CODEWITHSUNDEEP
rmatrixglmmtmbglmmtmb
yhentati
yhentati

Reputation: 1

Error in glmmTMB after Matrix update: negative log-likelihood is NaN at starting parameter values

I'm modeling camera trap data (counts) with camera site as a random effect and landscape-level covariates + season as fixed effects, and an offset for number of days the camera station was active, as below:

mod2 <- glmmTMB(Coyote ~ develop + PERCENTILE + season + (1|site), 
                offset = log(days.active),
                data = all_dat, family = "nbinom2")

and am, as of today, encountering the following error message:

Error in fitTMB(TMBStruc) : negative log-likelihood is NaN at starting parameter values

I have about 400 observations in my data. I only get this error when the offset is included, and it happens with both a negative binomial and Poisson model. My code was working just fine until it suddenly didn't, so I assume the issue was related to the recent Matrix changes and associated dependencies in TMB, as detailed in similar questions here with the same error in glmmTMB and here with a similar issue in lme4. However, none of the solutions in the links have worked for me (I tried commenting instead of opening a new question but wasn't able to as my account is new). I'm on macOS 12.5.

I have tried:

  1. Reinstall Matrix to version 1.6-2 as suggested by an answer in the first link using install_versions from the remotes package
  2. Reinstall Matrix from source (version 1.6-3) and also reinstall TMB and glmmTMB as suggested by the Matrix dev in the second link -- I tried this both ways suggested in the answer, by installing TMB/glmmTMB from the source and then trying to install a compatible binary from CRAN
  3. Reinstall Matrix from the source (version 1.6-3) and only reinstall glmmTMB from source, as installing TMB from source resulted in version discrepancies between glmmTMB and TMB
  4. Updated R and tried all of the above again

Both reinstallation methods didn't work at all for me until I installed gfortran, but now they are at least installing -- just not solving my issue.

Thanks in advance for any help!

Upvotes: 0

Views: 1531

Answers (1)

Ben Bolker
Ben Bolker

Reputation: 226742

This is just a guess, but I can easily get this error if I have a 0 in the variable that I am logging in order to get the offset (which would be any(all_dat$days.active == 0) in your case):

library(glmmTMB)
data("sleepstudy", package = "lme4")
ss <- transform(sleepstudy, offset = c(0, rep(1, nrow(sleepstudy)-1)))
m1 <- glmmTMB(round(Reaction)~ Days, family = nbinom2, data = ss)  ## works fine
m2 <- update(m1, offset = log(offset))

Error in fitTMB(TMBStruc) : negative log-likelihood is NaN at starting parameter values

(However, this is not consistent with "it was working before but then it stopped working", as this error should have happened with previous versions of the package too ...)

Upvotes: 0

Related Questions