Model convergence problem singular convergence (7). 'giveCsparse' has been deprecated setting 'repr = 'T'' for youĢ: In Matrix::sparseMatrix(dims = c(0, 0), i = integer(0), j = integer(0), :ģ: In Matrix::sparseMatrix(dims = c(0, 0), i = integer(0), j = integer(0), : Call() for package 'glmmTMB'ġ: In Matrix::sparseMatrix(dims = c(0, 0), i = integer(0), j = integer(0), : Call('FreeADFunObject', ptr, PACKAGE = DLL) : > glmmTMB_m = glmmTMB(y ~ x + (1 + x | group ), family = binomial(link = 'logit'), data = dat)Įrror in. Optimizer (Nelder_Mead) convergence code: 0 (OK) Generalized linear mixed model fit by maximum likelihood (Laplace Approximation) Ĭontrol: glmerControl(nAGQ0initStep = FALSE) These issues often make indispensable resources. Thank you to Ben Bolker and other contributors here and elsewhere. As a side note, for complete information, I'll note that I used nAGQ = 0 in a larger, non-converging glmer model, and it did not yield convergence in that case. An nAGQ > 1 'is only available for models with a single, scalar random-effects term', and it can sharply increase the running time.Īnother good aspect is the fact that the nAGQ = 0 model is faster and not less precise than the others, which have the default nAGQ = 1. The last model was the only one that did not produce any convergence warnings.Īside from convergence warnings, the results are nearly identical across the four models (β x = 0.039, p 1 was instead required to facilitate convergence (see nAGQ > 1 advice but also see nAGQ = 0 advice, and a related explanation). (2019), I ran a model using nAGQ = 0 (called 'nAGQ_m'). In addition, following Daniel Lüdecke's advice elsewhere, I ran a model using glmmTMB::glmmTMB() (called 'glmmTMB_m'). I have run the two models discussed above using the 'case1' data (see 'nAGQ0initStep_m' and 'tolPwrss_m' below). Crucially, though, the present convergence warnings are also common with more varied data. Firstly, I'd like to note that the 'case1' data may be an extreme case, in that it presents very scarce variation, with most values of x and y being 0. I am very interested in the present issue. Glmer2 <- glmer(y ~ x + (1 + x | group ), family = binomial(link = 'logit'), data = data, control = glmerControl(tolPwrss=1e-3))Įstimates and random variances are very different

Glmer1 <- glmer(y ~ x + (1 + x | group ), family = binomial(link = 'logit'), data = data, start=list(fixef=coef(glm1)), control=glmerControl(nAGQ0initStep=FALSE))

Glm1 <- glm(y ~ x, family = binomial(link = 'logit'), data = data) Model <- glmer(y ~ x + (1 + x | group ), data = data, family = binomial(link = 'logit'))

The variances of the random slope are also very different (lower in the second approach). As you will see from the data, estimates are different, even with different sign in the averaged slope. The model using tolPwrss seems to be better based on the log-likelihood (less negative value), but they were not too different: -1224.79 for your approach vs -1218.102 using tolPwrss. How different are the log-likelihoods? Can you provide me with the data or a reproducible sure! I attach the data because couldn't get a reprex to show this case. Which approach gave you a better log-likelihood? In general, the approach with the highest log-likelihood is preferred.

Gamma control loop