Performs leave-one-out cross-validation (LOOCV) to empirically
select the optimal ridge penalty (lambda) for a given count matrix. Models
are fit on $N-1$ samples using a specified lambda, and unpenalized
likelihood is evaluated on the held-out sample.
Usage
select_optimal_lambda(count_matrix, lambdas = c(0.1, 0.5, 1, 5, 10, 50, 100))
Arguments
- count_matrix
A numeric matrix of mutation counts with subtypes
as rows and samples as columns.
- lambdas
A numeric vector of lambda penalty values to evaluate.
Default is c(0.1, 0.5, 1, 5, 10, 50, 100).
Value
A list containing two items:
best_lambda: Numeric scalar. The lambda value from the
provided vector that yielded the highest overall cross-validation
log-likelihood.
scores: A named numeric vector detailing the cumulative
log-likelihood score achieved by each tested lambda.
