Error when trying to fit glmer in R with Poisson distribution: PIRLS step-halvings failed to reduce deviance in pwrssUpdate

Question

The models I'm running won't converge / show an error. Do you have any advice about working with non-normal focal data? Most of the data are zero inflated or highly skewed.

The model I want to run is this (the response is just an example):

model <- glmer(Duration_behavior_1 ~ 
                 status * days_since_event +
                 focal_age + 
                 focal_rank +
                 pink +
                 year + 
                 hour +
                 (1|FocalID) + 
                 (1|partner.ID),
               family = 'poisson',
               offset = Focal_duration,
               data = data)

But I get this error:

Error in (function (fr, X, reTrms, family, nAGQ = 1L, verbose = 0L, maxit = 100L,  :
  (maxstephalfit) PIRLS step-halvings failed to reduce deviance in pwrssUpdate

I think the problem is the offset. If I remove it, I get this warning instead:

Warning messages:
1: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  Model failed to converge with max|grad| = 0.00252314 (tol = 0.002, component 1)
2: In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
  Model is nearly unidentifiable: very large eigenvalue
 - Rescale variables?;Model is nearly unidentifiable: large eigenvalue ratio
 - Rescale variables?

But, of course, the offset is necessary to control for the duration of the observation...

Do you think it would be recommended to change the distribution to a negative binomial, instead of using a Poisson?

Answer 1

You almost certainly want to use log(Focal_duration) as your offset, rather than Focal_duration.

Suppose eta is the linear predictor not including the offset, e.g. beta0 + beta1*x1 + beta2*x2 + ... (also including the components based on the random effect). Then the predicted value is eta + offset, and the expected mean is exp(eta)*exp(offset). If the values of the offset (Focal_duration) are reasonably large (say, 60 seconds?), then the values of exp(offset) are huge (e.g. exp(60) ~ 1.1e26), which [besides being not what you want] messes up numerical computations.

In fact, the GLMM FAQ says something very similar:

a common error is to specify an offset to a log-link model as a raw searching-effort value, i.e. offset(effort) rather than offset(log(effort)). While the intention is to fit a model where counts ∝ effort, specifying offset(effort) leads to a model where counts ∝ exp(effort) instead; exp(effort) is often a huge (and model-destabilizing) number.