| Title: | Stabilising Variable Selection |
| Version: | 1.0.7 |
| Description: | A stable approach to variable selection through stability selection and the use of a permutation-based objective stability threshold. Lima et al (2021) <doi:10.1038/s41598-020-79317-8>, Meinshausen and Buhlmann (2010) <doi:10.1111/j.1467-9868.2010.00740.x>. |
| License: | MIT + file LICENSE |
| Encoding: | UTF-8 |
| LazyData: | true |
| RoxygenNote: | 7.3.3 |
| Config/testthat/edition: | 3 |
| Depends: | R (≥ 3.0.0) |
| Suggests: | rmarkdown, testthat (≥ 3.0.0), markdown, knitr |
| Imports: | glmnet, dplyr, bigstep, rsample, tibble, purrr, tidyr, stringr, ggplot2, broom, caret, ncvreg, Hmisc, expss, lme4, matrixStats, recipes, stats, lmerTest, furrr, future |
| VignetteBuilder: | knitr |
| NeedsCompilation: | no |
| Packaged: | 2026-03-09 10:01:09 UTC; svzrh2 |
| Author: | Robert Hyde  [aut,
cre],
Eliana Lima [aut],
Matthew Barden [aut],
Kate Lewis [aut],
Martin Green [aut] |
| Maintainer: | Robert Hyde <robert.hyde4@nottingham.ac.uk> |
| Repository: | CRAN |
| Date/Publication: | 2026-03-09 10:40:02 UTC |
stabiliser
Description
This package uses bootstrap resampling and an objective selection stability threshold to provide a robust method of selecting variables truly associated with an outcome.
Author(s)
Robert Hyde robert.hyde4@nottingham.ac.uk
Eliana Lima
Matthew Barden
Martin Green
boot_model
Description
Function to calculate stability of variables' association with an outcome for a given model over a number of bootstrap repeats
Arguments
data |
a dataframe containing an outcome variable to be permuted |
outcome |
the outcome as a string (i.e. "y") |
boot_reps |
the number of bootstrap samples |
model |
the model to be used (i.e. model_mbic) |
model_enet
Description
Function to model elastic net selection process on a given dataframe
Arguments
data |
a dataframe containing an outcome variable to be permuted (usually coming from nested bootstrap data) |
outcome |
the outcome as a string (i.e. "y") |
type |
model type, either "linear" or "logistic" |
model_lasso
Description
Function to model lasso selection process on a given dataframe
Arguments
data |
a dataframe containing an outcome variable to be permuted (usually coming from nested bootstrap data) |
outcome |
the outcome as a string (i.e. "y") |
type |
model type, either "linear" or "logistic" |
model_mbic
Description
Function to model mbic selection process on a given dataframe
Arguments
data |
a dataframe containing an outcome variable to be permuted (usually coming from nested bootstrap data) |
outcome |
the outcome as a string (i.e. "y") |
type |
model type, either "linear" or "logistic" |
model_mcp
Description
Function to model mcp selection process on a given dataframe
Arguments
data |
a dataframe containing an outcome variable to be permuted (usually coming from nested bootstrap data) |
outcome |
the outcome as a string (i.e. "y") |
type |
model type, either "linear" or "logistic" |
model_selector
Description
Determines which models to call.
perm_stab
Description
Main function to call both permutation and bootstrapping functions; to be looped over multiple models selected by the user.
permute
Description
Calculates permutation threshold for null model, where a specified model is run over multiple bootstrap resamples of multiple permuted version of the dataset.
Arguments
data |
a dataframe containing an outcome variable to be permuted |
outcome |
the outcome to be permuted as a string (i.e. "y") |
permutations |
the number of times to be permuted per repeat |
perm_boot_reps |
the number of times to repeat each set of permutations |
quantile |
The quantile of null stabilities to use as a threshold. |
rep_selector_boot
Description
wrapper function to determine the number of bootstrap repeats
Usage
rep_selector_boot(data, boot_reps)
Arguments
data |
the dataset to analyse. |
boot_reps |
the number of bootstrap samples |
rep_selector_boot
Description
wrapper function to determine the number of permutations
Usage
rep_selector_perm(data, permutations)
Arguments
data |
the dataset to analyse. |
permutations |
the number of times to be permuted per repeat |
selection_bias_inner
Description
An function to illustrate the risk of selection bias in conventional modelling approaches by simulating a dataset with no information and conducting conventional modelling with prefiltration.
Arguments
nrows |
The number of rows to simulate. |
ncols |
The number of columns to simulate. |
p_thresh |
The p-value threshold to use in univariate pre-filtration. |
Value
A list including a dataframe of results, a dataframe of the median number of variables selected and a plot illustrating false positive selection.
simulate_data
Description
Simulate a dataset. This can optionally include variables with a given associated with the outcome.
Usage
simulate_data(nrows, ncols, n_true = 0, amplitude = 0)
Arguments
nrows |
The number of rows to simulate. |
ncols |
The number of columns to simulate. |
n_true |
The number of variables truly associated with the outcome. |
amplitude |
The strength of association between true variables and the outcome. |
Value
A simulated dataset
simulate_data_re
Description
Simulate a 500x500 dataset with 8 true fixed effects, 492 junk variables and a clustered outcome suitable for a 2 level random effects analysis. The strength of association between true variables and the outcome is governed by the error added at level 1 (defined by parameter sd_level_1) and level 2 (sd_level_2).
Arguments
sd_level_1 |
Standard deviation of level 1 variables |
sd_level_2 |
Standard deviation of level 2 variables |
Value
A simulated dataset with a clustered outcome sutable for random effects analysis
simulate_data_re_glmer
Description
Simulate a dataset where some variables are associated with the outcome and some are unk
Usage
simulate_glmer_re_data(
n_subjects = 100,
obs_per_subject = 10,
n_signal = 2,
n_noise = 3,
beta0 = -1,
beta_signal = NULL,
sigma_u = 1
)
Arguments
n_subjects |
The number of individual subjects, e.g. participations |
obs_per_subject |
The number of observations per subject |
n_signal |
The number of causal predictors |
n_noise |
The number of junk predictors |
beta0 |
Intercept |
beta_signal |
signal size for causal parameters |
sigma_u |
standard deviation for random intercepts |
Value
A simulated dataset with a clustered outcome suitable for random effects analysis with a binary outcome
simulate_selection_bias
Description
An function to illustrate the risk of selection bias in conventional modelling approaches by simulating a dataset with no information and conducting conventional modelling with prefiltration.
Arguments
nrows |
A vector of the number of rows to simulate (i.e., c(100, 200)). |
ncols |
A vector of the number of columns to simulate (i.e., c(100, 200)). |
p_thresh |
A vector of the p-value threshold to use in univariate pre-filtration (i.e., c(0.1, 0.2)). |
Value
A list including a dataframe of results, a dataframe of the median number of variables selected and a plot illustrating false positive selection.
stab_plot
Description
Plot from stability object
Arguments
stabiliser_outcome |
Outcome from stabilise() or triangulate() function. |
Value
A ggplot object.
stabilise
Description
Function to calculate stability of variables' association with an outcome for a given model over a number of bootstrap repeats
Arguments
data |
A dataframe containing an outcome variable to be permuted. |
outcome |
The outcome as a string (i.e. "y"). |
boot_reps |
The number of bootstrap samples. Default is "auto" which selects number based on dataframe size. |
permutations |
The number of times to be permuted per repeat. Default is "auto" which selects number based on dataframe size. |
perm_boot_reps |
The number of times to repeat each set of permutations. Default is 20. |
models |
The models to select for stabilising. Default is elastic net (models = c("enet")), other available models include "lasso", "mbic", "mcp". |
type |
The type of model, either "linear" or "logistic" |
quantile |
The quantile of null stabilities to use as a threshold. |
normalise |
Normalise numeric variables (TRUE/FALSE) |
dummy |
Create dummy variables for factors/characters (TRUE/FALSE) |
impute |
Impute missing data (TRUE/FALSE) |
Value
A list for each model selected. Each list contains a dataframe of variable stabilities, a numeric permutation threshold, and a dataframe of coefficients for both bootstrap and permutation.
stabilise_re
Description
Function to calculate stability of variables' association with an outcome for a given model over a number of bootstrap repeats using clustered data.
Arguments
data |
A dataframe containing an outcome variable to be permuted. |
outcome |
The outcome as a string (i.e. "y"). |
intercept_level_ids |
A vector names defining which variables are random effect, i.e., c("level_2_column_name", "level_3_column_name"). |
n_top_filter |
The number of variables to filter for final model (Default = 50). |
boot_reps |
The number of bootstrap samples. Default is "auto" which selects number based on dataframe size. |
permutations |
The number of times to be permuted per repeat. Default is "auto" which selects number based on dataframe size. |
perm_boot_reps |
The number of times to repeat each set of permutations. Default is 20. |
normalise |
Normalise numeric variables (TRUE/FALSE) |
dummy |
Create dummy variables for factors/characters (TRUE/FALSE) |
impute |
Impute missing data (TRUE/FALSE) |
Value
A list containing a table of variable stabilities and a numeric permutation threshold.
stabilise_re_glmer
Description
Function to calculate stability of variables' association with an outcome for a given model over a number of bootstrap repeats using clustered data.
Arguments
data |
A dataframe containing an outcome variable to be permuted. |
outcome |
The outcome as a string (i.e. "y"). |
intercept_level_ids |
A vector names defining which variables are random effect, i.e., c("level_2_column_name", "level_3_column_name"). |
n_top_filter |
The number of variables to filter for final model (Default = 50). |
boot_reps |
The number of bootstrap samples. Default is "auto" which selects number based on dataframe size. For glmer models, these are subsamples of the dataset, set to 80%. |
permutations |
The number of times to be permuted per repeat. Default is "auto" which selects number based on dataframe size. |
perm_boot_reps |
The number of times to repeat each set of permutations. Default is 20. |
normalise |
Normalise numeric variables (TRUE/FALSE) |
dummy |
Create dummy variables for factors/characters (TRUE/FALSE) |
impute |
Impute missing data (TRUE/FALSE) |
base_id |
level of the random effect to bootstrap by, e.g individual. This is likely the lower level of random effect specified |
parallel |
TRUE or FALSE, whether to set up parallel processing |
num_cores |
Number of cores to use if parallel processing required |
Value
A list containing a table of variable stabilities and a numeric permutation threshold.
stabiliser_example
Description
A simulated dataset
Usage
stabiliser_example
Format
A data frame with 50 rows and 100 variables.
The stabiliser_example dataset is a simulated example with the following properties:
1 simulated outcome variable: y
4 variables simulated to be associated with y: causal1, causal2...
95 variables simulated to have no association with y: junk1, junk2...
stabiliser_prep
Description
Prepares dataset using recipes framework
Arguments
normalise |
Normalise numeric variables (TRUE/FALSE) |
dummy |
Create dummy variables for factors/characters (TRUE/FALSE) |
impute |
Impute missing data (TRUE/FALSE) |
triangulate
Description
Triangulate multiple models using a stability object
Arguments
object |
An object generated through the stabilise() function. |
quantile |
The quantile of null stabilities to use as a threshold. |
Value
A combined list of model results including a dataframe of stability results for variables and a numeric permutation threshold.