4

I’ve have been working with a mixed model (glmmTMB) to analyse the abundance of snails in dependency of several categorical predictors. The data was measured twice in the same sample sites in two different years (YEAR) and at 11 different farms (FARM).

So far, I have been using YEAR and FARM as random effects, but the number of two levels seems not sufficient for a random effect, as it has been discussed by several authors, e.g., http://bbolker.github.io/mixedmodels-misc/glmmFAQ.html#should-i-treat-factor-xxx-as-fixed-or-random.

I’ve come across the recommendation to use a low-level explanatory factor as a fixed effect instead, however, as I have done a repeated measurement in the same sample sites (YEAR), I think this would violate the assumption of independence of observations. I have been looking for alternative ways to handle this, but without success so far, so any advice or corrections would be highly appreciated.

kjetil b halvorsen
  • 77,844
Sonja
  • 41

1 Answers1

1

Based on your description, I would expect correlations on the farm level, i.e., measurements in the same farm will be correlated. And then on the site level, i.e., measurements within the same site will be correlated. Sites seem to be nested within farms. Hence, you could see if nested random effects are needed. More on the model-specification in lme4 and glmmTMB in these cases here: https://bbolker.github.io/mixedmodels-misc/glmmFAQ.html#nested-or-crossed

Are measurements in the same year but from different sites and farms expected to be correlated? If not, then you need to include year only as fixed-effects. Otherwise, it could be included as a random factor.

Dimitris Rizopoulos
  • 21,024
  • 2
  • 20
  • 47
  • Thanks for the prompt response and advice. I don’t think the model would profit from a nested design, as FARM does not seem to be a strong grouping factor. I considered dropping it the model altogether.

    Observations from the same farm are very variable, as well as observations from the same year. Intraclass correlation coefficients for these groups are also very low. This could justify using YEAR as a fixed effect if I understand correctly?

     – Sonja Mar 24 '21 at 09:47