mcor function

Compute (Large) Correlation Matrix

Compute a correlation matrix, possibly by robust methods, applicable also for the case of a large number of variables.

mcor(dm, method = c("standard", "Qn", "QnStable",
                    "ogkScaleTau2",  "ogkQn", "shrink"))

Arguments

• dm: numeric data matrix; rows are observiations (samples ), columns are variables.

• method: a string; "standard" (default), "Qn", "QnStable", "ogkQn" and "shrink" envokes standard, elementwise robust (based on $Q_n$ scale estimator, see Qn), robust ($Q_n$ using OGK, see covOGK) or shrinked

  correlation estimate respectively.

Returns

A correlation matrix estimated by the specified method.

Details

The "standard" method envokes a standard correlation estimator. "Qn"

envokes a robust, elementwise correlation estimator based on the Qn scale estimte. "QnStable" also uses the Qn scale estimator, but uses an improved way of transforming that into the correlation estimator. "ogkQn" envokes a correlation estimator based on Qn using OGK. "shrink" is only useful when used with pcSelect. An optimal shrinkage parameter is used. Only correlation between response and covariates is shrinked.

See Also

Qn and covOGK

from package robustbase. pcorOrder for computing partial correlations.

References

See those in the help pages for Qn and covOGK from package robustbase.

Author(s)

Markus Kalisch kalisch@stat.math.ethz.ch and Martin Maechler

Examples

## produce uncorrelated normal random variables
set.seed(42)
x <- rnorm(100)
y <- 2*x + rnorm(100)
## compute correlation of var1 and var2
mcor(cbind(x,y), method="standard")

## repeat but this time with heavy-tailed noise
yNoise <- 2*x + rcauchy(100)
mcor(cbind(x,yNoise), method="standard") ## shows almost no correlation
mcor(cbind(x,yNoise), method="Qn")       ## shows a lot correlation
mcor(cbind(x,yNoise), method="QnStable") ## shows still much correlation
mcor(cbind(x,yNoise), method="ogkQn")    ## ditto