in R?" -type question.

You know what I think, @Gavin :-)

It sounds very much like a "help me find the right switch" sort of question

@Macro OK - still calibrating my Cross Validated filter. Will vote to close.

It could've had a statistical component to it, but the user seems to know what s/he wants to do

What I'm unclear on (because I'm not a SO regular) is whether this should be migrated or not.

@Macro This could escape scrutiny on Stack Overflow but it could just as well get closed very quickly as it is basically amounts to a misunderstanding of what the OP thinks nlme can do.

@Gavin, well, if you were the moderator, would you have migrated it with your binding vote?

@Macro My gut feeling is this would be considered a poor Q on Stack Overflow and I would not migrate this one. I'd just close it here.

Cool

@Macro :-)

That wasn't meant to be a quiz. btw

Our convo last week was illuminating in terms of our migrated questions being a nuisance on SO

Because I hardly ever go there, I just sort of think of the question as "gone" when it gets migrated, not knowing/caring where it went

So that question was more directed at clarifying that policy (that policy that doesn't quite exist yet)

I think that view of migrating questions on Stack Exchange sites is quite common. There was some discussion on [meta.so] a while back about this issue. High traffic sites like Stack Overflow were getting lots of questions of a Sys Admin or general computing nature that were all being migrated to ServerFault or one of the other SE sites without the SO mods or the users voting to migrated knowing what the rules were about the site they were passing a question to.

Many questions migrated just got closed on the receiving site thus requiring more work for users there. There was general grumbling all round.

In the end, the mods were asked to consider whether a Q was suitable for the indicated migration. If not the Q was just closed.

That said, I think the mods here do a pretty good job in general with their migrations to Stack Overflow.

sure

that interaction did change my approach though. In the few relevant reviews I've done since then, I've been just closing instead of migrating

@Macro That was a general conclusion on Stack Overflow too. Close more than migrate, unless clearly on topic and appropriate for the intended site. That helped my approach a lot.

The SE chat notification sound is very unpleasant

@Macro I know what you mean. It's a nice ping on Stack Overflow chat.