The GxE kernel has a flat eigenvalue tail that behaves like an identity matrix
and competes with the residual. Dropping it reduces sigma2_GE inflation (mean
abs bias 0.333 -> 0.130) at essentially no cost in cross-validated prediction
(CV2 0.678 full vs 0.669 truncated).
Usage
truncate_gxe_kernel(random, keep_prop = 0.75, pattern = "GE")
Arguments
- random
a named list of kernels (the random argument of kernel_model).
- keep_prop
numeric in (0, 1]. Proportion of the GxE eigenvalue mass to retain.
Default 0.75.
- pattern
character. Regular expression selecting the GxE kernels to truncate.
Default "GE".
Value
The input random list of kernels, with the selected GxE
kernels replaced by their reduced-rank (truncated) version and any cached
spectral decomposition removed so the kernels are re-decomposed on the next
fit. When no kernel matches pattern, or keep_prop >= 1, the
list is returned unchanged.
Details
CAUTION: truncation distorts the G:GxE ratio – rg moved to 0.72 vs a true
0.595 in testing. Use only when sigma2_GE specifically matters.
Any cached decomposition is dropped, since the kernel changes.
