PIRLS step-halvings failed to reduce deviance in pwrssUpdate with probit relmatGlmer

[zoeluo15]

Hi, I am using the relmatGlmer function with the probit link function in lme4qtl. Here is my data and code

kinship.txt genotype_phenotype.txt

test<-read.table("genotype_phenotype.txt")
kinship<-as.matrix(read.table("kinship.txt")) colnames(kinship)<-rownames(kinship) relmatGlmer(phenotype ~ genotype + (1|ID), test, relmat = list(ID = kinship), family = binomial(link="probit"))

And I got the following error message:

Error in pwrssUpdate(pp, resp, tol = tolPwrss, GQmat = GQmat, compDev = compDev, : (maxstephalfit) PIRLS step-halvings failed to reduce deviance in pwrssUpdate

I found a relevant question in lme4 https://github.com/lme4/lme4/issues/579, and tried their solution

relmatGlmer(phenotype ~ genotype + (1|ID), test, relmat = list(ID = kinship), family = binomial(link="probit"),nAGQ=20)

It returns the error message:

Error in pwrssUpdate(pp, resp, tol = tolPwrss, GQmat = GQmat, compDev = compDev, : AGQ only defined for a single scalar random-effects term

I think my random effect term is a scalar random effect. Am I doing anything wrong here?

Any help would be appreciated!

Zoe

[variani]

Hi Zoe,

I'm not sure what is exactly happening here. Looking at the issue for lme4 (you mentioned), the cause seems to be in a low number of observations. I wonder if you can simulate data, increase the number of observations and check if the issue is gone.

Best, Andrey

[zoeluo15]

Hi Andrey,

There is indeed no problem once I increase the number of observations in each group. But I would like to fit the model with an individual random effect (i.e. one observation per level of the random effect) because this is actually a real bird dataset.

I think it might be the distribution of the genotype (the fixed effect) as there is no problem fitting the same model with genotypes at another position (test2.txt).

table(test$genotype) #this is the dataset I attached in the original post

0 1 2 82 64 8

table(test2$genotype) #this is the new one that has no problem in model fitting

0 1 2 136 17 1

Many thanks, Zoe

[variani]

Hi Zoe,

Unfortunately, I cannot help too much here, as I was not the person who implemented the model fitting algorithm involved.

Perhaps, you can play with glmerControl. In particular, there are two algorithms for the optimizer (the first argument of glmerControl).

Best, Andrey

[zoeluo15]

Thanks! I will have a try with glmerControl. - Zoe