GLMM model failing to converge
On Fri, Oct 16, 2015 at 3:24 PM, Shadiya Al Hashmi <saah500 at york.ac.uk> wrote:
Hello, I?m novice in using R in general and generalized logistic regression models with mixed effects in particular. At any rate, I?m testing how close the linguistic perception (response vowels) of different Turkish listeners (T [monolingual Turkish speakers], TA [bilingual Turkish-Arabic speakers] and TQ [Turkish speakers with some knowledge of Arabic through Quran recitation]) is to observed mappings (predicted vowels) in my research qualitative corpus. In the data, this is reflected in the binary variable match (1=match, 0=mismatch). Having said this, my dependent variable is ?match? which interacts with some +20 independent variables, some of which are factors with up to 12 levels. Now, the basic model I?ve used is as follows and works just fine. m0.1 <- glmer(match ~ Listgp + (1|Listener), data = PATdata1, family = "binomial") However, all subsequent models such as the one below crash. cf. m0.4 <- glmer(match~ Listgp + stimulus + st.context + st.length + age + gender + level.of.education + reading.A + comprehension.A + speaking.A + writing.A + (1|Listener), data = PATdata1, family = "binomial")
What does "crash" mean? Precision is important here -- you could mean: * a warning (which should certainly concern you, but it might be a false positive ...) -- in this case you *will* get a result, which you can use if you conclude that the warnings don't actually represent a serious problem; * an error -- in this case you won't get an answer at all, you need to deal with/work around the error before you can get results; * a true crash, where the R process actually stops. This is by definition a bug in the package, or (much more unlikely) in R itself.
Once I start parsing in the other factors especially the ones with mutli-levels such as ?stimulus? , the model fails to converge and I get a number of warning messages as follows. 1. fixed-effect model matrix is rank deficient so dropping 4 columns / coefficients?
This means you have collinear predictors -- most likely, some combinations of factors are aliased with each other.
2. In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv, : Model failed to converge with max|grad| = 0.151201 (tol = 0.001, component 7).
This is a "medium-sized" gradient; it may be OK, hard to know. How big is your data set? As suggested in ?convergence, the gold standard is to try your model with one or more different optimizers and see if it gets to a sufficiently (for your purposes) similar answer.
3. (function (fn, par, lower = rep.int(-Inf, n), upper = rep.int(Inf, : failure to converge in 10000 evaluations
See ?lmerControl for advice about how to increase the number of evaluations.
Any advice on how to go about this?
Thank you,
Shadiya
[[alternative HTML version deleted]]
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models