What you’ll learn
- How to launch
sb_gamlss()with correlated-selectBoost bootstraps for μ and σ. - How to interpret
selection_table(),plot_sb_gamlss(), andeffect_plot()outputs. - How to mix engines across parameters (
engine,engine_sigma,engine_tau) and keep factor/spline terms grouped. - How to automate c0 exploration with
sb_gamlss_c0_grid(),autoboost_gamlss(), andfastboost_gamlss().
Conference highlight. SelectBoost for GAMLSS was presented at the Joint Statistics Meetings 2024 (Portland, OR) in the talk “An Improvement for Variable Selection for Generalized Additive Models for Location, Shape and Scale and Quantile Regression” by Frédéric Bertrand and M. Maumy. The presentation stressed how correlation-aware resampling sharpens variable selection for GAMLSS and quantile regression problems even when predictors are numerous and strongly correlated.
set.seed(1)
library(gamlss)
library(SelectBoost.gamlss)
n <- 400
x1 <- rnorm(n); x2 <- rnorm(n); x3 <- rnorm(n)
mu <- 1 + 1.5*x1 - 1.2*x3
y <- gamlss.dist::rNO(n, mu = mu, sigma = 1)
dat <- data.frame(y, x1, x2, x3)
fit <- sb_gamlss(
y ~ 1,
mu_scope = ~ x1 + x2 + x3,
sigma_scope = ~ x1 + x2,
family = gamlss.dist::NO(),
data = dat,
B = 50,
sample_fraction = 0.7,
pi_thr = 0.6,
k = 2,
pre_standardize = TRUE,
trace = FALSE
)
fit$final_formula
#> $mu
#> y ~ x1 + x2 + x3
#>
#> $sigma
#> ~1
#> <environment: 0x1377c04a8>
#>
#> $nu
#> ~1
#> <environment: 0x1377c04a8>
#>
#> $tau
#> ~1
#> <environment: 0x1377c04a8>
sel <- selection_table(fit)
sel <- sel[order(sel$parameter, -sel$prop), , drop = FALSE]
head(sel, n = min(8L, nrow(sel)))
#> parameter term count prop
#> 1 mu x1 50 1.00
#> 3 mu x3 50 1.00
#> 2 mu x2 37 0.74
#> 4 sigma x1 10 0.20
#> 5 sigma x2 1 0.02
stable <- sel[sel$prop >= fit$pi_thr, , drop = FALSE]
if (nrow(stable)) {
stable
} else {
cat("No terms reached the stability threshold of", fit$pi_thr, "for this run.\n")
}
#> parameter term count prop
#> 1 mu x1 50 1.00
#> 3 mu x3 50 1.00
#> 2 mu x2 37 0.74
plot_sb_gamlss(fit, top = 10)
if (requireNamespace("ggplot2", quietly = TRUE)) {
print(effect_plot(fit, "x1", dat, what = "mu"))
}
selection_table() ranks stable terms for each parameter;
the chunk above prints the eight strongest effects (sorted by stability)
alongside the terms exceeding the stability threshold.
plot_sb_gamlss() overlays selection frequency
vs. coefficient magnitude, and effect_plot() visualizes
partial effects (using ggplot2 if available, otherwise
base R).
