simcity

The goal of simcity is to streamline simulating data to evaluate and compare methods for some common statistical tasks.

Installation

You can install the development version of simcity from GitHub with:

# install.packages("devtools")
devtools::install_github("joftius/simcity")

Example

The first motivating example for this package was simulating high-dimensional linear regression models, fitting them with the lasso, and analyzing various performance metrics.

library(simcity)
n <- 100
p <- 200
s0 <- 5
one_lasso_fit <- instance_hdr(n, p, s0)
one_lasso_fit |> head()
which(one_lasso_fit$true_beta != 0)
which(one_lasso_fit$estimate != 0)

The above example generates one instance of simulated data, fits a regression model using glmnet::cv.glmnet and the lambda.1se option by default. The simcity package streamlines doing processes like this many times, and typically finishes in about half the time (or less) by using parallel processing.

niters <- 200
many_lasso_fits <- simulate_hdr(niters, n, p, s0, cores = 4)
sim_summary <- simmary_coefs(many_lasso_fits)
head(sim_summary)
ggplot(sim_summary, aes(screened, beta_min)) + geom_boxplot()
ggplot(sim_summary, aes(beta_min, mse)) + geom_point()

The above example generates a number niters of instances of simulated data, fits each of them using cv.glmnet, and computes some interesting summaries about the overall results.