The two examples below implement Bayesian linear and logistic
regression models using the horseshoe prior over parameters to encourage
a sparse model. Models are fitted using the hsstan
R
package, which performs full Bayesian inference through a
Stan
implementation. In Bayesian inference model
meta-parameters such as the amount of shrinkage are also given prior
distributions and are thus directly learned from the data through
sampling. This bypasses the need to cross-validate results over a grid
of values for the meta-parameters, as would be required to find the
optimal lambda in a lasso or elastic net model.
However, Bayesian inference is computationally intensive. In
high-dimensional settings, with e.g. more than 10,000 biomarkers,
pre-filtering of inputs based on univariate measures of association with
the outcome may be beneficial. If pre-filtering of inputs is used then a
cross-validation procedure is needed to ensure that the data points used
for pre-filtering and model fitting differ from the data points used to
quantify model performance. The outercv()
function is used
to perform univariate pre-filtering and cross-validate model performance
in this setting.
CAUTION should be used when setting the number of cores available for
parallelisation. The default setting in hsstan
is to use 4
cores to parallelise Markov chains in the Bayesian inference procedure.
This can be switched off either by setting
options(mc.cores = 1)
.
Argument cv.cores
in outercv()
controls
parallelisation over the outer CV folds. On unix/mac setting
cv.cores
to >1 will induce nested parallelisation which
will generate an error, unless parallelisation of the chains is disabled
by setting options(mc.cores = 1)
.
Nested parallelisation is feasible if cv.cores
is >1
and multicore_fork = FALSE
is set as this uses cluster
based parallelisation instead of mclapply
. Beware that
large numbers of processes may be spawned: the code will use the product
of cv.cores
and mc.cores
. If we are performing
10-fold cross-validation with 4 chains and set
cv.cores = 10
then 40 processes will be invoked
simultaneously.
We use cross-validation and apply univariate filtering of predictors and model fitting in one part of the data (training fold), followed by evaluation of model performance on the left-out data (testing fold), repeated in each fold.
Only one cross-validation split is needed (function
outercv()
) as the Bayesian model does not require
cross-validation for meta-parameters.
Note, in the examples below length of sampling and number of chains
is curtailed for speed. We recommend 4 chains, warmup
1000
and iter
2000 in practice.
library(hsstan)
# number of cores for parallelising hsstan sampling
# store original options in oldopt
# at the end reset options to old configuration
oldopt <- options(mc.cores = 2)
# load iris dataset and simulate a continuous outcome
data(iris)
dt <- iris[, 1:4]
colnames(dt) <- c("marker1", "marker2", "marker3", "marker4")
dt <- as.data.frame(apply(dt, 2, scale))
dt$outcome.cont <- -3 + 0.5 * dt$marker1 + 2 * dt$marker2 + rnorm(nrow(dt), 0, 2)
# unpenalised covariates: always retain in the prediction model
uvars <- "marker1"
# penalised covariates: coefficients are drawn from hierarchical shrinkage prior
pvars <- c("marker2", "marker3", "marker4") # penalised covariates
# run cross-validation with univariate filter and hsstan
res.cv.hsstan <- outercv(y = dt$outcome.cont, x = dt[, c(uvars, pvars)],
model = "model.hsstan",
filterFUN = lm_filter,
filter_options = list(force_vars = uvars,
nfilter = 2,
p_cutoff = NULL,
rsq_cutoff = 0.9),
chains = 2, # chains parallelised via options(mc.cores)
n_outer_folds = 3, cv.cores = 1, # CV folds not parallelised
unpenalized = uvars, warmup = 100, iter = 200)
We can then view prediction performance based on the testing folds
and examine the Bayesian model using the hsstan
package.
summary(res.cv.hsstan)
#> Single cross-validation to measure performance
#> Outer loop: 3-fold CV
#> No inner loop
#> 150 observations, 4 predictors
#>
#> Model: model.hsstan
#> Filter: lm_filter
#> n.filter
#> Fold 1 3
#> Fold 2 3
#> Fold 3 3
#>
#> Final fit: mean sd 2.5% 97.5% n_eff Rhat
#> (Intercept) -3.17 0.14 -3.40 -2.89 221 0.99
#> marker1 0.37 0.32 -0.36 0.90 209 1.00
#> marker2 2.07 0.22 1.70 2.47 196 1.00
#> marker3 0.11 0.31 -0.43 0.92 104 1.00
#>
#> Result:
#> RMSE Rsquared MAE
#> 2.1226 0.4772 1.6971
sampler.stats(res.cv.hsstan$final_fit)
#> accept.stat stepsize divergences treedepth gradients warmup sample
#> chain:1 0.9843 0.0241 0 8 14844 0.25 0.12
#> chain:2 0.9722 0.0316 0 8 11356 0.12 0.09
#> all 0.9783 0.0279 0 8 26200 0.37 0.21
print(projsel(res.cv.hsstan$final_fit), digits = 4) # adding marker2
#> Model KL ELPD
#> Intercept only 0.32508 -374.99055
#> Initial submodel 0.32031 -374.72416
#> marker2 0.00151 -325.82289
#> marker3 0.00000 -325.88824
#> var kl rel.kl.null rel.kl elpd delta.elpd
#> 1 Intercept only 0.325077 0.00000 NA -375.0 -49.10230
#> 2 Initial submodel 0.320306 0.01468 0.0000 -374.7 -48.83591
#> 3 marker2 0.001506 0.99537 0.9953 -325.8 0.06535
#> 4 marker3 0.000000 1.00000 1.0000 -325.9 0.00000
Here adding marker2
improves the model fit: substantial
decrease of KL-divergence from the full model to the submodel. Adding
marker3
does not improve the model fit: no decrease of
KL-divergence from the full model to the submodel.
We use cross-validation and apply univariate filtering of predictors and model fitting in one part of the data (training fold), followed by evaluation of model performance on the left-out data (testing fold), repeated in each fold.
Only one cross-validation split is needed (function
outercv
) as the Bayesian model does not require
cross-validation for meta-parameters.
# sigmoid function
sigmoid <- function(x) {1 / (1 + exp(-x))}
# load iris dataset and create a binary outcome
set.seed(267)
data(iris)
dt <- iris[, 1:4]
colnames(dt) <- c("marker1", "marker2", "marker3", "marker4")
dt <- as.data.frame(apply(dt, 2, scale))
rownames(dt) <- paste0("sample", c(1:nrow(dt)))
dt$outcome.bin <- sigmoid(0.5 * dt$marker1 + 2 * dt$marker2) > runif(nrow(dt))
dt$outcome.bin <- factor(dt$outcome.bin)
# unpenalised covariates: always retain in the prediction model
uvars <- "marker1"
# penalised covariates: coefficients are drawn from hierarchical shrinkage prior
pvars <- c("marker2", "marker3", "marker4") # penalised covariates
# run cross-validation with univariate filter and hsstan
res.cv.hsstan <- outercv(y = dt$outcome.bin,
x = as.matrix(dt[, c(uvars, pvars)]),
model = "model.hsstan",
filterFUN = ttest_filter,
filter_options = list(force_vars = uvars,
nfilter = 2,
p_cutoff = NULL,
rsq_cutoff = 0.9),
chains = 2, # parallelise over chains
n_outer_folds = 3, cv.cores = 1,
unpenalized = uvars, warmup = 100, iter = 200)
We view prediction performance based on testing folds and examine the model.
summary(res.cv.hsstan)
#> Single cross-validation to measure performance
#> Outer loop: 3-fold CV
#> No inner loop
#> 150 observations, 4 predictors
#> FALSE TRUE
#> 78 72
#>
#> Model: model.hsstan
#> Filter: ttest_filter
#> n.filter
#> Fold 1 3
#> Fold 2 3
#> Fold 3 3
#>
#> Final fit: mean sd 2.5% 97.5% n_eff Rhat
#> (Intercept) -0.12 0.23 -0.56 0.27 200 1
#> marker1 0.50 0.30 -0.10 1.16 208 1
#> marker2 1.91 0.36 1.26 2.66 263 1
#> marker3 0.01 0.24 -0.55 0.69 171 1
#>
#> Result:
#> Reference
#> Predicted FALSE TRUE
#> FALSE 56 23
#> TRUE 22 49
#>
#> AUC Accuracy Balanced accuracy
#> 0.8284 0.7000 0.6993
# examine the Bayesian model
print(projsel(res.cv.hsstan$final_fit), digits = 4) # adding marker2
#> Model KL ELPD
#> Intercept only 0.20643 -104.26957
#> Initial submodel 0.20118 -104.17411
#> marker2 0.00060 -73.84232
#> marker3 0.00000 -73.93210
#> var kl rel.kl.null rel.kl elpd delta.elpd
#> 1 Intercept only 2.064e-01 0.00000 NA -104.27 -30.33748
#> 2 Initial submodel 2.012e-01 0.02543 0.000 -104.17 -30.24201
#> 3 marker2 5.964e-04 0.99711 0.997 -73.84 0.08977
#> 4 marker3 9.133e-18 1.00000 1.000 -73.93 0.00000
options(oldopt) # reset options
Here adding marker2
improves the model fit: substantial
decrease of KL-divergence from the full model to the submodel. Adding
marker3
does not improve the model fit: no decrease of
KL-divergence from the full model to the submodel.
At time of writing, there appears to be a bug in rstan
(used by hsstan
) leading to it ignoring the pass-thru
argument cores
and instead spawning multiple processes as
specified by the chain
argument. This behaviour can be
limited by setting options(mc.cores = ..)
.
A key problem with parallelisation in R is that errors, warnings and
user input have to be suppressed during multicore processing. If a
nestedcv
call is not working, we recommend that you try it
with cv.cores = 1
first to check it starts up without error
messages.
If you use this package, please cite as:
Lewis MJ, Spiliopoulou A, Goldmann K, Pitzalis C, McKeigue P, Barnes MR (2023). nestedcv: an R package for fast implementation of nested cross-validation with embedded feature selection designed for transcriptomics and high dimensional data. Bioinformatics Advances. https://doi.org/10.1093/bioadv/vbad048
Carpenter, B., et al. Stan: A Probabilistic Programming Language. Journal of Statistical Software 2017;76(1):1-32.
Piironen, J. and Vehtari, A. Sparsity information and regularization in the horseshoe and other shrinkage priors. Electronic Journal of Statistics 2017;11(2):5018-5051, 5034.