random_spec {dynamite}

Additional Specifications for the Group-level Random Effects of the DMPM

Description

This function can be used as part of dynamiteformula() to define whether the group-level random effects should be modeled as correlated or not.

Usage

random_spec(correlated = TRUE, noncentered = TRUE)

Arguments

correlated

[logical(1)]
If TRUE (the default), correlations of random effects are modeled as multivariate normal.

noncentered

[logical(1)]
If TRUE (the default), use a noncentered parameterization for random effects. Try changing this if you encounter divergences or other problems in sampling.

Details

With a large number of time points random intercepts can become challenging sample with default priors. This is because with large group sizes the group-level intercepts tend to be behave similarly to fixed group-factor variable so the model becomes overparameterized given these and the common intercept term. Another potential cause for sampling problems is relatively large variation in the intercepts (large sigma_nu) compared to the sampling variation (sigma) in the Gaussian case.

Value

An object of class random_spec.

See Also

Model formula construction dynamiteformula(), dynamite(), lags(), lfactor(), splines()

Examples

data.table::setDTthreads(1) # For CRAN
# two channel model with correlated random effects for responses x and y
obs(y ~ 1 + random(~1), family = "gaussian") +
  obs(x ~ 1 + random(~1 + z), family = "poisson") +
  random_spec(correlated = TRUE)