Fits a univariate or multivariate linear animal model with fixed effects and relatedness
matrix using a single-chain Gibbs sampler. This is intended for quick
simulations and checks; its priors differ from those in lmm_animal.
Usage
lmm_animal_gibbs(Y, X = NULL, A, iter = 1000, warmup = floor(iter/2), thin = 1)Arguments
- Y
Numeric response vector or matrix, with individuals in rows. Every trait must vary, and is checked as in
lmm_animal.- X
Optional numeric full-column-rank fixed-effects design matrix, with the same rows as
Y. The defaultNULLfits an intercept-only model: one column of ones, named"(Intercept)". Unlike the Stan models, it needs no column representing a constant: the sampler does not shift the response mean through the design.- A
Symmetric positive-semidefinite relationship matrix, with rows and columns in the same individual order as the rows of
Y. When bothAandYhave row names they are used to putAin the order ofY; names that disagree are an error.- iter
Total number of Gibbs iterations, a whole number.
- warmup
Number of initial iterations to discard, a whole number smaller than
iter.- thin
Keep every
thinth iteration after burn-in, a whole number;warmup + thinmust not exceediter.
Value
A list of posterior draws: G and R are covariance
arrays, beta contains fixed-effect matrix draws, h2
contains trait-wise variance ratios, and iterations gives the
retained iteration numbers. Quantities are indexed by trait first and draw
last, from named dimensions, the layout that
back_transform returns as well. fit_draws
accepts this list and returns the same layout with R named
E, plus the quantities derived from it.
Details
Fixed effects have an improper flat prior. Genetic and residual
covariance matrices have inverse-Wishart priors with degrees of freedom
K + 1 and scale matrix (K + 1) I / 2, where K is
the number of traits. These differ from the priors used by
lmm_animal; the resulting posterior draws are not directly
interchangeable.
