gen_data: Simulate data for regression models

In pbreheny/hdrm: High-Dimensional Regression Modeling

Simulate data for regression models

Description

This function is designed to scale efficiently to high dimensions, and therefore imposes some restrictions. For example, correlation must be positive.

Usage

gen_data(
  n,
  p,
  p1 = floor(p/2),
  beta,
  family = c("gaussian", "binomial"),
  SNR = 1,
  signal = c("homogeneous", "heterogeneous"),
  corr = c("exchangeable", "autoregressive"),
  rho = 0
)

Arguments

n	Sample size
p	Number of features
p1	Number of nonzero features
beta	Vector of regression coefficients in the generating model, or, if a scalar, the value of each nonzero regression coefficient.
family	Generate y according to linear "gaussian" or logistic "binomial" model
SNR	Signal to noise ratio
signal	Should the beta coefficients be homogeneous (default) or heterogeneous
corr	Correlation structure between features ('exchangeable' | 'autoregressive')
rho	Correlation coefficient

Details

Note that if beta is not supplied, this function must calculate the SNR to determine an appropriate coefficient size. This will be slow if the dimension is large and beta is not sparse.

Examples

dat <- gen_data(100, 100, 10)
dim(dat$X)
head(dat$y)
head(dat$beta)

gen_data(100, 10, 5)$beta
gen_data(100, 10, 5, SNR=2)$beta
gen_data(100, 10, 5, SNR=2, corr='exch', rho=0.7)$beta
gen_data(100, 10, 5, SNR=2, corr='auto', rho=0.7)$beta
gen_data(100, 10, 5, SNR=2, corr='auto', rho=0.7, signal='het')$beta
gen_data(100, 10, 5, SNR=2, corr='auto', rho=0.1, signal='het')$beta
gen_data(100, 10, 5, SNR=2, corr='auto', rho=0.1, signal='het', b=1)$beta

gen_data(10, 10, 5, family='binomial')$y

gen_data(1000, 10, rho=0.0, corr='exch')$X |> cor() |> round(digits=2)
gen_data(1000, 10, rho=0.7, corr='exch')$X |> cor() |> round(digits=2)
gen_data(1000, 10, rho=0.7, corr='auto')$X |> cor() |> round(digits=2)
gen_data(1000, 3, 3, rho=0)$X |> cor() |> round(digits=2)

pbreheny/hdrm documentation built on Jan. 17, 2024, 8:53 p.m.