This is a repost of an example that I posted last year but at the time I only had the PDF document (written in ). I’m reposting it directly into WordPress and I’m including the graphs.

From time-to-time a researcher needs to develop a script or an application to collect and analyze data. They may also need to test their application under a variety of scenarios prior to data collection. However, because the data has not been collected yet it is necessary to create test data. Creating continuous data is relatively simple and is fairly straight forward using the Cholesky (pronounced kol-eh-ski) decomposition. This approach takes an original X variable (or matrix) and uses the Cholesky transformation to create a new, correlated, Y variable. To make things simple and straight forward this example will generate data from the a random normal distribution N(0,1).

The reason this approach is so useful is that that correlation structure can be specifically defined. The scripts can be used to create many different variables with different correlation structures. The method to transform the data into correlated variables is seen below using the correlation matrix R.

Once the correlation matrix is set the researcher takes the Cholesky decomposition of the correlation matrix. Multiplying the Cholesky decomposition of the correlation matrix by the data matrix the resulting matrix is a transformed dataset with the specified correlation.

The R code from below will generate a correlation matrix of:

```
```R = matrix(cbind(1,.80,.2, .80,1,.7, .2,.7,1),nrow=3)
U = t(chol(R))
nvars = dim(U)[1]
numobs = 100000
set.seed(1)
random.normal = matrix(rnorm(nvars*numobs,0,1), nrow=nvars, ncol=numobs);
X = U %*% random.normal
newX = t(X)
raw = as.data.frame(newX)
orig.raw = as.data.frame(t(random.normal))
names(raw) = c("response","predictor1","predictor2")
cor(raw)
plot(head(raw, 100))
plot(head(orig.raw,100))

### Like this:

Like Loading...

## Steve Walker

/ March 11, 2013Cool stuff. Here’s my take on it:

http://stevencarlislewalker.wordpress.com/2012/06/05/simulating-random-variables-with-a-particular-correlation-structure/

And here’s a little package for doing this stuff (on R-forge not CRAN):

install.packages(“rmv”, repos=”http://R-Forge.R-project.org”)

Here’s the syntax:

rmv(n, covmat, rfunc = rnorm,

method = c(“chol”, “eigen”), …)

## fd

/ March 11, 2013Thanks, this is very useful. I know I’ve learned this in the past but I’m fuzzy on the details. Do you have a citation or link that provides more details on why this works?

## Steve Walker

/ March 11, 2013Here are my notes on why it works

http://stevencarlislewalker.files.wordpress.com/2012/06/covariancematrixsqrtmethod.png

Hope it helps.

## Bob Muenchen

/ March 12, 2013That’s some very useful code. Thanks!

## ddaa

/ March 14, 2013What if I need different distributions on different variables (e.g. first column normal distributed, second one uniform, third one gamma).

Thanks in advance for your inputs.

## Wesley

/ March 14, 2013When you create the variable ”random.normal <-” you can construct the matrix for each dimension. So rather than having ”nrows=nvars” just set it equal to one (1). Then you can rbind the different distributions together. So that way you can have, for example, a Normal(0,1), a Uniform(1,2), and a Gamma(9,.5). Each will be on a different row of the matrix random.normal.

## ddaa

/ March 18, 2013It works but this is what you get as correlation matrix:

response predictor1 predictor2

response 1.00 0.84 0.09

predictor1 0.84 1.00 0.27

predictor2 0.09 0.27 1.00

In fact this was the solution I gave to myself but also the answer comply with my suspects.

## RogerS

/ March 30, 2014Any chance you can post some code with different distributions for each row/variable?

Thanks

## ddaa

/ March 18, 2013P.S.: a more appropriate answer can be Copula

## itfeature.com

/ June 9, 2013Thanks for data generation. Statistical Simulation

## Vince Diego

/ May 28, 2015This is great! Do you have any recommendations for what to do when the Cholesky decomposition cannot be computed?

## Vince Diego

/ May 28, 2015Here’s an approach in R that I found after asking my question:

http://www.r-bloggers.com/dealing-with-non-positive-definite-matrices-in-r/

## Tony Kamsi

/ June 4, 2015Pls,I want to generate random multivariate data in R for two data sets ie canonical sets using exponential and gamma.Help