Skip to contents

Variable importance data object

Source: R/gg_boost_vimp.R
gg_boost_vimp.Rd

Extract variable importance from a vimp.boostmtree object, or from a BoostMLR predict object, for both the main effect of each covariate and its interaction with time.

Usage

gg_boost_vimp(object, components = c("main", "interaction"), ...)

Arguments

object

A vimp.boostmtree object, or a BoostMLR predict object made with importance = TRUE.

components

Character vector naming the components to extract, any of "main" and "interaction". Defaults to both.

...

Not used; present for S3 consistency.

Value

A gg_boost_vimp data.frame with columns:

variable

Factor covariate name.

importance

Numeric importance.

component

Factor, main or interaction.

response

Factor naming the response.

with attributes metric (the importance metric) and time.effect, plus interaction.by.interval for BoostMLR.

Details

boostmtree reports importance in two parts. The main effect asks what is lost by perturbing a covariate; the interaction asks what is lost by perturbing its interaction with time. For a longitudinal model the second is often the interesting one, since it is where a covariate whose effect changes over follow-up shows up.

boostmtree labels the interaction rows x1:time, x2:time and so on. The suffix is stripped here, so a variable carries one label across both components and can be compared or dodged against itself.

The importance metric is a string on the source object rather than a fixed quantity, so it travels as the metric attribute and the renderer uses it to label the axis. The overall time effect has no variable to attach to and travels as the time.effect attribute.

A joint importance object (vimp.boostmtree(fit, joint = TRUE)) reports a single combined value, labelled joint.vimp. Note that CRAN boostmtree 2.0.0 cannot produce one at all; the patched build this package requires is needed to compute it.

BoostMLR computes importance inside prediction: pass the object returned by BoostMLR::predictBoostMLR(fit, x, tm, id, y, importance = TRUE), whose $vimp holds one matrix per response. BoostMLR splits the covariate-time interaction across its overlapping time intervals (columns Int_Eff.1, Int_Eff.2, ...); these are summed here to one interaction value per covariate so the result has the same shape as for boostmtree. The per-interval matrices travel unsummed as the interaction.by.interval attribute, one per response. The metric is "Standardized VIMP", and no time.effect is recorded. A joint importance from BoostMLR::vimp.BoostMLR(joint = TRUE) is a bare list and is not accepted.

See also

plot.gg_boost_vimp, gg_boost_effect

Examples

# \donttest{
sim <- boostmtree::simLong(n = 25, n.time = 4, model = 1)$data.list
fit <- boostmtree::boostmtree(
  x = sim$features, tm = sim$time, id = sim$id, y = sim$y,
  M = 50, cv.flag = TRUE, verbose = FALSE
)
plot(gg_boost_vimp(boostmtree::vimp.boostmtree(fit)))

# }