Bootstrap W with nonstationary covariate

Like bootstrap_W_detrended but uses an arbitrary covariate (e.g., SST) instead of a linear time trend, and retains the covariate sensitivity mu1 in the bootstrap parameter vector. NOAA sites are fitted with nonstationary GEV; ADCIRC sites use stationary GEV.

Usage

bootstrap_W_ns(dat, df, covariate, B = 500, ref_value = NULL, seed = NULL)

Arguments

dat

An evfuse_data object.

df

The raw data frame with covariate column.

covariate

String column name in df.

B

Number of bootstrap replications.

ref_value

Centering value (default: mean across NOAA site-years).

seed

Random seed.

Value

A list with W_bs (4L x 4L), Gamma (B x 4L), and n_failures.

Details

The bootstrap Gamma matrix has L * 4 columns in parameter-major ordering: (mu0, mu1, log_sigma, xi) at all sites. ADCIRC mu1 entries are NA; the resulting W_bs has NA in those positions. This is by design: embed_W skips unobserved entries.

Note

A small fraction of bootstrap resamples may produce degenerate data that causes warnings or convergence failures in extRemes::fevd. Failed fits are recorded in n_failures and excluded via pairwise-complete covariance estimation (failure rates are typically well below 1 percent).

See also

bootstrap_W_detrended for the time-trend version.