sobolnp 0.2.0

Bug fixes

• Request standard errors explicitly in npreg() (se = TRUE). np 0.80-1 made them opt-in, which broke the residual bootstrap.
• The residual bootstrap now resamples the residuals of each variable separately. Previously all variables were bootstrapped with the residuals of the first variable, and sobolnp() failed with two input variables. Bootstrap estimates (Sboot, bwsboot) differ from version 0.1.0.
• A bandwidth vector of length ncol(X) is now used one value per variable instead of failing.
• The package no longer masks base::print() and graphics::plot().

Changes

• Improved bootstrap bandwidth selection. The bootstrap responses are built from residuals standardized with a kernel estimate of their scale. For each candidate bandwidth, the bootstrap curves give a bias-corrected curve (twice the CV curve minus the mean bootstrap curve). The bandwidth that keeps the bootstrap indices closest to the CV index is searched between bws / 1000 and bws. In simulations with the default second-order kernel, Sboot reduces the downward bias of the CV estimate. Results differ from version 0.1.0, and the minqa dependency is no longer needed.
• sobolnp() always returns a sobolnp object, also with bootstrap = FALSE (then Sboot and bwsboot are NULL).
• sobolnp() validates its inputs, accepts data frames and names unnamed variables X1, ..., Xp. Negative bandwidths raise an error instead of returning NA.
• New argument verbose to silence progress messages and bars.
• plot() draws a grouped bar plot, accepts ylim, xlab, ylab, col and legend.pos, and returns its input invisibly.
• print() gains a digits argument and returns its input invisibly.

sobolnp 0.1.0

Initial version based in the paper

Solís, Maikol. “Nonparametric estimation of the first order Sobol indices with bootstrap bandwidth.” (2018).