Crossed or nested random effects structure or both for mixed effects model using glmer?

Question

I conducted an experiment where i walked transects down polytunnels on a farm. On each transect i was recording the number of flower visitation events that i saw. The farm is divided up into 3 fields, and I walked 3 transects down 3 seperate polytunnels in each field (so 9 transects in total). I then repeated these transect walks on a weekly basis for 6 weeks.

I using mixed effects models to analyse the data, with 'number of flower visits' as the response variable, and 'temperature' as a fixed effect. I am wondering how the random effects structure should look. I want to nest the transect within the field, but then each transect was repeatedly sampled on a weekly basis, so this aspect would be crossed. To represent this, i currently have the following random effects structure in R:

(1|Field/Transect) + (1|Week)

Is this correct for what i want to represent, as explained above?

Any help would be greatly appreciated.

Answer 1

This looks reasonable. (1|Field/Transect) + (1|Week) is equivalent to (1|Field) + (1|Field:Transect) + (1|Week). In principle given your experimental design you could fit the interaction of week with your experimental blocking factors (field and transect), i.e.

(1|Field/Transect) + (1|Week) + (1|Week:Field) + (1|Week:Field:Transect)

assuming you have one observation per week/field/transect combination, the last term in this would be an observation-level random effect (which would be unidentifiable in a model that includes a dispersion parameter, such as a linear (Gaussian) mixed model or a negative binomial model, but which can help account for overdispersion in a Poisson model). (In theory this model could be specified as (1|(Field/Transect)*Week), but I don't think lme4 is actually quite smart enough to handle that syntax ...)

You might well find that the more complex model is overspecified (i.e., some of your variances are estimated as exactly zero), in which case I would fall back to the slightly simpler model you've suggested. (In any case don't forget to check/account for overdispersion ...)