Samples the priors that lmm_animal places on the genetic and
residual covariance matrices and summarises them as heritabilities and
genetic correlations, which do not depend on the scale of the data.
Usage
animal_prior_draws(
n_traits,
n = 4000,
lkj_prior = 2,
trait_names = NULL,
beta_prior = NULL,
correlation_prior = c("lkj", "size", "shared_size")
)Arguments
- n_traits
Number of traits, a whole number of at least 1.
- n
Number of prior draws, a whole number of at least 1.
- lkj_prior
LKJ shape used in the fit, a finite number above 0. Shapes below 1 put mass on correlations near +/-1, and a shape of 0 or less is not a distribution.
- trait_names
Optional trait names.
- beta_prior
Optional shapes of the beta prior that the
parametrization = "h2"models put on the heritability of each trait. When given, the heritabilities come from that beta instead of the scaled half-normal ratio, because that is the prior the fit uses. The genetic correlations are drawn from the LKJ in either case.- correlation_prior
Prior on the genetic correlations, as in
lmm_animal:"lkj"(default), or"size"and"shared_size", which draw each correlation matrix as \(D(ll^T + R)D\) with \(R \sim\) LKJ(lkj_prior), size loadings \(l_k = \mu + \sigma z_k\), \(\mu^2/(1+\mu^2) \sim\) Beta(1, 1), \(\sigma \sim\) half-normal(0, 1) and \(z_k \sim N(0, 1)\). The two size priors differ only in howEshares the loading pattern, so they give the same draws here. Both are fitted only withparametrization = "h2", so the heritabilities come frombeta_prior, withc(1, 1)when it isNULL. They need at least two traits.
Value
A long data frame with columns type ("Heritability" or
"Genetic correlation"), parameter and value.
Details
Standard deviations are half-normal(0, 1) and correlation matrices
are LKJ(lkj_prior), as in the Stan models. Fits with
parametrization = "variance" use that prior unless beta_prior
is given, in which case that beta is the prior on each heritability.
The residual correlation structure of the fit does not enter these draws:
heritability divides the genetic variance by the phenotypic variance, which
only uses the diagonal of the residual covariance, and the genetic
correlations come from the correlation matrix of G. So
residual_structure makes no difference here, and there is no
argument for it.
