I'm running a glmm of pathogens in edible crabs with a pathogen as my dependent variable and I have sex, width, external disease, season, damage and biofouling as explanatory variables and site as a random effect. The code for the model looks like this:
glmmTMB(Paramikrocytos ~ Width + Sex + Condition + Disease + Biofouling + (1 | Site),
data = crab, family = "binomial")
When I run this I get the following error message:
Error in fitTMB(TMBStruc) :
negative log-likelihood is NaN at starting parameter values
If I try using a Poisson family instead of a binomial like the troubleshooter suggests I get the same error message and the output looks like this:
Conditional model:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -3.182e+00 NaN NaN NaN
Width 5.834e-03 NaN NaN NaN
SexM -2.001e-01 NaN NaN NaN
Condition2 -4.017e-01 NaN NaN NaN
Condition3 -3.297e+01 NaN NaN NaN
Condition4 -6.188e-01 NaN NaN NaN
Condition5 -1.037e+02 NaN NaN NaN
Condition7 -2.773e-01 NaN NaN NaN
Disease2 4.384e-01 NaN NaN NaN
Disease3 2.375e-01 NaN NaN NaN
Disease4 2.486e+00 NaN NaN NaN
Disease5 5.011e+01 NaN NaN NaN
Disease6 -1.242e+01 NaN NaN NaN
Biofouling2 2.794e-01 NaN NaN NaN
Biofouling3 -4.770e+01 NaN NaN NaN
Biofouling4 -8.677e+01 NaN NaN NaN
I tried using a Poisson family as was suggested when I ran the function "diagnose" on the model but I got the same error message. Is there a way of running a negative binomial in this? What could possibly be causing this error message?
I've solved it! I needed to set the priors using
cprior <- data.frame(prior = rep("normal(0,3)", 2),
class = rep("beta", 2),
coef = c("(Intercept)", ""))
If I add priors = cprior into the model then it works!
If anyone else is having this issue, try this and see if it helps.