# mixed effect binomial model over time

I have a data set of seed germination (germinated / failure) and I'm trying to understand whether germination rates of treatments differed from control. We made a randomised blocked experiment in which in each experimental unit (Plots) we had five seeds. We have two time points where germination data was measured (T1 and T2). Since we observed that several seeds germinated in the second time point, it is of our interest to include "time" as a variable and show this pattern. We are modeling the data through glmm in which we have successes and failures for each sampling unit:

cbind(successes,failures) ~ treatment * time + (1 | Blocks), family = binomial, link = logit


The question lies within the germination data. Should I keep what was computed in T1 or only account what was recorded for T2?

The successes of T2 include the successes of T1, therefore T2 will never have less successes than T1 (basically, T1 is within T2). For example, Pot 1 had 2 out of 5 germinated seeds in T1 (40%), and in T2 it had 4 out of 5 (80%). But two of them were already in T1, so actually T2 had 2 germinated seeds out of 3 (66.6%).

Hope you'll understand the question.

• I think it's fine to remove time from the model and count seeds as having 'succeeded' if they germinated by T2. So your model would be (germination ~ treatment + (1 | Blocks), family = binomial, link = logit) – mkt - Reinstate Monica Nov 27 '17 at 15:59

Unless you are specifically interested in the time-dependence of germination, I think it's fine to remove time from the model and count seeds as having 'succeeded' if they germinated by T2. So your model would be
glmer(germination ~ treatment + (1 | Blocks), family = binomial)

This has the added advantage of simplifying your model. glmer can be tricky if you are not familiar with GLMMs; I would recommend reading a book or consulting a statistician if the fitting is challenging or the output difficult to interpret.