Dear colleagues, I have performed several dozens of glm.nb(response ~ variable) analyses weeks ago, and when I looked through the results today I saw that many of the results have quite different intercept values despite the response part remained the same. I'm quite sure I did same kind of analysis when the intercept values were around consistently around 2.2 and when they were above 3. When I repeated the analyses today, the intercept values were normal, they were between 2.1 to 2.3 instead of being above 3. I'm standing in front of a puzzle... they surely aren't glm() results, for they would give intercept values well above 9. Is there anything like a set.seed() thing that could have changed some properties inside R? On a second look, I discovered that the init.theta value is much lower in those analyses I have to perform again. Does anybody have a clue to this problem? It isn't that important that I have an answer (because I simply have to repeat the analyses), but still... David
glm.nb() giving strongly different results
4 messages · David Croll, Bill Venables
From what you tell us it is impossible even to see if there is a problem, let alone what it might be if there is one. There are all kinds of reasons why intercepts may change and it is only unexpected if you do not fully understand what the intercept parameter really is. For example, if you change a predictor variable to have a different centre, x -> x-c, you will not change the regression coefficient with respect to x, but by varying c you can make the intercept anything you like. Literally. Anything. And this is nothing whatever to do with glm.nb, it applies equally to glm, lm, aov, ...
I can console you on one point, though. glm.nb does not use a stochastic algorithm, and so no random numbers are involved. So unless you are generating fake data, the random number generator should play no part. Bill Venables http://www.cmis.csiro.au/bill.venables/ -----Original Message----- From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-project.org] On Behalf Of David Croll Sent: Wednesday, 25 March 2009 12:36 PM To: r-help at r-project.org Subject: [R] glm.nb() giving strongly different results Dear colleagues, I have performed several dozens of glm.nb(response ~ variable) analyses weeks ago, and when I looked through the results today I saw that many of the results have quite different intercept values despite the response part remained the same. I'm quite sure I did same kind of analysis when the intercept values were around consistently around 2.2 and when they were above 3. When I repeated the analyses today, the intercept values were normal, they were between 2.1 to 2.3 instead of being above 3. I'm standing in front of a puzzle... they surely aren't glm() results, for they would give intercept values well above 9. Is there anything like a set.seed() thing that could have changed some properties inside R? On a second look, I discovered that the init.theta value is much lower in those analyses I have to perform again. Does anybody have a clue to this problem? It isn't that important that I have an answer (because I simply have to repeat the analyses), but still... David ______________________________________________ R-help at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
Thank you, Bill, for your answer! I am also at a total loss when looking for an explanation. I just can't remember what I did differently... At least the errors are confined to a rather small dataset so the repetition of all the glm.nb() analyses won't take much time. The only thing I found out so far is that the problem appeared at binary explanatory variables to which the full set of study participants contributed their answers, but the problem did not occur when analysing binary variables where only the employed people contributed to the dataset... Either employed people are some kind of magicians or I drank too little coffee to understand that I *really* did something different during my R work... Thanks again, David
From what you tell us it is impossible even to see if there is a problem, let alone what it might be if there is one. There are all kinds of reasons why intercepts may change and it is only unexpected if you do not fully understand what the intercept parameter really is. For example, if you change a predictor variable to have a different centre, x -> x-c, you will not change the regression coefficient with respect to x, but by varying c you can make the intercept anything you like. Literally. Anything. And this is nothing whatever to do with glm.nb, it applies equally to glm, lm, aov, ...
I can console you on one point, though. glm.nb does not use a stochastic algorithm, and so no random numbers are involved. So unless you are generating fake data, the random number generator should play no part. Bill Venables http://www.cmis.csiro.au/bill.venables/ -----Original Message----- From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-project.org] On Behalf Of David Croll Sent: Wednesday, 25 March 2009 12:36 PM To: r-help at r-project.org Subject: [R] glm.nb() giving strongly different results Dear colleagues, I have performed several dozens of glm.nb(response ~ variable) analyses weeks ago, and when I looked through the results today I saw that many of the results have quite different intercept values despite the response part remained the same. I'm quite sure I did same kind of analysis when the intercept values were around consistently around 2.2 and when they were above 3. When I repeated the analyses today, the intercept values were normal, they were between 2.1 to 2.3 instead of being above 3. I'm standing in front of a puzzle... they surely aren't glm() results, for they would give intercept values well above 9. Is there anything like a set.seed() thing that could have changed some properties inside R? On a second look, I discovered that the init.theta value is much lower in those analyses I have to perform again. Does anybody have a clue to this problem? It isn't that important that I have an answer (because I simply have to repeat the analyses), but still... David
______________________________________________ R-help at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
I don't think I made myself clear, though. With a series of models fitted using the same form fm <- glm.nb(response ~ variable, data) why would you expect them to have similar intercepts? I just don't get it. You could try fitting the same model in a different way using, e.g. fm <- glm.nb(response ~ I(variable - mean(variable)), data) Now the intercepts really should all be exactly the same, and the slope parameters should not have change from the previous form. Can you check that? The problem with fitting models in this form, though, is that you cannot easily predict from the resulting fitted model object. Bill Venables http://www.cmis.csiro.au/bill.venables/ -----Original Message----- From: David Croll [mailto:david.croll at gmx.ch] Sent: Wednesday, 25 March 2009 7:28 PM To: Venables, Bill (CMIS, Cleveland); r-help at r-project.org Subject: Re: [R] glm.nb() giving strongly different results Thank you, Bill, for your answer! I am also at a total loss when looking for an explanation. I just can't remember what I did differently... At least the errors are confined to a rather small dataset so the repetition of all the glm.nb() analyses won't take much time. The only thing I found out so far is that the problem appeared at binary explanatory variables to which the full set of study participants contributed their answers, but the problem did not occur when analysing binary variables where only the employed people contributed to the dataset... Either employed people are some kind of magicians or I drank too little coffee to understand that I *really* did something different during my R work... Thanks again, David
From what you tell us it is impossible even to see if there is a problem, let alone what it might be if there is one. There are all kinds of reasons why intercepts may change and it is only unexpected if you do not fully understand what the intercept parameter really is. For example, if you change a predictor variable to have a different centre, x -> x-c, you will not change the regression coefficient with respect to x, but by varying c you can make the intercept anything you like. Literally. Anything. And this is nothing whatever to do with glm.nb, it applies equally to glm, lm, aov, ...
I can console you on one point, though. glm.nb does not use a stochastic algorithm, and so no random numbers are involved. So unless you are generating fake data, the random number generator should play no part. Bill Venables http://www.cmis.csiro.au/bill.venables/ -----Original Message----- From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-project.org] On Behalf Of David Croll Sent: Wednesday, 25 March 2009 12:36 PM To: r-help at r-project.org Subject: [R] glm.nb() giving strongly different results Dear colleagues, I have performed several dozens of glm.nb(response ~ variable) analyses weeks ago, and when I looked through the results today I saw that many of the results have quite different intercept values despite the response part remained the same. I'm quite sure I did same kind of analysis when the intercept values were around consistently around 2.2 and when they were above 3. When I repeated the analyses today, the intercept values were normal, they were between 2.1 to 2.3 instead of being above 3. I'm standing in front of a puzzle... they surely aren't glm() results, for they would give intercept values well above 9. Is there anything like a set.seed() thing that could have changed some properties inside R? On a second look, I discovered that the init.theta value is much lower in those analyses I have to perform again. Does anybody have a clue to this problem? It isn't that important that I have an answer (because I simply have to repeat the analyses), but still... David
______________________________________________ R-help at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.