The goal of sim2Dpredictr
is to facilitate
straightforward simulation of spatially dependent predictors (continuous
or binary), which may then be used to simulate continuous, binary,
count, or categorical (sim2Dpredictr
, which was used to simulate data to evaluate
the performance of methods for high-dimensional data analysis and
prediction (Leach, Aban, and Yi 2022; Leach et al. 2022).
In the first step, we simulate the predictors, i.e., the
Continuous predictors are simulated using Multivariate Normal (MVN)
distributions with a focus on specific correlation structures;
alternatively, one can specify conditional dependence via a precision
matrix, specifically for a Conditional Autoregressive (CAR) model. Tools
are included for easily constructing and taking the Cholesky
decomposition of a covariance or precision matrix with either base
R
or the R
package spam
, which
makes this process faster when the matrix is sparse. The Boolean Model
and thresholding of MVN’s are used to simulate spatially dependent
binary maps. The package also includes a tool for easily specifying a
parameter vector with spatially clustered non-zero elements. These
simulation tools are designed for, but not limited to, testing the
performance of variable selection methods when predictors are spatially
correlated.
In the second step we use the predictor vectors
sim2Dpredictr
is available on CRAN:
install.packages("sim2Dpredictr")
You can install the latest version of sim2Dpredictr
from
GitHub with:
::install_github("jmleach-bst/sim2Dpredictr") devtools
A simple demonstration is as follows; suppose each subject has a
beta_builder()
, simulate and take the Cholesky
decomposition of a correlation (or covariance) matrix with
chol_s2Dp()
, and generate both the images and outcomes with
sim_Y_MVN_X()
.
library(sim2Dpredictr)
# Construct spatially clusterd non-zero parameters.
<- sim2Dpredictr::beta_builder(row.index = c(1, 1, 2),
Bex col.index = c(1, 2, 1),
im.res = c(3, 3),
B0 = 0, B.values = rep(1, 3))
# Construct and take Cholesky decomposition of correlation matrix.
<- sim2Dpredictr::chol_s2Dp(corr.structure = "ar1",
Rex im.res = c(3, 3), rho = 0.5,
use.spam = TRUE)
# Simulate a dataset with spatially dependent design matrix and binary outcomes.
<- sim2Dpredictr::sim_Y_MVN_X(N = 3, B = Bex$B,
sim.dat R = Rex$R, S = Rex$S,
dist = "binomial")
sim.dat#> Y X1 X2 X3 X4 X5 X6 X7
#> 1 0 0.5198678 -1.1035062 -0.805563 -1.3767652 -1.70525968 -0.6388318 -1.5858393
#> 2 0 0.1798315 0.5728877 1.463024 -0.9348662 -0.05594379 0.8592733 0.2619289
#> 3 1 0.1996273 1.7450479 2.500030 1.4882150 1.61903239 1.2503542 0.3439712
#> X8 X9 subjectID
#> 1 -1.72102736 0.06016316 1
#> 2 0.08300277 -0.28904666 2
#> 3 0.63695327 0.85655952 3
Once the dependence framework and non-zero parameter vector is set,
sim_Y_MVN_X()
can be used to draw as many data sets as
necessary, upon each of which variable selection methods are applied;
summaries from each analyzed data set can be obtained and then used to
evaluate variable selection performance. The documentation provides
details about how to use these functions (and others) to create desired
simulations.