I am try to see if my model is valid. I ran a beta glmm in brms.
brm.formula.1 <- bf(
prop ~ 0 + group +
# Group-specific independent random intercepts and slopes
(1 | gr(ID, by = group)) +
# OU correlation structure
gp(time,
by = ID,
k = 10, # optional number of basis functions for computing Hilbert-space approximate GPs
cov = "exponential", # Ornstein–Uhlenbeck
iso = TRUE, # (default)
gr = TRUE, # (default)
cmc = TRUE # (default)
),
phi ~ 0 + group + group:x1
)
# fit model
fit.brms.ou.1 <- brm(formula = brm.formula.1,
data = ddd,
#prior = prior,
family = Beta(link = "logit"),
chains = 4,
iter = 8000,
thin = 2,
warmup = 4000,
cores = 4,
control = list(
adapt_delta = 0.95,
max_treedepth = 14
),
backend = "cmdstanr"
)
I did this model with and without the ou process. then ran a loo_compare
I then check the density overlay
ppc_loo_overlay
ppc_loo_pit_overlay(
y = y_obs,
yrep = y_rep,
psis_object = loo.1$psis_object,
samples = 100
)
Which this shows that it fits into the other simulated unifrom, but there is a frowny face
I then included a spline
brm.formula.1 <- bf(
prop ~ 0 + group+
# Group-specific independent random intercepts and slopes
(1 | gr(ID, by = group)) +
gp(time,
by = ID,
k = 10, # optional number of basis functions for computing Hilbert-space approximate GPs
cov = "exponential", # Ornstein–Uhlenbeck
iso = TRUE, # (default)
gr = TRUE, # (default)
cmc = TRUE # (default)
),
phi ~ 0 + group + s(x1, by = group, k = 5)
)
and my ppc_loo_overlay looks better
So overall this seems good to me, but Im not entirely sure.
Thanks
