mixed mutlinomial regression for count data with, overdisperion & zero-inflation
Hi Dave, Thanks for all this detailed info. I'm on Mac unfortunately. But as the model with nbinom2 runs and is a better fit, I'll inspect the rep file fo this one. Or is there a way to access this directly on R? Best, Stephanie
On 20 May 2016 at 05:05, dave fournier <davef at otter-rsch.com> wrote:
I found a version of glmmadmb for Windows 64bit ( I sadly assume) that almost does the job out of the box. Using the information from R Forge http://glmmadmb.r-forge.r-project.org/ which points to http://www.admb-project.org/buildbot/glmmadmb/ where I found this exe glmmadmb-mingw64-r3274-windows10-mingw64.exe Using your glmmadmb.pin and glmmadmb.dat files (which you renamed to glmmadmb1.pin and glmmadmb1.dat) I ran the following script. ./glmmadmb-mingw64-r3274-windows10-mingw64.exe -crit 1.e-6 -ind glmmadmb1.dat -ainp glmmadmb.par -maxph 5 -shess -noinit -phase 7 (I have added two extra phases and modified the convergence criterion to 1.e-6.) which converged with nice estimates etc. Now one of the really nice things about actually fitting the model rather than resorting to other diagnostics is that if you are successful you can look at the fit. In this case the squared residuals divided by the predicted variance. These are in the output glmmadmb.rep file. I sorted them in R and looked at the large ones at the end. index obs pred whatever its called 1206 1206 319 3.47536e+01 1.24498e+01 305 305 4490 2.18655e+03 1.29949e+01 1074 1074 413 5.13552e+01 1.36380e+01 1385 1385 4002 1.68879e+03 1.69679e+01 854 854 224 1.22219e+01 1.96515e+01 1691 1691 2713 8.33316e+02 2.27056e+01 1427 1427 1732 3.92621e+02 2.44684e+01 1433 1433 1612 3.25266e+02 2.72590e+01 1313 1313 1815 3.52356e+02 3.25137e+01 341 341 2031 3.55824e+02 4.22336e+01 191 191 5814 7.18097e+02 1.93656e+02 599 599 3586 2.68911e+02 2.19118e+02 So you have a few gigantic outliers. This is fairly common with data for which the model has problems converging. But rather than celebrating the failure of the model as a useful diagnostic for bad data it is really useful to coax it to fit the data and investigate the residuals, If you can explain them it should help.
*St?phanie PERIQUET (PhD) * - Bat-eared Fox Research Project *Dept of Zoology & Entomology* *University of the Free State, Qwaqwa Campus* *Cell: +27 79 570 2683* ResearchGate profile <https://www.researchgate.net/profile/Stephanie_Periquet> Kalahari bat-eared foxes on Twitter <https://twitter.com/kal_batearedfox> [[alternative HTML version deleted]]