An embedded and charset-unspecified text was scrubbed... Name: not available URL: <https://stat.ethz.ch/pipermail/r-sig-mixed-models/attachments/20111213/59ea8200/attachment.pl>
False convergence in a mixed model.
5 messages · ONKELINX, Thierry, Luca Borger, staffan at myrica.se +1 more
1 day later
Dear Staffan, False convergence is often due to a model that is too complex for the data. Do you have enough data to support that four-way interaction. And if so do you really want a four-way interaction? Another problem might be complete separation: for a given combination of factors all responses are either 0 or 1 Best regards, Thierry ir. Thierry Onkelinx Instituut voor natuur- en bosonderzoek / Research Institute for Nature and Forest team Biometrie & Kwaliteitszorg / team Biometrics & Quality Assurance Kliniekstraat 25 1070 Anderlecht Belgium Thierry.Onkelinx at inbo.be www.inbo.be To call in the statistician after the experiment is done may be no more than asking him to perform a post-mortem examination: he may be able to say what the experiment died of. ~ Sir Ronald Aylmer Fisher The plural of anecdote is not data. ~ Roger Brinner The combination of some data and an aching desire for an answer does not ensure that a reasonable answer can be extracted from a given body of data. ~ John Tukey -----Oorspronkelijk bericht----- Van: r-sig-mixed-models-bounces at r-project.org [mailto:r-sig-mixed-models-bounces at r-project.org] Namens staffan at myrica.se Verzonden: dinsdag 13 december 2011 10:34 Aan: r-sig-mixed-models at r-project.org Onderwerp: [R-sig-ME] False convergence in a mixed model. Hi there! I?m a quite new R-user, ecpecially when it comes to mixed models. Currently I?m doing my master thesis and I have built a glmer model as seen below. But I recive a warning message "Warning message: In mer_finalize(ans): false convergence (8)" and can?t get rid of it. I have gone trough the mailing list on this subject and googled, but the answers I find (which offcourse are good) are to difficult for me to interpret. Now I?m sending you this email in hope of a more simple answer that I can interpret. My initial data looks like this; X Y Status Size Density Trial Summarize( I have 3 densities 8,16 and 32, 12 trials A-L, status is either 1 or 0, and X,Y and size are measures. 27,7 903 1 2.85454545454545 32 C 1 855 0 3.5 8 D 54,7 796 0 3.12727272727273 32 C 113,6 3 1 2.94545454545455 16 B 32,5 863 0 3.09433962264151 32 C 283,5 808 0 3.07142857142857 8 D 281,5 898 1 3.16363636363636 8 D 8,6 845 0 2.90909090909091 32 C and I use the model: model1<-glmer(Status~1+(X*Y*Size*Density)+(1|Trial),data=data,family=binomial) In other answers I have read about "verbose=T" but this does not change the warning. Neither does "scale=T". Something about "mer_optimize" is also spoken about, but I do not understand what I am supposed to do with that. I get some significans in the model, and the very pin pointed question is....can I trust these results even tough the warning? I?m soon about to present my project, so I would very much appritiate a quick answer. A Huge thanks! //Staffan svanberg
Hello, I've found that sometimes you can get the models to converge by increasing the value for nAGQ (the default is equal to one, see ?glmer). However, even if this will lead your model to (apparently) converge well, I would seriously consider Thierry's remarks. HTH Luca ------------------------------------------- Luca Borger Postdoctoral Research Fellow CNRS - Centre d'Etudes Biologiques de Chiz? Villiers-en-Bois 79360 Beauvoir-sur-Niort France Tel: +33 (0)549 09 96 13 Fax: +33 (0)549 09 65 26 email: lborger at cebc.cnrs.fr http://www.researcherid.com/rid/C-6003-2008 http://cnrs.academia.edu/LucaBorger ------------------------------------------- Le 14/12/2011 11:03, ONKELINX, Thierry a ?crit :
Dear Staffan, False convergence is often due to a model that is too complex for the data. Do you have enough data to support that four-way interaction. And if so do you really want a four-way interaction? Another problem might be complete separation: for a given combination of factors all responses are either 0 or 1 Best regards, Thierry ir. Thierry Onkelinx Instituut voor natuur- en bosonderzoek / Research Institute for Nature and Forest team Biometrie& Kwaliteitszorg / team Biometrics& Quality Assurance Kliniekstraat 25 1070 Anderlecht Belgium Thierry.Onkelinx at inbo.be www.inbo.be To call in the statistician after the experiment is done may be no more than asking him to perform a post-mortem examination: he may be able to say what the experiment died of. ~ Sir Ronald Aylmer Fisher The plural of anecdote is not data. ~ Roger Brinner The combination of some data and an aching desire for an answer does not ensure that a reasonable answer can be extracted from a given body of data. ~ John Tukey -----Oorspronkelijk bericht----- Van: r-sig-mixed-models-bounces at r-project.org [mailto:r-sig-mixed-models-bounces at r-project.org] Namens staffan at myrica.se Verzonden: dinsdag 13 december 2011 10:34 Aan: r-sig-mixed-models at r-project.org Onderwerp: [R-sig-ME] False convergence in a mixed model. Hi there! I?m a quite new R-user, ecpecially when it comes to mixed models. Currently I?m doing my master thesis and I have built a glmer model as seen below. But I recive a warning message "Warning message: In mer_finalize(ans): false convergence (8)" and can?t get rid of it. I have gone trough the mailing list on this subject and googled, but the answers I find (which offcourse are good) are to difficult for me to interpret. Now I?m sending you this email in hope of a more simple answer that I can interpret. My initial data looks like this; X Y Status Size Density Trial Summarize( I have 3 densities 8,16 and 32, 12 trials A-L, status is either 1 or 0, and X,Y and size are measures. 27,7 903 1 2.85454545454545 32 C 1 855 0 3.5 8 D 54,7 796 0 3.12727272727273 32 C 113,6 3 1 2.94545454545455 16 B 32,5 863 0 3.09433962264151 32 C 283,5 808 0 3.07142857142857 8 D 281,5 898 1 3.16363636363636 8 D 8,6 845 0 2.90909090909091 32 C and I use the model: model1<-glmer(Status~1+(X*Y*Size*Density)+(1|Trial),data=data,family=binomial) In other answers I have read about "verbose=T" but this does not change the warning. Neither does "scale=T". Something about "mer_optimize" is also spoken about, but I do not understand what I am supposed to do with that. I get some significans in the model, and the very pin pointed question is....can I trust these results even tough the warning? I?m soon about to present my project, so I would very much appritiate a quick answer. A Huge thanks! //Staffan svanberg [[alternative HTML version deleted]]
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
__________ Information from ESET Mail Security, version of virus signature database 6710 (20111214) __________ The message was checked by ESET Mail Security. http://www.eset.com
An embedded and charset-unspecified text was scrubbed... Name: not available URL: <https://stat.ethz.ch/pipermail/r-sig-mixed-models/attachments/20111214/03f8999f/attachment.pl>
1 day later
staffan at ... <staffan at ...> writes:
#I tried to increase the nAGQ without success however. #The reason I have the four-way interaction is that I wanted to start off with a full model which would be the model: model1<-glmer(Status~1+(X*Y*Size*Density)+(1|Trial), data=data,family=binomial)
the "1" is redundant (the intercept is included by default), but also harmless
I have then used stepwise backward elimination of non-significant terms, the four-way interaction is the first one to go. Even though I take away this term and some more terms the warning message is still there. I believe that I have enough data, with 112 observations on each measured variable (X, Y and size).
It depends: "enough data" can depend on the structure of the data as well as the raw amount. See various discussions on this list of the potential dangers of backward stepwise regression ...
Another problem might be complete separation: for a given combination of factors all responses are either 0 or 1 I do not fully understand your point (I appologize). Do you mean that e.g. the interaction X*Y explains the probability of survival to 100 % (1) or not at al (0)? If it is not much of a problem, I would appreciate an example. The response variable survival is either 1 (alive) or 0 (dead).
Here is an example: X Y Status 1.1 1.1 0 1.2 1.2 0 1.1 5 1 4.3 1.2 0 4.2 1.7 1 5.2 2.4 1 5.4 1.7 0 5.7 6.5 1 6.2 6.1 1 6.5 6.4 0 If you plot this out, you'll see that if X and Y are both small, then the Status is always zero. This leads to
My variables are very highly correlated in the model summary, in a range 0.806-0.998, I guess this has to do with the warning message and Thierry's point above, and another proof to my model being incorrect? I would love to understand how I could improve my model and get rid of that warning, however, the basic question is: Can I interpret the results I have and feel confident about the significant terms? Best regards //Staffan (since the data did not make sense in the reply, I ad it once more so that you
can see the structure)
X Y survival size density trial 1 391 0 3.41818181818182 8 A 167.9 874 1 3.01785714285714 8 A 40.6 874 1 2.95374013014014 8 A 54.3 739 0 2.76363636363636 32 C 27.7 903 1 2.85454545454545 32 C 1 855 0 3.5 8 D 54.7 796 0 3.12727272727273 32 C 113.6 3 1 2.94545454545455 16 B 32.5 863 0 3.09433962264151 32 C 283.5 808 0 3.07142857142857 8 D 281.5 898 1 3.16363636363636 8 D
Summarize( I have 3 densities 8,16 and 32, 12 trials A-L, status is either 1 or 0, and X,Y and size are measures.
Can you post the full data set somewhere?