Stability selection - Using stabs"

knitr::opts_chunk$set(echo = TRUE)

required <- c("lars", "mboost")
if (!all(sapply(required, function(pkg) requireNamespace(pkg, quietly = TRUE))))
    knitr::opts_chunk$set(eval = FALSE)

stabs implements resampling procedures to assess the stability of selected variables with additional finite sample error control for high-dimensional variable selection procedures such as Lasso or boosting. Both, standard stability selection (Meinshausen & B├╝hlmann, 2010, doi:10.1111/j.1467-9868.2010.00740.x) and complementarty pairs stability selection with improved error bounds (Shah & Samworth, 2013, doi:10.1111/j.1467-9868.2011.01034.x) are implemented. The package can be combined with arbitrary user specified variable selection approaches.



To be able to use the install_github() command, one needs to install devtools first:


Using stabs

A simple example of how to use stabs with package lars:

## make data set available
data("bodyfat", package = "")
## set seed

## lasso
(stab.lasso <- stabsel(x = bodyfat[, -2], y = bodyfat[,2],
                       fitfun = lars.lasso, cutoff = 0.75,
                       PFER = 1))

## stepwise selection
(stab.stepwise <- stabsel(x = bodyfat[, -2], y = bodyfat[,2],
                          fitfun = lars.stepwise, cutoff = 0.75,
                          PFER = 1))

Now plot the results

## plot results
par(mfrow = c(1, 2))
plot(stab.lasso, main = "Lasso")
plot(stab.stepwise, main = "Stepwise Selection")

We can see that stepwise selection seems to be quite unstable, even in this low dimensional example!

User-specified variable selection approaches

To use stabs with user specified functions, one can specify an own fitfun. These need to take arguments x (the predictors), y (the outcome) and q the number of selected variables as defined for stability selection. Additional arguments to the variable selection method can be handled by .... In the function stabsel() these can then be specified as a named list which is given to args.fitfun.

The fitfun function then needs to return a named list with two elements selected and path: selected is a vector that indicates which variable was selected. path is a matrix that indicates which variable was selected in which step. Each row represents one variable, the columns represent the steps. The latter is optional and only needed to draw the complete selection paths.

The following example shows how lars.lasso is implemented:


To see more examples simply print, e.g., lars.stepwise, glmnet.lasso, or glmnet.lasso_maxCoef. Please contact me if you need help to integrate your method of choice.

Using boosting with stability selection

Instead of specifying a fitting function, one can also use stabsel directly on computed boosting models from mboost.

### low-dimensional example
mod <- glmboost(DEXfat ~ ., data = bodyfat)

## compute cutoff ahead of running stabsel to see if it is a sensible
## parameter choice.
##   p = ncol(bodyfat) - 1 (= Outcome) + 1 ( = Intercept)
stabsel_parameters(q = 3, PFER = 1, p = ncol(bodyfat) - 1 + 1,
                   sampling.type = "MB")
## the same:
stabsel(mod, q = 3, PFER = 1, sampling.type = "MB", eval = FALSE)

## now run stability selection
(sbody <- stabsel(mod, q = 3, PFER = 1, sampling.type = "MB"))

Now let us plot the results, as paths and as maximum selection frequencies:

opar <- par(mai = par("mai") * c(1, 1, 1, 2.7), mfrow = c(1, 2))
plot(sbody, type = "paths")
plot(sbody, type = "maxsel", ymargin = 6)


To cite the package in publications please use


which will currently give you


To obtain BibTeX references use


Try the stabs package in your browser

Any scripts or data that you put into this service are public.

stabs documentation built on Jan. 29, 2021, 5:14 p.m.