Correct estimation of arguments for glmmLasso function I am using glmmLasso for variable selection. In my case, n is slightly less than p and p are bioclimatic variables for different time periods, so are highly correlated. 
How do I choose the right values for the arguments: lambda and control. I have tried with different values of lambda, maxIter (even to 10,000) and control (start and steps). But, the algorithm never converges. The p.values of all the selected variables are 0 and that makes me wonder if the non-convergence is the cause. What factor could help achieve convergence in a reasonable number of iterations? 
Among the variables selected, since p.values are all 0 (the rest of the variables are NA), can I estimate the relative variable importance based on the StdErr?
Also, the lambda resulting in the lowest BIC values selects too many variables as being significant. Would it be OK in this case,(as I am only doing a rough variable selection for modelling) to not worry about the BIC, but choose the lambda that gives me a reasonable number of variables that also make sense given my data.
What is the start arguments in control, anyway? All zeros are not acceptable, and so are vectors below a certain length. What do the values and vector length depend on? Apologies, if that shows ignorance of the mathematics behind mixed-models...
For details on the data, please refer to Variable selection using mixed-models (lme4)
I went through the GMMBoost, also by Groll, but did not find something to guide me in this case. Has anyone used glmmLasso for analysis and faced similar situations? Glad to hear any suggestions
 A: the glmmLasso package contains a demo file, where several strategies are shown, how the optimal tuning parameter lambda. I suggest to start with a high values for lambda, such that all covariates are set to zero and then reduce it step by step. It increase speed, if the final coefficient estimates corresponding to a lambda are used as starting values for the next smaller lambda.
The argument start of the control list is a vector of starting values for all model parameters, i.e. all fixed effects (usually including the intercept) followed by all random effects (and thus depending on the length of your gouping factor and on the structure of your random terms), see the documentation of glmmLassoControl. For example, if you fit the following GLMM with a simple random intercept:
fit <- glmmLasso(y ~ x1 + ... + x10, rnd = list(id=~1), data = ..., ... ) 
and suppose your grouping factor id contains 80 individuals, then the start vector would have to be of length 1(intercept)+10(fixed effects)+80(random effects)=91.
In my experience, the model corresponding to the lambda with the lowest BIC is usually quite sparse. So what I would suggest is to check first, if really all covariates have been standardized in adavnace to have variance one. Furthermore, have a look on the coefficient paths for your sequence of lambda values to check if they build up as one would typically expect for LASSO.
