Extracts the draws of an animal model and reshapes every parameter so that
its first dimensions are the traits. Covariance matrices come out as trait
by trait by draw arrays with named dimensions, ready to average, take
correlations of, or compare with a simulated truth, instead of one Stan
variable per element such as G[1,1].
Arguments
- fit
A fit from
lmm_animal,lm_phen_covorlmm_animal_gibbs, aposteriordraws object, or a named matrix of draws such asfit$draws(format = "draws_matrix"). Objects fromback_transformalso work.- traits
Optional trait names, one per trait. The default uses the column names of the response when the fit kept them, and
trait1,trait2, ... otherwise.- terms
Optional names for the columns of
Xin thebetadraws, in their order in the fit. The default uses the column names ofXwhen the fit kept them, and the names the fit reports otherwise.
Value
A list of class fit_draws, with one element per quantity
the fit holds: G, E, P, corrG, corrE,
corrP as trait by trait by draw arrays, h2, sigma_G,
sigma_E, sigma_P, uniqueness, communality,
genetic_uniqueness, residual_uniqueness, s_G and,
for correlation_prior = "size" or "shared_size" fits,
size_loadings_G and size_loadings_E as trait by draw
matrices, loadings as a trait by factor by draw array
for factor fits, factor_heritability and factor_magnitude
as factor by draw matrices, and beta as a trait by term by draw
array. Every dimension
is named, and draws are last, the same layout as
lmm_animal_gibbs and back_transform.
Quantities the fit does not report are derived when they follow from the
others: P = G + E, corrG, corrE and corrP from
the covariances, sigma_G, sigma_E and sigma_P from
their diagonals, and h2 from G and P.
Details
The Gibbs sampler calls its residual covariance R; it is
returned as E so that the two engines give the same names. The
priors differ between the engines, so the draws themselves are not
interchangeable.
loadings comes back **sign-aligned**: a column of \(\Lambda\) is
identified only up to its sign, so a chain may visit both orientations, and
averaging the raw draws then cancels the factor out of the posterior mean and
out of any heatmap built from it. Each draw is flipped so that the trait with
the largest mean absolute loading on that factor is positive, which is a
property of the whole posterior rather than of one draw. Nothing else is
changed: communality, uniqueness, factor_magnitude and
the covariances are invariant to the sign anyway. A permutation of the
columns or a genuine rotation between draws is *not* repaired, which is why
those are read on G and communality instead. For the draws
exactly as the sampler wrote them, use fit$draws("loadings").
Examples
k <- 2
vars <- sprintf("G[%d,%d]", rep(1:k, k), rep(1:k, each = k))
m <- cbind(matrix(stats::rnorm(40), 10, 4, dimnames = list(NULL, vars)),
matrix(stats::rnorm(40), 10, 4,
dimnames = list(NULL, sub("G", "E", vars))))
d <- fit_draws(m, traits = c("t1", "t2"))
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: diag(V) had non-positive or NA entries; the non-finite result may be dubious
#> Warning: NaNs produced
#> Warning: NaNs produced
#> Warning: NaNs produced
names(d)
#> [1] "G" "E" "P" "corrG" "corrE" "corrP" "h2"
#> [8] "sigma_G" "sigma_E" "sigma_P"
dim(d$G) # traits x traits x draws
#> [1] 2 2 10
round(draws_mean(d, "G"), 2) # posterior mean G, as a matrix
#> t1 t2
#> t1 0.07 0.27
#> t2 0.24 0.12
round(draws_mean(d, "corrG"), 2)
#> t1 t2
#> t1 1 NaN
#> t2 NaN 1
