Type:	Package
Title:	Stability-Selection via Correlated Resampling for Beta-Regression Models
Version:	0.4.5
Date:	2025-11-04
Depends:	R (≥ 4.0)
Imports:	betareg, gamlss, gamlss.dist, glmnet, MASS, Rcpp, rlang, stats, withr
Suggests:	future, future.apply, gamlss.lasso, ggplot2, knitr, rmarkdown, testthat (≥ 3.0.0)
LinkingTo:	Rcpp, RcppArmadillo
Author:	Frederic Bertrand [cre, aut]
Maintainer:	Frederic Bertrand <frederic.bertrand@lecnam.net>
Description:	Adds variable-selection functions for Beta regression models (both mean and phi submodels) so they can be used within the 'SelectBoost' algorithm. Includes stepwise AIC, BIC, and corrected AIC on betareg() fits, 'gamlss'-based LASSO/Elastic-Net, a pure 'glmnet' iterative re-weighted least squares-based selector with an optional standardization speedup, and 'C++' helpers for iterative re-weighted least squares working steps and precision updates. Also provides a fastboost_interval() variant for interval responses, comparison helpers, and a flexible simulator simulation_DATA.beta() for interval-valued data. For more details see Bertrand and Maumy (2023) <doi:10.7490/f1000research.1119552.1>.
License:	GPL-3
Encoding:	UTF-8
Classification/MSC:	62H11, 62J12, 62J99
VignetteBuilder:	knitr
RoxygenNote:	7.3.3
URL:	https://fbertran.github.io/SelectBoost.beta/, https://github.com/fbertran/SelectBoost.beta/
BugReports:	https://github.com/fbertran/SelectBoost.beta/issues/
Config/testthat/edition:	3
NeedsCompilation:	yes
Packaged:	2025-11-04 09:49:20 UTC; bertran7
Repository:	CRAN
Date/Publication:	2025-11-11 09:20:09 UTC

SelectBoost.beta: Stability-Selection via Correlated Resampling for Beta-Regression Models

Description

Adds variable-selection functions for Beta regression models (both mean and phi submodels) so they can be used within the 'SelectBoost' algorithm. Includes stepwise AIC, BIC, and corrected AIC on betareg() fits, 'gamlss'-based LASSO/Elastic-Net, a pure 'glmnet' iterative re-weighted least squares-based selector with an optional standardization speedup, and 'C++' helpers for iterative re-weighted least squares working steps and precision updates. Also provides a fastboost_interval() variant for interval responses, comparison helpers, and a flexible simulator simulation_DATA.beta() for interval-valued data. For more details see Bertrand and Maumy (2023) doi:10.7490/f1000research.1119552.1.

Author(s)

This package was written by Frédéric Bertrand. Maintainer: Frédéric Bertrand frederic.bertrand@lecnam.net

References

Bertrand, F. and Maumy, M. (2023). Improving variable selection in Beta regression models using correlated resampling. Presented at Joint Statistical Meetings 2023, Toronto, Canada.

Bertrand, F. and Maumy, M. (2023). SelectBoost4Beta: Improving variable selection in Beta regression models. Presented at BioC2023, Boston, USA.

See Also

Useful links:

• https://fbertran.github.io/SelectBoost.beta/

• https://github.com/fbertran/SelectBoost.beta/

• Report bugs at https://github.com/fbertran/SelectBoost.beta/issues/

Examples

set.seed(1)
n <- 150; p <- 6
X <- matrix(rnorm(n*p), n, p); colnames(X) <- paste0("x",1:p)
eta <- 0.4 + X[,1] - 0.7*X[,3]
mu  <- plogis(eta)
Y   <- rbeta(n, mu*25, (1-mu)*25)

betareg_step_aic(X, Y)    # should return (Intercept) + x1,x3 nonzero often
betareg_step_bic(X, Y)
betareg_step_aicc(X, Y)

Beta regression Elastic-Net via GAMLSS (gamlss.lasso)

Description

Uses gamlss.lasso::gnet() to fit ENet on the mean submodel of gamlss(dist = BE). The routine assumes complete cases and does not expose offsets or precision-model terms.

Usage

betareg_enet_gamlss(
  X,
  Y,
  method = c("IC", "CV"),
  ICpen = c("BIC", "AIC", "HQC"),
  alpha = 1,
  trace = FALSE
)

Arguments

Value

Named numeric vector of coefficients as in betareg_lasso_gamlss().

See Also

gamlss.lasso::gnet(), gamlss::gamlss(), gamlss.dist::BE()

Pure glmnet IRLS selector for Beta regression

Description

Runs an IRLS loop with Beta working responses/weights and calls glmnet on the weighted least-squares surrogate. Supports BIC/AIC/CV model choice and an optional prestandardize speedup. The helper uses only the mean submodel, requires complete cases, and does not expose offset terms.

Usage

betareg_glmnet(
  X,
  Y,
  alpha = 1,
  choose = c("bic", "aic", "cv"),
  nfolds = 5,
  n_iter = 6,
  tol = 1e-05,
  standardize = TRUE,
  lambda = NULL,
  phi_init = 20,
  update_phi = TRUE,
  phi_maxit = 5,
  prestandardize = FALSE,
  trace = FALSE
)

Arguments

Value

Named numeric vector (Intercept) + colnames(X) with zeros for unselected variables.

See Also

glmnet::glmnet(), glmnet::cv.glmnet()

Examples

set.seed(1); X <- matrix(rnorm(500), 100, 5); Y <- plogis(X[,1]-0.5*X[,3])
Y <- rbeta(100, Y*40, (1-Y)*40)
betareg_glmnet(X, Y, alpha = 1, choose = "bic", prestandardize = TRUE)

Beta regression LASSO via GAMLSS

Description

Uses gamlss::ri() (L1 penalty) in a gamlss(dist = BE) mean submodel to select variables. The helper works on complete cases of X/Y, targets the mean component, and does not yet expose offset handling.

Usage

betareg_lasso_gamlss(
  X,
  Y,
  method = c("ML", "GAIC"),
  k = 2,
  degf = NULL,
  lambda = NULL,
  trace = FALSE
)

Arguments

Value

Named numeric vector of coefficients (Intercept) + colnames(X), with 0 for unselected variables.

See Also

gamlss::gamlss(), gamlss::ri(), gamlss.dist::BE()

Examples

set.seed(1); X <- matrix(rnorm(300), 100, 3); Y <- plogis(X[,1]); Y <- rbeta(100, Y*30, (1-Y)*30)
betareg_lasso_gamlss(X, Y, method = "GAIC", k = 2)

Stepwise Beta regression by AIC

Description

Fits a Beta regression with optional joint selection of the mean and precision (phi) submodels using betareg::betareg(). The routine performs greedy forward/backward search using the requested information criterion and returns coefficients aligned with the supplied design matrix. The selectors currently target the mean submodel only, require complete cases, and do not expose offsets. Observation weights are passed through to betareg() when provided.

Usage

betareg_step_aic(
  X,
  Y,
  direction = "both",
  link = "logit",
  link.phi = "log",
  type = "ML",
  trace = FALSE,
  max_steps = NULL,
  epsilon = 1e-08,
  X_phi = NULL,
  direction_phi = c("none", "both", "forward", "backward"),
  weights = NULL
)

Arguments

Value

Named numeric vector of length p_mean + p_phi + 1 containing the intercept, mean coefficients, phi-intercept (prefixed by "phi|"), and phi coefficients (also prefixed by "phi|"). Non-selected variables have coefficient 0.

See Also

betareg::betareg()

Examples

set.seed(1)
X <- matrix(rnorm(200), 100, 2);
Y <- plogis(0.5 + X[,1]-X[,2]);
betareg_step_aic(X, Y)
Y <- rbeta(100, Y*20, (1-Y)*20)
betareg_step_aic(X, Y)

Stepwise Beta regression by AICc (finite-sample corrected AIC)

Description

Greedy forward/backward search minimizing AICc computed on betareg fits with optional precision-submodel selection and observation weights.

Usage

betareg_step_aicc(
  X,
  Y,
  direction = "both",
  link = "logit",
  link.phi = "log",
  type = "ML",
  trace = FALSE,
  max_steps = NULL,
  epsilon = 1e-08,
  X_phi = NULL,
  direction_phi = c("none", "both", "forward", "backward"),
  weights = NULL
)

Arguments

Value

See betareg_step_aic().

Examples

set.seed(1);
X <- matrix(rnorm(400), 100, 4);
Y <- plogis(X[,1]+0.5*X[,2])
betareg_step_aicc(X, Y)
Y <- rbeta(100, Y*25, (1-Y)*25);
betareg_step_aicc(X, Y)

Stepwise Beta regression by BIC

Description

Stepwise Beta regression by BIC

Usage

betareg_step_bic(
  X,
  Y,
  direction = "both",
  link = "logit",
  link.phi = "log",
  type = "ML",
  trace = FALSE,
  max_steps = NULL,
  epsilon = 1e-08,
  X_phi = NULL,
  direction_phi = c("none", "both", "forward", "backward"),
  weights = NULL
)

Arguments

Value

See betareg_step_aic().

Examples

set.seed(1); X <- matrix(rnorm(300), 100, 3);
Y <- plogis(X[,1]);
betareg_step_bic(X, Y)
Y <- rbeta(100, Y*30, (1-Y)*30)
betareg_step_bic(X, Y)

Bootstrap selection frequencies across selectors

Description

Bootstraps the dataset B times and records how often each variable is selected by each selector. Observations containing NA in either X or Y are removed prior to resampling. Column names are abbreviated internally and mapped back to the originals in the output just like in compare_selectors_single().

Usage

compare_selectors_bootstrap(X, Y, B = 50, include_enet = TRUE, seed = NULL)

Arguments

Value

Long data frame with columns selector, variable, freq in ⁠[0,1]⁠, n_success, and n_fail. The freq column reports the share of bootstrap replicates where a variable was selected by the corresponding selector. Values near 1 signal high stability whereas small values indicate weak evidence. n_success counts the successful fits contributing to the frequency estimate (excluding failed replicates), while n_fail records the number of unsuccessful fits. A "failures" attribute attached to the returned data frame lists the replicate indices and messages for any encountered errors.

Examples

set.seed(1)
X <- matrix(rnorm(300), 100, 3); Y <- plogis(X[, 1])
Y <- rbeta(100, Y * 30, (1 - Y) * 30)
freq <- compare_selectors_bootstrap(X, Y, B = 10, include_enet = FALSE)
head(freq)
subset(freq, freq > 0.8)


# Increase B until the reported frequencies stabilise. For example,
freq_big <- compare_selectors_bootstrap(X, Y, B = 200, include_enet = FALSE)
stats::aggregate(freq ~ selector, freq_big, summary)

Run all selectors once on a dataset

Description

Convenience wrapper that runs AIC/BIC/AICc stepwise, GAMLSS LASSO (and ENet when available), and the pure glmnet IRLS selector, then collates coefficients into a long table for comparison. Observations containing NA in either X or Y are removed prior to fitting. Column names are temporarily shortened to satisfy selector requirements and avoid clashes; the outputs remap them to the original labels before returning so the reported variables always match the input design.

Usage

compare_selectors_single(X, Y, include_enet = TRUE)

Arguments

Value

A list with:

coefs

Named coefficient vectors for each selector.

table

Long data frame with columns selector, variable, coef, selected.

Examples

set.seed(1)
X <- matrix(rnorm(300), 100, 3); Y <- plogis(X[, 1])
Y <- rbeta(100, Y * 30, (1 - Y) * 30)
single <- compare_selectors_single(X, Y, include_enet = FALSE)
head(single$table)

Merge single-run results and bootstrap frequencies

Description

Merge single-run results and bootstrap frequencies

Usage

compare_table(single_tab, freq_tab = NULL)

Arguments

Value

Merged data frame.

Examples

single_tab <- data.frame(
  selector = rep(c("AIC", "BIC"), each = 3),
  variable = rep(paste0("x", 1:3), times = 2),
  coef = c(0.5, 0, -0.2, 0.6, 0.1, -0.3)
)
single_tab$selected <- single_tab$coef != 0
freq_tab <- data.frame(
  selector = rep(c("AIC", "BIC"), each = 3),
  variable = rep(paste0("x", 1:3), times = 2),
  freq = c(0.9, 0.15, 0.4, 0.85, 0.3, 0.25)
)
compare_table(single_tab, freq_tab)

Interval-response stability selection (fastboost variant)

Description

Repeats selection on interval-valued responses by sampling a pseudo-response from each interval (uniformly or midpoint), tallying variable selection frequencies across B replicates.

Usage

fastboost_interval(
  X,
  Y_low,
  Y_high,
  func,
  B = 100,
  sample = c("uniform", "midpoint"),
  version = "glmnet",
  use.parallel = FALSE,
  seed = NULL,
  ...
)

Arguments

Value

A list with:

betas

⁠B × (p+1)⁠ matrix of coefficients over replicates.

freq

Named vector of selection frequencies for each predictor.

Examples

# suppose you have interval data (Y_low, Y_high)
set.seed(1)
n <- 120; p <- 6
X <- matrix(rnorm(n*p), n, p); colnames(X) <- paste0("x",1:p)
mu <- plogis(X[,1] - 0.5*X[,2]); Y <- rbeta(n, mu*25, (1-mu)*25)
Y_low <- pmax(0, Y - 0.05); Y_high <- pmin(1, Y + 0.05)
fb <- fastboost_interval(X, Y_low, Y_high,
       func = function(X,y) betareg_glmnet(X,y, choose="bic", prestandardize=TRUE),
       B = 40)
sort(fb$freq, decreasing = TRUE)

Side-by-side coefficient heatmap

Description

Visual comparison of coefficients returned by each selector. Requires ggplot2.

Usage

plot_compare_coeff(single_tab)

Arguments

Value

A ggplot object when ggplot2 is available; otherwise draws a base R image.

Examples

demo_tab <- data.frame(
  selector = rep(c("AIC", "BIC"), each = 3),
  variable = rep(paste0("x", 1:3), times = 2),
  coef = c(0.6, 0, -0.2, 0.55, 0.05, -0.3)
)
demo_tab$selected <- demo_tab$coef != 0
plot_compare_coeff(demo_tab)

Side-by-side selection-frequency heatmap

Description

Visual comparison of bootstrap selection frequencies by selector. Requires ggplot2.

Usage

plot_compare_freq(freq_tab)

Arguments

Value

A ggplot object when ggplot2 is available; otherwise draws a base R image.

Examples

freq_tab <- data.frame(
  selector = rep(c("AIC", "BIC"), each = 3),
  variable = rep(paste0("x", 1:3), times = 2),
  freq = c(0.85, 0.2, 0.45, 0.75, 0.35, 0.3)
)
plot_compare_freq(freq_tab)

Apply a selector to a collection of resampled designs

Description

Apply a selector to a collection of resampled designs

Usage

sb_apply_selector_manual(
  X_norm,
  resamples,
  Y,
  selector,
  ...,
  keep_template = TRUE
)

Arguments

Value

A numeric matrix of coefficients with one column per resample.

SelectBoost for beta-regression models

Description

sb_beta() orchestrates all SelectBoost stages—normalisation, correlation analysis, grouping, correlated resampling, and stability tallying—while using the beta-regression selectors provided by this package. It can operate on point-valued or interval-valued responses and automatically squeezes the outcome into ⁠(0, 1)⁠ unless instructed otherwise.

Usage

sb_beta(
  X,
  Y = NULL,
  selector = betareg_step_aic,
  corrfunc = "cor",
  B = 100,
  step.num = 0.1,
  steps.seq = NULL,
  version = c("glmnet", "lars"),
  squeeze = TRUE,
  use.parallel = FALSE,
  seed = NULL,
  verbose = FALSE,
  threshold = 1e-04,
  interval = c("none", "uniform", "midpoint"),
  Y_low = NULL,
  Y_high = NULL,
  ...
)

Arguments

Details

The returned object carries a rich set of attributes:

• "c0.seq" – the grid of absolute-correlation thresholds explored during resampling.

• "steps.seq" – the raw sequence (if any) used to construct the grid.

• "selector" – the selector identifier (function name or expression).

• "B" – number of resampled designs passed to the selector.

• "interval" – the interval sampling mode ("none", "uniform", or "midpoint").

• "resample_diagnostics" – per-threshold data frames with summary statistics on the cached correlated draws.

These attributes mirror the historical SelectBoost beta implementation so the object can be consumed by existing plotting and reporting utilities.

Value

Matrix of selection frequencies with one row per c0 level and class "sb_beta". See Details for the recorded attributes.

Examples

set.seed(42)
sim <- simulation_DATA.beta(n = 80, p = 4, s = 2)
# increase B for real applications
res <- sb_beta(sim$X, sim$Y, B = 5)
res

SelectBoost workflow for interval responses

Description

sb_beta_interval() forwards to sb_beta() while activating interval sampling so that beta-regression SelectBoost runs can ingest lower/upper response bounds directly. It mirrors fastboost_interval() but reuses the correlated resampling pipeline of sb_beta().

Usage

sb_beta_interval(
  X,
  Y_low,
  Y_high,
  selector = betareg_step_aic,
  sample = c("uniform", "midpoint"),
  Y = NULL,
  ...
)

Arguments

Value

See sb_beta(). The returned object carries the same "sb_beta"-class attributes describing the correlation thresholds, resampling diagnostics, selector, and number of replicates.

Examples

set.seed(1)
sim <- simulation_DATA.beta(n = 120, p = 5, s = 2)
y_low <- pmax(sim$Y - 0.05, 0)
y_high <- pmin(sim$Y + 0.05, 1)
interval_fit <- sb_beta_interval(
  sim$X,
  Y_low = y_low,
  Y_high = y_high,
  B = 5,
  step.num = 0.4
)
attr(interval_fit, "interval")

User-friendly methods for sb_beta() results

Description

These S3 helpers make it easier to inspect and visualise the correlation-threshold grid returned by sb_beta(). They surface the stored attributes, reshape the selection frequencies into tidy summaries, and produce quick ggplot2 visualisations for interactive use.

Usage

## S3 method for class 'sb_beta'
print(x, digits = 3, ...)

## S3 method for class 'sb_beta'
summary(object, ...)

## S3 method for class 'summary.sb_beta'
print(x, digits = 3, n = 10, ...)

autoplot.sb_beta(object, variables = NULL, ...)

Arguments

Value

summary.sb_beta() returns an object of class summary.sb_beta containing a tidy data frame of selection frequencies. The plotting and printing methods are invoked for their side effects and return the input object invisibly.

Examples

set.seed(42)
sim <- simulation_DATA.beta(n = 50, p = 4, s = 2)
fit <- sb_beta(sim$X, sim$Y, B = 5, step.num = 0.5)
print(fit)
summary(fit)
if (requireNamespace("ggplot2", quietly = TRUE)) {
  autoplot.sb_beta(fit)
}

Core helpers for SelectBoost-style beta regression

Description

These helpers expose the individual stages of the SelectBoost workflow so that beta-regression selectors can be combined with correlation-aware resampling directly from SelectBoost.beta. They normalise the design matrix, derive correlation structures, form groups of correlated predictors, generate Gaussian surrogates that mimic the observed dependency structure, and apply a user-provided selector on each resampled design.

Usage

sb_normalize(X, center = NULL, scale = NULL, eps = 1e-08)

sb_compute_corr(X, corrfunc = "cor")

sb_group_variables(corr_mat, c0)

Arguments

Value

sb_normalize() returns a centred, \ell_2-scaled copy of X.

sb_compute_corr() returns the association matrix.

sb_group_variables() returns a list of integer vectors, one per variable, describing the correlated group it belongs to.

Examples

sb_normalize(matrix(rnorm(20), 5))

Generate correlated design replicates for a set of groups

Description

Generate correlated design replicates for a set of groups

Usage

sb_resample_groups(
  X_norm,
  groups,
  B = 100,
  jitter = 1e-06,
  seed = NULL,
  use.parallel = FALSE,
  cache = NULL
)

Arguments

Details

When every group has size one (no correlated variables) the function simply returns B copies of X_norm. A warning is issued in that situation so downstream code can avoid mistaking the replicated designs for genuinely resampled surrogates. The covariance matrices underpinning each correlated draw are cached in the supplied cache environment; reusing the environment across calls lets sb_resample_groups() skip redundant covariance decompositions for identical groups and speeds up iterative workflows.

Value

An object of class sb_resamples, i.e. a list of length B whose elements are resampled design matrices. The object exposes per-group diagnostics in its "diagnostics" attribute and returns the cache via the "cache" attribute for reuse.

Compute selection frequencies from coefficient paths

Description

Compute selection frequencies from coefficient paths

Usage

sb_selection_frequency(
  coef_matrix,
  version = c("glmnet", "lars"),
  threshold = 1e-04
)

Arguments

Value

Numeric vector of selection frequencies.

Simulate interval Beta-regression data (flexible)

Description

Simulate interval Beta-regression data (flexible)

Usage

simulation_DATA.beta(
  n,
  p,
  s = min(5L, p),
  beta_size = 1,
  a0 = 0,
  X_dist = c("gaussian", "t", "bernoulli"),
  corr = c("indep", "ar1", "block"),
  rho = 0,
  block_size = 5L,
  df = 5,
  prob = 0.5,
  active_idx = NULL,
  phi = 20,
  mechanism = c("jitter", "quantile", "mixed"),
  mix_prob = 0.5,
  delta = 0.05,
  delta_low = NULL,
  delta_high = NULL,
  alpha = 0.1,
  alpha_low = NULL,
  alpha_high = NULL,
  na_rate = 0,
  na_side = c("random", "left", "right"),
  centerX = FALSE,
  scaleX = FALSE,
  seed = NULL
)

Arguments

Value

list with ⁠X, Y, Y_low, Y_high, mu, beta, a0, phi, info, settings⁠.