Pollinator abundance---sampled two years, fixed or random effect?

Question

I'm trying to construct several GLMMs with my thesis data and I can't solve this question. I have data on pollinator abundance and fruit set from plant individuals (with different treatments - related to a pest infection -) sampled at six study sites (spatial replicas) and during two consecutive years (time replicas). The problem is that not all individuals (plants) were sampled in both years (because of lack of bloom). My question is: How should i use "Year", as random effect or as fixed effect? I constructed three candidate models and compared the AIC:

a) Response variable ~Predictor variables + (1 | year/plant) + (1 | site/plant)

b) Response variable ~Predictor variables + (1 | year) + (1 | site/plant)

c) Response variable ~Predictor variables * year + (1 | site/plant)

Option a) had the lowest AIC and the highest R2, but I read in some papers that is not correct use a variable with less than five levels as random effect ... so, anyone could help me?

Answer 1

As year has only two distinct values, there is nothing to gain by representing it as a random effect, see What is the minimum recommended number of groups for a random effects factor? . That leaves you with option c).

About using aic to choose between these models, see Comparing between random effects structures in a linear mixed-effects model . There is doubt about wether aic should be used with mixed models with different random effects . One reason is that it is not obvious how to count degrees of freedom in such cases. I am not an expert on this issue, other users here seem to be, hopefully one can chime in!