getPartialCor: Calculates partial correlation coefficients after retrieving data matrix froma fitted regression model

Description

The input is a fitted regression model, from which the design matrix is retrieved, along with the dependent variable. The partial correlation is calculated using matrix algebra that has not been closely inspected for numerical precision. That is to say, it is in the stats book style, rather than the numerically optimized calculating format that functions like lm() have adopted.

Usage

getPartialCor(model, dvonly = TRUE)

Arguments

model

A fitted regression model, such as output from lm(). Any object that has methods model.matrix and model.frame will be sufficient.

dvonly

Default = TRUE. Only show first column of the full partial correlation matrix. That corresponds to the partial correlation of each predictor with y. I mean, r[yx].[others]

Value

A column or matrix of partial correlation coefficients

Details

I often criticize partial correlations because they change in a very unstable way as terms are added or removed in regression models. Nevertheless, I teach with books that endorse them, and in order to have something to criticize, I need to have a function like this. There are other packages that offer partial correlation calculations, but they are either 1) not easy to install from CRAN because of dependencies or 2) do not directly calculate the values we want to see.

To students. 1) This gives the same result as the function cov2pcor in gRbase, so far as I can tell. Why use this? Simply for convenenience. We have found that installing gRbase is a problem because it depends on packages in Bioconductor. 2) By default, I show only one column of output, the partial correlations involving the dependent variable as something being explained. The other columns that would depict the dependent variable as a predictor of the independent variables have been omitted. You can let me know if you think that's wrong.

Please note I have not gone out of my way to make this calculation "numerically stable." It does not use any orthogonal matrix calculations; it is using the same textbook theoretical stats formula that is used by cov2pcor in gRbase and in every other package or online source I could find. I prepared a little WorkingExample file matrix-partial-correlations-1.R that discusses this, in case you are interested (http://pj.freefaculty.org/R).