I have an experimental dataset with 160 conditions, each lasting 3.0 hours. During each condition, the subjects had to express their vote on a (single-item) rating scale with ordered categories. They voted after 1.0h, 1.5h, 2.0h, 2.5h, and 3.0h, for a total of 5 repeated measurements for each subject. In total, there were 800 males and 800 females. Only 10 subjects (5 males and 5 females) participated in each of the 160 conditions; no subject participated in more than one condition.
However, the subject ID is missing; therefore, knowing who voted in each category over time is impossible.
My idea to circumvent (to some extent) this problem was to model the aggregate data. So, for each measurement at time t model the number of votes in each ordered category. I am starting easy, and for the moment, I am not considering either the repeated measurements or the autocorrelation in the response variable (‘vote’).
The model is:
fit_logit <-
brm(data = data.sim,
family = cumulative(logit),
formula = vote | trials(n_vote) ~ 1 + x + female + (1 | condition),
iter = 4000, warmup = 1000,
chains = 4, cores = 4,
seed = 2024,
file = 'sim')
However, this model does not run because the argument ‘trials’ is not supported for family ‘cumulative(logit)’. The model works with family = multinomial(); however, doing so loses the information contained in the ordering. Is there a way to use the family = cumulative() to model aggregate data, considering that then should be extended to model repeated measurements?