There are currently two strategies for the variational covariance matrix of latent variables:
- unstructured
- diagonal
Experience tells me we can "get away" with something more simple than 1) without resorting to 2). My thought is to collect all off-diagonal VA parameters into a single parameter, which might speed up estimation, especially for a large number of sites.
There are currently two strategies for the variational covariance matrix of latent variables:
Experience tells me we can "get away" with something more simple than 1) without resorting to 2). My thought is to collect all off-diagonal VA parameters into a single parameter, which might speed up estimation, especially for a large number of sites.