Package 'ggBoostedTrees'

Plot a boostmtree fit

Description

A shortcut to the error trajectory, which is the first thing worth looking at after a fit. Every other figure is reached through its own ⁠gg_boost_*()⁠ extractor.

Usage

## S3 method for class 'boostmtree'
autoplot(object, ...)

Arguments

object	A fitted boostmtree object grown with cv.flag = TRUE.
...	Passed to autoplot.gg_boost_error.

Value

A ggplot object, as produced by autoplot.gg_boost_error.

See Also

gg_boost_error, gg_boost_path, gg_boost_trajectory

Examples

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
)
ggplot2::autoplot(fit)

---

Plot a gg_boost_calibration object

Description

Observed against fitted over follow-up: bin means of the observed values with their intervals, the fitted bin means beside them, and the cohort mean curve when the object carries one.

Usage

## S3 method for class 'gg_boost_calibration'
autoplot(object, ci = TRUE, ...)

## S3 method for class 'gg_boost_calibration'
plot(x, ci = TRUE, ...)

Arguments

object	A gg_boost_calibration object.
ci	Logical. Draw 95\ to TRUE; bins holding one observation have none.
...	Not used; present for S3 consistency.
x	A gg_boost_calibration object.

Details

Each bin is drawn as a thin horizontal segment spanning its time range at the observed mean. With equal-count bins that segment is the tail-support signal: where observations are sparse, a bin has to stretch further to collect its share, so a long segment late in follow-up means a thin tail.

Observed bin means are points with vertical 95\ fitted bin means are a second point series in another shape, joined by a dashed line. Where the two agree bin by bin, the model is calibrated over that stretch of follow-up. The cohort curve, drawn as a heavier line, is the mean predicted curve across all subjects; it can depart from the bin means late in follow-up when the subjects still being measured differ from the cohort, which is a property of follow-up rather than of the model.

The returned plot carries no theme.

Value

A ggplot object.

See Also

gg_boost_calibration

Examples

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$y,
  M = 50, verbose = FALSE
)
plot(gg_boost_calibration(fit, pred = predict(fit, x = fit$x)))

---

Plot a gg_boost_effect object

Description

Covariate effect curves, one line per time point, faceted by variable and, for a multi-response fit, by response.

Usage

## S3 method for class 'gg_boost_effect'
autoplot(object, ...)

## S3 method for class 'gg_boost_effect'
plot(x, ...)

Arguments

object	A gg_boost_effect object.
...	Passed to geom_line (continuous covariates) or geom_point (discrete covariates).
x	A gg_boost_effect object.

Details

Time is mapped to colour rather than to a facet because the question this figure answers is how a covariate's effect changes across follow-up, and that change is legible only when the curves share one panel. Variables get the facets instead, since their x scales are unrelated.

A continuous covariate is drawn as a line per time point across the covariate grid. A discrete (factor) covariate has one row per level, so a "curve" would connect points along an axis with no ordering; it is drawn instead as a point per level per time point, on an axis labelled with the level names. Because ggplot2 permits only one scale type per aesthetic across all facets, an object mixing continuous and discrete covariates cannot be drawn as a single figure and is refused with an error asking the caller to plot them one at a time.

The returned plot carries no theme.

Value

A ggplot object.

See Also

gg_boost_effect

Examples

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, verbose = FALSE
)
pp <- boostmtree::partial.plot(
  fit, x.var.names = "x1", output = "data", verbose = FALSE
)
plot(gg_boost_effect(pp))

---

Plot a gg_boost_error object

Description

The error path of a boosted multivariate tree fit against the boosting iteration, with the cross-validated optimal iteration marked.

Usage

## S3 method for class 'gg_boost_error'
autoplot(object, optimal = TRUE, ...)

## S3 method for class 'gg_boost_error'
plot(x, optimal = TRUE, ...)

Arguments

object	A gg_boost_error object.
optimal	Logical. Draw a vertical rule at the optimal iteration. Defaults to TRUE.
...	Passed to geom_line.
x	A gg_boost_error object.

Details

The returned plot carries no theme, so it composes with any ggplot2 theme or scale. Several responses are drawn as facets rather than as coloured series, because the responses do not share a y scale.

Value

A ggplot object.

See Also

gg_boost_error

Examples

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_error(fit))

---

Plot a gg_boost_path object

Description

The estimated correlation, variance, and smoothing parameter paths of a boosted multivariate tree fit against the boosting iteration.

Usage

## S3 method for class 'gg_boost_path'
autoplot(object, ...)

## S3 method for class 'gg_boost_path'
plot(x, ...)

Arguments

object	A gg_boost_path object.
...	Passed to geom_line.
x	A gg_boost_path object.

Details

The three parameters live on entirely different scales, so they are always faceted with a free y axis rather than drawn together. With several responses the facet grid is parameter by response.

The returned plot carries no theme.

Value

A ggplot object.

See Also

gg_boost_path

Examples

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_path(fit))

---

Plot a gg_boost_trajectory object

Description

Observed and fitted subject trajectories over time: one line per subject through the fitted values, with the observed values as points.

Usage

## S3 method for class 'gg_boost_trajectory'
autoplot(
  object,
  subset = NULL,
  n_max = 100,
  observed = TRUE,
  alpha = NULL,
  ...
)

## S3 method for class 'gg_boost_trajectory'
plot(x, subset = NULL, n_max = 100, observed = TRUE, alpha = NULL, ...)

Arguments

object	A gg_boost_trajectory object.
subset	Character or numeric vector of subject identifiers to keep. NULL (default) keeps all of them.
n_max	Maximum number of subjects to draw. Defaults to 100; Inf draws every subject.
observed	Logical. Draw the observed values as points. Defaults to TRUE, and is ignored when the fit records no observed values.
alpha	Numeric transparency for both layers. NULL (default) computes one from the number of subjects drawn.
...	Passed to geom_line.
x	A gg_boost_trajectory object.

Details

The extractor returns every subject, so thinning happens here and the tidy data frame always represents the whole fit. subset names the subjects to keep; n_max caps how many are drawn, sampling at random and saying so. Set n_max = Inf to draw all of them. Random sampling is not seeded – call set.seed() first for a reproducible figure.

Transparency is doing real work in this figure rather than decorating it. A cohort of any size overplots, and partial transparency turns the tangle into something you can read densities off: where many subjects follow the same path the ink accumulates. The default alpha therefore falls as the number of subjects drawn rises. Pass alpha explicitly to override.

The returned plot carries no theme.

Value

A ggplot object.

See Also

gg_boost_trajectory

Examples

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, verbose = FALSE
)
plot(gg_boost_trajectory(fit))

---

Plot a gg_boost_vimp object

Description

Variable importance as a horizontal bar chart, ordered by importance and faceted by component.

Usage

## S3 method for class 'gg_boost_vimp'
autoplot(object, ...)

## S3 method for class 'gg_boost_vimp'
plot(x, ...)

Arguments

object	A gg_boost_vimp object.
...	Passed to geom_col.
x	A gg_boost_vimp object.

Details

Bars run horizontally because variable names are text and read better along the y axis than rotated beneath it, and variables are ordered by importance because an importance chart in arbitrary order is close to unreadable.

The x axis is labelled with the metric attribute carried by the extracted object. That metric is a property of how importance was computed rather than a fixed quantity, so a hard-coded label would misdescribe some objects.

A negative importance is not an error: with a permutation-style metric a variable that contributes nothing can score slightly below zero by chance.

The returned plot carries no theme.

Value

A ggplot object.

See Also

gg_boost_vimp

Examples

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)))

---

Partial effects from a BoostMLR fit

Description

Compute partial effect surfaces for one or more covariates of a BoostMLR fit, in a form gg_boost_effect accepts.

Usage

boostmlr_partial(object, xvar.names = NULL, ...)

Arguments

object	A BoostMLR grow object, as returned by BoostMLR::BoostMLR().
xvar.names	Character vector of covariate names. Defaults to every covariate in the fit.
...	Passed to BoostMLR::partial.BoostMLR(), for example n.x, n.tm or Mopt.

Details

BoostMLR::partial.BoostMLR() takes one covariate per call and returns a bare list with neither the covariate's name nor the responses' names, so it cannot be told apart from any other list. This wrapper calls it once per covariate with plotting switched off, and records the covariate names and the fit's response names (y_Names) alongside the results. It is the BoostMLR counterpart of boostmtree::partial.plot(output = "data").

Each result holds the raw partial effect (pList) and a lowess smooth of it over time (sList), as covariate-by-time matrices, one per response. gg_boost_effect extracts the smooth.

Value

A boostmlr_partial object: a list with elements curves (one partial.BoostMLR() result per covariate, named by covariate), x.var.names and response.labels.

See Also

gg_boost_effect

Examples

if (requireNamespace("BoostMLR", quietly = TRUE)) {
  sim <- BoostMLR::simLong(
    n = 20, N = 3, rho = 0.8, model = 1, q_x = 2, q_y = 0
  )$dtaL
  fit <- BoostMLR::BoostMLR(
    x = sim$features, tm = sim$time, id = sim$id, y = sim$y,
    M = 50, Verbose = FALSE
  )
  plot(gg_boost_effect(boostmlr_partial(fit, "x1")))
}

---

Calibration over follow-up data object

Description

Compare observed and fitted values over follow-up time, binned by time, with an optional cohort mean curve.

Usage

gg_boost_calibration(object, pred = NULL, n_bins = 10, breaks = NULL, ...)

Arguments

object	A fitted boostmtree object, or a fitted BoostMLR object. Not a predict object: pass that as pred.
pred	Optional boostmtree predict object from predict(fit, x = ...), used for the cohort curve.
n_bins	Number of equal-count time bins. Defaults to 10.
breaks	Optional numeric vector of bin edges in the fit's time units. Overrides n_bins.
...	Not used; present for S3 consistency.

Details

A cohort's predicted curves say what the model expects; they cannot say whether it matches the data. This figure answers that. Observations are binned by time, and each bin reports the mean observed value, the mean fitted value at the same observations, and how many observations and distinct subjects stand behind it.

Bins hold roughly equal numbers of observations by default: the edges are quantiles of observed time, computed per response. Tied times, common in clinical data where many measurements share a visit time, merge quantile edges; duplicates are removed and a message says how many bins survived. Supply breaks, in the fit's own time units, to choose the edges yourself. Observations outside the range of breaks are dropped, with a message giving the count.

With equal-count bins, n_obs is nearly constant by construction, so it cannot show where data thins out. Two columns can: bin width (time_hi - time_lo), which grows where observations are sparse, and n_subject, which falls when a late bin rests on repeat measurements of a few subjects.

The interval on observed is a normal approximation that treats observations as independent. Repeated measurements within a subject make it too narrow. Read it as a visual guide, not an inferential claim.

Supplying pred adds the cohort curve: the mean over subjects of the predicted curves on a common time grid, attached as the cohort attribute. Make pred with predict(fit, x = ...) and no tm or id; with them, each subject is predicted at its own times and a mean across subjects changes composition from one time to the next. The cohort curve is available for boostmtree fits only.

This figure accepts either a boostmtree or a BoostMLR grow fit. It reads its rows from gg_boost_trajectory, so both backends' storage layouts are handled there.

Value

A gg_boost_calibration data.frame with one row per response and bin, ordered by response then time, with columns:

response

Factor naming the response.

bin

Factor bin label, ordered by time within each response.

time_lo, time_hi

Numeric bin edges.

time

Numeric median observed time within the bin.

n_obs

Integer count of observations in the bin.

n_subject

Integer count of distinct subjects in the bin.

observed

Numeric mean observed value.

observed_lo, observed_hi

Numeric 95\ NA when the bin holds one observation.

fitted

Numeric mean fitted value at the same observations.

When pred is supplied, the attribute cohort holds a data.frame with columns response (factor), time and fitted (numeric).

See Also

plot.gg_boost_calibration, gg_boost_trajectory

Examples

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$y,
  M = 50, verbose = FALSE
)
pred <- predict(fit, x = fit$x)
gg <- gg_boost_calibration(fit, pred = pred)
plot(gg)

# The predicted curves themselves, for a look at who drives the tail:
plot(gg_boost_trajectory(pred), n_max = 10)

---

Partial and marginal effect data object

Description

Extract covariate effect curves over time from a partial.plot or marginal.plot object, or from the boostmlr_partial result for a BoostMLR fit.

Usage

gg_boost_effect(object, ...)

Arguments

object	A partial.plot.boostmtree or marginal.plot.boostmtree object, as returned by boostmtree::partial.plot() or boostmtree::marginal.plot() with ⁠output = "data", verbose = FALSE⁠, or a boostmlr_partial object.
...	Not used; present for S3 consistency.

Details

The two differ in what they hold constant. A partial effect varies one covariate while averaging over the others; a marginal effect reads the fitted surface as the data actually distribute it. Both are covariate-by-time surfaces, so both land in this one class, distinguished by kind.

marginal.plot() returns a raw scatter alongside its smoothed curve. That scatter is not the raw observations: it holds one unsmoothed fitted prediction per subject at the subject's own observed covariate value, not the observed response and not the stored fitted values. It is also not reconstructable from gg_boost_trajectory, which carries no covariate column. gg_boost_effect extracts the smoothed curve instead, so that both levels of kind mean the same thing: the fitted effect.

For BoostMLR, boostmlr_partial returns both the raw partial effect and a lowess smooth of it over time; the smooth is extracted, matching the marginal choice above. The result has kind partial and one block per response, labelled from the fit's y_Names.

No source computes a confidence interval, so none is reported here.

boostmtree accepts factor covariates. For those, partial.plot() and marginal.plot() return a character (or, for marginal.plot()$data, factor) x column with one row per level rather than a numeric grid. gg_boost_effect detects this and maps each level to an integer position in x, carrying the level itself in x_label; a continuous covariate keeps its numeric value in x and leaves x_label NA. The grid is resolved once per variable so every time point shares the same level ordering.

A multi-response fit (family = "ordinal" or "nominal") nests ⁠$curves⁠ / ⁠$smooth⁠ as ⁠[[response]][[variable]]⁠; a single-response fit flattens the outer level. gg_boost_effect accepts both and records the response in a response column, labelled from the object's ⁠$response.labels⁠ ("response" for a single-response fit). Each response of an ordinal or nominal fit is one model component, so its curves are on that component's scale.

Value

A gg_boost_effect data.frame with columns:

variable

Factor covariate name.

x

Numeric covariate value. For a continuous covariate this is the covariate itself; for a discrete (factor) covariate this is an integer position, one per level.

x_label

Character level label for a discrete covariate, NA for a continuous one.

time

Numeric time point.

estimate

Numeric fitted effect.

kind

Factor, partial or marginal.

response

Factor naming the response.

See Also

plot.gg_boost_effect, gg_boost_vimp, boostmlr_partial

Examples

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, verbose = FALSE
)
pp <- boostmtree::partial.plot(
  fit, x.var.names = "x1", output = "data", verbose = FALSE
)
plot(gg_boost_effect(pp))

---

Boosting error trajectory data object

Description

Extract the error path of a boosted multivariate tree fit as a function of the boosting iteration, together with the optimal iteration selected by cross-validation.

Usage

gg_boost_error(object, use.rmse = TRUE, ...)

Arguments

object	A fitted boostmtree object, or a fitted BoostMLR object.
use.rmse	Logical. When TRUE (default) return the standardized l2 error; when FALSE return the unstandardized squared error.
...	Not used; present for S3 consistency.

Details

The fit must have been grown with cv.flag = TRUE. boostmtree records err.rate and m.opt only when cross-validation ran; a default fit carries neither, and there is no error path to extract. If this function reports that the fit has no error path, refit with cv.flag = TRUE.

boostmtree stores the error on the standardized response scale in the l2 column. use.rmse = FALSE returns (l2 * y.sd)^2, the squared error on the original scale.

This figure also accepts a BoostMLR grow fit (predict objects are not supported – see below). BoostMLR records Error_Rate as an M-by-response matrix already on a single scale, so there is no use.rmse argument for that backend; use.rmse = FALSE is refused, while TRUE, the default, names the scale it already returns. BoostMLR grow objects also select no optimal iteration anywhere in the object – partial.BoostMLR() takes Mopt as a user-supplied argument instead – so optimal is FALSE for every row; deriving an argmin here would report a choice the backend never made.

A BoostMLR predict object is refused outright, even though it shares the "BoostMLR" class and also carries an Error_Rate. Its Error_Rate is test error rather than the training path documented above, and it records a real Mopt that this extractor would otherwise silently discard and overwrite with optimal = FALSE.

Value

A gg_boost_error data.frame with columns:

iteration

Integer boosting iteration, from 1.

value

Numeric error at that iteration.

response

Factor naming the response.

optimal

Logical, TRUE at the cross-validated optimal iteration.

See Also

plot.gg_boost_error, gg_boost_path

Examples

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
)
gg_dta <- gg_boost_error(fit)
plot(gg_dta)

---

Boosting parameter path data object

Description

Extract the estimated correlation, variance, and penalized-spline smoothing parameters of a boosted multivariate tree fit as functions of the boosting iteration.

Usage

gg_boost_path(object, parameters = c("rho", "phi", "lambda"), ...)

Arguments

object	A fitted boostmtree object, or a fitted BoostMLR object.
parameters	Character vector naming the paths to extract. For a boostmtree fit, any of "rho", "phi", and "lambda", defaulting to all three. For a BoostMLR fit, "rho" and/or "phi" only, defaulting to both.
...	Not used; present for S3 consistency.

Details

boostmtree re-estimates three quantities at every boosting iteration: rho, the within-subject correlation; phi, the variance component; and lambda, the P-spline smoothing parameter for the time-covariate interaction. Their paths diagnose whether the variance structure settled or is still drifting when boosting stops.

Unlike the error path, these are recorded on every fit and do not require cv.flag = TRUE. A parameter absent from the fit is dropped silently; it is an error only when none of the requested parameters is present.

This figure also accepts a BoostMLR fit, which records only Rho and Phi as M-by-response matrices – the method's default parameters is c("rho", "phi") rather than all three. BoostMLR has no comparable lambda: its Lambda_List holds per-iteration basis coefficients rather than a scalar smoothing parameter per response, a different quantity, so requesting "lambda" from a BoostMLR fit is refused with that reason instead of silently dropped.

Value

A gg_boost_path data.frame with columns:

iteration

Integer boosting iteration, from 1.

value

Numeric parameter value at that iteration.

parameter

Factor naming the parameter.

response

Factor naming the response.

See Also

plot.gg_boost_path, gg_boost_error

Examples

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_path(fit))

---

Subject trajectory data object

Description

Extract observed and fitted longitudinal trajectories from a boosted multivariate tree fit, one row per observation.

Usage

gg_boost_trajectory(object, ...)

Arguments

object	A fitted boostmtree object, or a fitted BoostMLR object.
...	Not used; present for S3 consistency.

Details

This is the figure longitudinal boosting exists for: whether the model tracks individual subjects over time, rather than only fitting the population mean.

boostmtree stores time, mu and y.org as parallel lists of per-subject vectors, in the order observations were supplied rather than in time order. Rows here are sorted by time within each subject, and the assembled frame is response-major (all subjects for one response, then the next), so a line drawn through one subject's rows follows the trajectory instead of zigzagging. Subject identifiers come from the fit's own id.unique, not from a positional index.

observed is NA throughout when the fit carries no observed response, which happens for a prediction on new data.

This figure accepts either a boostmtree or a BoostMLR fit. BoostMLR stores mu and y as flat observation-by-response matrices, with tm and id as parallel flat vectors, rather than boostmtree's nested per-subject lists – entirely different layout for the same information. Response labels come from y_Names, since mu carries no column names. BoostMLR is natively multi-response, so a fit typically yields several response blocks even for a "single" model.

Value

A gg_boost_trajectory data.frame with columns:

id

Factor subject identifier, taken from id.unique.

time

Numeric observation time, ascending within subject.

fitted

Numeric fitted value.

observed

Numeric observed value, or NA.

response

Factor naming the response.

See Also

plot.gg_boost_trajectory, gg_boost_error

Examples

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, verbose = FALSE
)
plot(gg_boost_trajectory(fit))

---

Variable importance data object

Description

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.

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.

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.

See Also

plot.gg_boost_vimp, gg_boost_effect

Examples

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)))

---