Empirically choose the number of latent factors K

Applies the centred log-ratio (clr) transformation to the count matrix and performs PCA on the transformed data, then returns the smallest number of components whose eigenvalues explain at least prop of the total variance. This is the data-driven rule for setting K described in the paper.

Usage

choose_K(Y, prop = 0.95, pseudocount = 0.5, plot = FALSE)

Arguments

Y

An n x J count matrix.

prop

Proportion of variance to explain. Default 0.95.

pseudocount

Added before the log for the clr transform. Default 0.5.

plot

Logical; draw a scree plot. Default FALSE.

Value

An integer, the suggested K. The full eigenvalue vector is attached as attribute "eigenvalues".

Examples

sim <- simulate_bcaia(seed = 1)
choose_K(sim$Y)
#> [1] 6
#> attr(,"eigenvalues")
#>  [1] 2.874167e+01 5.916151e+00 3.625730e+00 2.408626e+00 2.110645e+00
#>  [6] 6.095488e-01 5.103859e-01 4.583081e-01 3.389422e-01 2.673896e-01
#> [11] 1.845620e-01 1.007298e-01 9.736230e-02 6.822095e-02 5.060356e-16