Skip to content
Prev 53899 / 63421 Next
Original Raw

Wrongly converging glm()

Harm-Jan Westra 
My apologies if I seemed to ?blame R?. This was in no way my intention. I get the feeling that you?re missing my point as well.

I observed something that I thought was confusing, when comparing two more or less identical methods (when validating the C code), and wanted to make a suggestion as to how to help future R users. Note that I already acknowledged that my data was bad. Note that I also mention that the way R determines convergence is a valid approach.

What strikes me as odd is that R would warn you when your data is faulty for a function such as cor(), but not for glm(). I don?t see why you wouldn?t want to check both convergence criteria if you know multiple of such criteria exist. It would make the software more user friendly in the end.

It may be true that there are millions of edge cases causing issues with glm(), as you say, but here I am presenting an edge case that can be easily detected, by checking whether the difference in beta estimates between the current and previous iteration is bigger than a certain epsilon value.

I agree ?that everybody using R should first do the effort of learning what they're doing?, but it is a bit of a non-argument, because we all know that, the world just doesn?t work that way, plus this is one of the arguments that has held for example the Linux community back for quite a while (i.e. let?s not make the software more user friendly because the user should be more knowledgeable).

Harm-Jan


From: Joris Meys<mailto:jorismeys at gmail.com>
Sent: Thursday, July 20, 2017 13:16
To: Harm-Jan Westra<mailto:westra.harmjan at outlook.com>
Cc: r-devel at r-project.org<mailto:r-devel at r-project.org>
Subject: Re: [Rd] Wrongly converging glm()
On Thu, Jul 20, 2017 at 6:21 PM, Harm-Jan Westra <westra.harmjan at outlook.com<mailto:westra.harmjan at outlook.com>> wrote:
Dear Joris,


I agree that such a covariate should not be used in the analysis, and fully agree with your assessment. However, your response assumes that everybody who uses R knows what they're doing, which is a dangerous assumption to make. I bet there are a lot of people who blindly trust the output from R, even when there's clearly something wrong with the estimates.

You missed my point then. I don't assume that everybody who uses R knows what they're doing. Actually, I know for a fact quite a few people using R have absolutely no clue about what they are doing. My point is that everybody using R should first do the effort of learning what they're doing. And if they don't, one shouldn't blame R. There's a million different cases where both algorithms would converge and the resulting estimates are totally meaningless regardless. R cannot be held responsible for that.



In terms of your conclusion that the C++ estimate corresponds to a value within the R estimated confidence interval: if I allow the C++ code to run for 1000 iterations, it's estimate would be around -1000. It simply never converges.

I didn't test that far, and you're right in the sense that -100 is indeed not the final estimate. After looking at the C code, it appears as if the author of that code combines a Newton-Raphson approach with a different convergence rule. And then it's quite understandible it doesn't converge. You can wildly vary that estimate, the effect it has on the jacobian, log likelihood or deviance will be insignificant. So the model won't improve, it would just move all over the parameter space.



I think there's nothing wrong with letting the user know there might be something wrong with one of the estimates, especially if your code can easily figure it out for you, by adding an additional rule. Not everyone is always paying attention (even if they know what they're doing).

If R would do that, it wouldn't start the fitting procedure but just return an error "Your analysis died due to a lack of useable data." . Because that's the problem here.



With kind regards,


Harm-Jan
Message-ID: <DM3PR20MB05728541A9D1DD21D3354C0887A70@DM3PR20MB0572.namprd20.prod.outlook.com>
In-Reply-To: <CAO1zAVaWouvQ-p3kjxhH7V3RR5qP_3fwb4Je4KZXcyaZDfotqQ@mail.gmail.com>
Back to archive Original archive Thread
7 / 11

Thread (11 messages)

Harm-Jan Westra Wrongly converging glm() Jul 20 Joris Meys Wrongly converging glm() Jul 20 Harm-Jan Westra Wrongly converging glm() Jul 20 Simon Bonner Wrongly converging glm() Jul 20 Harm-Jan Westra Wrongly converging glm() Jul 20 Joris Meys Wrongly converging glm() Jul 20 Harm-Jan Westra Wrongly converging glm() Jul 20 Mark Leeds Wrongly converging glm() Jul 20 Harm-Jan Westra Wrongly converging glm() Jul 20 Joris Meys Wrongly converging glm() Jul 20 Harm-Jan Westra Wrongly converging glm() Jul 20