Help understanding warning message

[ChristelleJammar]

Hi,

I am currently working with functional trait data and using the gllvm package to perform a fourth-corner analysis. I am fitting multiple models (including negative binomial and zero-inflated negative binomial distributions) and comparing them using AIC to select the best-supported model.

However, when fitting the negative binomial models (both with 1 and 2 latent variables), I consistently obtain the following warning:

simpleWarning in EvalADFunObject(ADFun, theta, order = order, hessiancols = cols, : underflow occurred in 'lgammacor'

For example, my model specification is:

fit_NB <- gllvm(
y = abundance_matrix_up,
X = env_matrix_up,
TR = trait_matrix_up,
family = "negative.binomial",
num.lv = 1,
formula = y ~ (FSF + TOM + DTT + DEPTH + YSC + OWF) +
(FSF + TOM + DTT + DEPTH + YSC + OWF):(TR),
offset = log(Sampling_size_up),
seed = 123,
control.start = list(n.init = 5, jitter.var = 0.1)
)

Could you please help be understand this warning and if it's appearing here due to how my data looks like? Based on AIC, models with a negative binomial distribution perform better than those with a zero-inflated negative binomial distribution. However, my abundance data contain a high proportion of zeros, so I am wondering whether a ZINB model might ultimately be more appropriate despite the AIC results

Thank you in advance for your help.
Christelle

[BertvanderVeen]

Thanks for your message, Christelle.

However, when fitting the negative binomial models (both with 1 and 2 latent variables), I consistently obtain the following warning:

simpleWarning in EvalADFunObject(ADFun, theta, order = order, hessiancols = cols, : underflow occurred in 'lgammacor'

You are free to ignore the warning. It can occur if the optimisation visits a part of the parameter space that leads to numerical issues, but it does not necessarily have any significant repercussions for the converged model object.

ZNIB is not necessarily suitable for a high proportion of zeros in the data, only if those zeros cannot be represented by any of the structural model components. That means to determine if zero-inflation needs to be added to the model, you first need to check if any residual assumptions are violated for the negative-binomial model.