How to include autocorrelation in GLMM and what to do with, false convergence warning message appearing after including an, interaction term (Lucia)
I will try to explain my data as good as possible.
So we taged 13 different whales with a tag that records time, depth, speed, angle of descent and ascent 25 samples every second. The normal diving behaviour of these animals is one deep dive of one hour to 1200 meters followed by a series of 3-7 shallow dives of 20 minutes up to 300 m. Because the tag not always stays the same time in each animal, my data is unbalanced and some tag records have one deep dive and 6 shallow dives while other records have 7 deep dives and 26 shallow dives. I divided each dive in units of 30 seconds. for each unit I have the next data: whale number, dive number, total number of fluke strokes in the 30 seconds unit of analysis, mean of the sin of the angle during the 30 seconds unit, swim speed, dive type(ascent or descent), dive direction( if it is a descent or an ascent) and time since the start of the dive and finally my variable response which is presence or absence of one type of fluke stroke called stroke type B. I think there has to be some autocorrelation between each 30 seconds unit of analyis and need to include it in my model but do not know how! I am interested to know what affects the presence or absence of the type B stroke (which is a binomial variable with 0 and 1) so I decide to use a binomial glmm with whale number as a random effect. I included as well dive number within whale as a random effect. here is the model. glmm114<-lmer (StrokeB~ Time * Depth+SINP+flukes*Depth+speed+(1|whale_number)+(0+dive_number|whale_number),data=Luciadeepas, family = binomial)
13 tagged animals may represent a lot of work....but it is not that much for a random intercept (and slope!). Perhaps dropping the random slope may avoid the warning message. Also...depth and speed sound like "collinearity"??
this is my final model after taking out the non significative variables, the problem is that due to the interaction a problem appears saying
Check whether the quality/quantity of your data supports an interaction. There is always a reason if there is trouble.
The false convergence warning message (8) I looked on internet and it says is a common problem,
It is a common problem that people use too complicated models..:-) Or have data that are too limited for the type of models they want to use. and some people
says that it doesnt make any change in the output while others says that each variable has to be divided by 100.
Better not believe everything people say. Divide by 100...huh???? but when I do this then the variables that become
significant doesnt make any sense.
I don;t know what you divided by 100.....but in principle dividing the continuous covariates by 100 shouldn't change things that much. Unless you have some serious model misspecifications. Perhaps you want to upload some data. Alain >>
Thanks so much in advanced Lucia
Dr. Alain F. Zuur First author of: 1. Analysing Ecological Data (2007). Zuur, AF, Ieno, EN and Smith, GM. Springer. 680 p. URL: www.springer.com/0-387-45967-7 2. Mixed effects models and extensions in ecology with R. (2009). Zuur, AF, Ieno, EN, Walker, N, Saveliev, AA, and Smith, GM. Springer. http://www.springer.com/life+sci/ecology/book/978-0-387-87457-9 3. A Beginner's Guide to R (2009). Zuur, AF, Ieno, EN, Meesters, EHWG. Springer http://www.springer.com/statistics/computational/book/978-0-387-93836-3 4. Zero Inflated Models and Generalized Linear Mixed Models with R. (2012) Zuur, Saveliev, Ieno. http://www.highstat.com/book4.htm Other books: http://www.highstat.com/books.htm Statistical consultancy, courses, data analysis and software Highland Statistics Ltd. 6 Laverock road UK - AB41 6FN Newburgh Tel: 0044 1358 788177 Email: highstat at highstat.com URL: www.highstat.com URL: www.brodgar.com