Examples for the corrgram package"

In corrgram: Plot a Correlogram

Package overview

The corrgram package provides functions for creating corrgrams using three different graphics systems, base, grid, and lattice.

Base R graphics + single function corrgram() for dataframes or matrices. + Enables most features found in the paper by @friendly2002corrgrams. - No automatic legend. - Not easily combined with other graphics.

lattice graphics + Separate panel functions for lattice::levelplot() for dataframes and lattice::splom() for correlation matrices. + Enables automatic legend. + Enables corrgrams conditioned on other variables. + Can be combined with other lattice graphics for complex figures. - Not feature complete compared to base R.

grid graphics + single function corrgram2() for either dataframes or correlation matrices. + Enables automatic legend. + Can be combined with other grid graphics for complex figures. - Not feature complete compared to base R. + Faster than base R when evaluated inside Positron.

This vignette

This vignette demonstrates corrgrams with base R graphics and reproduces most of the figures in @friendly2002corrgrams.

Setup

library("knitr")
opts_chunk$set(fig.align="center", fig.width=6, fig.height=6)
options(width=90)

The data are 11 measures of performance and salary for 263 baseball players in the 1986 baseball season in the United States. The data were used in 1988 Data Expo at the Joint Statistical Meetings.

The first 6 rows of the data and the upper-left corner of the correlation matrix are given below.

library("corrgram")
head(baseball)
round(cor(baseball[, 5:14], use="pair"),2)

Figure 2

Figure 2 shows two ways to graphically display the correlation matrix using the panel.shade() and panel.pie() functions.

vars2 <- c("Assists","Atbat","Errors","Hits","Homer","logSal",
           "Putouts","RBI","Runs","Walks","Years")
corrgram(baseball[,vars2], order=TRUE,
         main="Baseball data PC2/PC1 order",
         lower.panel=panel.shade, upper.panel=panel.pie,
         diag.panel=panel.minmax, text.panel=panel.txt)

Figure 3

Figure 3 shows an eigenvector plot of the correlation matrix. This forms the basis of the orderings of the variables in the corrgram in Figure 4b in the next section, beginning with "AtBat" and then moving counter-clockwise up to "Years". Note: Eigenvectors are unique only up to a change of sign.

baseball.cor <- cor(baseball[,vars2], use='pair')
baseball.eig <- eigen(baseball.cor)$vectors[,1:2]
e1 <- baseball.eig[,1]
e2 <- baseball.eig[,2]
plot(e1,e2,col='white', xlim=range(e1,e2), ylim=range(e1,e2))
text(e1,e2, rownames(baseball.cor), cex=1)
title("Eigenvector plot of baseball data")
arrows(0, 0, e1, e2, cex=0.5, col="red", length=0.1)

Figure 4a, 4b

In figure 4a the variables are sorted in the alphabetical order as given in the data.

In figure 4b, the variables are sorted according to the principal component ordering in Figure 3 to look for possible clustering of the variables. It is not surprising to see that more times at bat is strongly correlated with a higher number of hits and a higher number of runs.

corrgram(baseball[,vars2], main="Baseball data (alphabetic order)")

corrgram(baseball[,vars2], order=TRUE,
         main="Baseball data (PC order)",
         panel=panel.shade, text.panel=panel.txt)

Figure 5

Figure 5 shows a corrgram for all numeric variables in the dataframe. Non-numeric columns in the data are ignored.

corrgram(baseball, order=TRUE, main="Baseball data (PC order)")

Figure 6.

Figure 6 shows a corrgram of automotive data on 74 different models of cars from 1979. There are two obvious groups of variables

Note, the arrangement is slightly different from Friendly.

corrgram(auto, order=TRUE, main="Auto data (PC order)")

Figure 7.

The inverse of the correlation matrix expresses conditional dependence and independence of the variables.

The variables are sorted in the same order as in figure 4. One example interpretation is: controlling for all other variables, there is still a large correlation between Years and log Salary.

rinv <- function(r){
  # r is a correlation matrix
  # calculate r inverse and scale to correlation matrix
  # Derived from Michael Friendly's SAS code

  ri <- solve(r)
  s <- diag(ri)
  s <- diag(sqrt(1/s))
  ri <- s %*% ri %*% s
  n <- nrow(ri)
  ri <- ri * (2*rep(1,n) - matrix(1, n, n))
  diag(ri) <- 1  # Should already be 1, but could be 1 + epsilon
  colnames(ri) <- rownames(ri) <- rownames(r)
  return(ri)
}

vars7 <- c("Years", "logSal", "Homer", "Putouts", "RBI", "Walks",
           "Runs", "Hits", "Atbat", "Errors", "Assists")
cb <- cor(baseball[,vars7], use="pair")
corrgram(-rinv(cb), main=expression(paste("Baseball data ", R^-1)))

Figure 8

Figure 8 shows a partial independence corrgram for the automotive data, when Price and MPG are partialed out.

require(Matrix) # For block diagonal function

partial <- function(r, xvar){
  # r is a correlation matrix
  # Calculate partial correlation of y|x
  yvar <- setdiff(colnames(r), xvar)
  ri <- r[yvar,yvar] - r[yvar,xvar] %*% solve(r[xvar,xvar]) %*% r[xvar,yvar]
  s <- diag(ri)
  s <- diag(sqrt(1/s))
  ri <- s %*% ri %*% s
  ri <- as.matrix(Matrix::bdiag(ri, r[xvar, xvar]))
  diag(ri) <- 1  # Should already be 1, but could be 1 + epsilon
  colnames(ri) <- rownames(ri) <- c(yvar, xvar)
  return(ri)
}

vars8a <- c("Gratio", "Rep78", "Rep77", "Hroom", "Trunk", "Rseat",
            "Length", "Weight", "Displa", "Turn")
vars8b <- c("MPG", "Price")
vars8 <- c(vars8a, vars8b)
auto.cor <- cor(auto[, vars8], use="pair")
auto.par <- partial(auto.cor, vars8b)
corrgram(auto.par,
         lower.panel=panel.pie, upper.panel=panel.pie,
         main="Auto data, partialing out Price,MPG")

Figure 11

Figure 11 provides another way to display the data, using both ellipses and loess lines. Long, narrow ellipses represent high correlations while circular ellipses represent low correlations.

corrgram(baseball[,vars2], order=TRUE,
         main="Baseball correlation ellipses",
         panel=panel.ellipse,
         text.panel=panel.txt, diag.panel=panel.minmax)

Further examples

Demonstrate density panel, correlation confidence panel

corrgram(iris,
         main="Iris data with example panel functions",
         lower.panel=panel.pts, upper.panel=panel.conf,
         diag.panel=panel.density)

Demonstrate panel.bar, panel.ellipse, panel.minmax, col.regions

corrgram(auto, order=TRUE,
         main="Auto data (PC order)",
         lower.panel=corrgram::panel.ellipse,
         upper.panel=panel.bar, diag.panel=panel.minmax,
         col.regions=colorRampPalette(c("darkgoldenrod4", "burlywood1",
                                        "darkkhaki", "darkgreen")))

Correlation matrix

The vote data is a matrix.

# 'vote' is a correlation matrix, not a data frame
corrgram(vote, order=TRUE,
         upper.panel=panel.cor, main="vote")

Ratings data

An example showing one way to plot ratings data.

load(url("https://github.com/alexanderrobitzsch/sirt/blob/master/data/data.ratings3.rda?raw=true"))

# jitter first, so the upper/lower panels are symmetric
data.ratings3 <- transform(data.ratings3,
                           c2=jitter(crit2), c3=jitter(crit3),
                           c4=jitter(crit4), c6=jitter(crit6))

library(corrgram)
panel.raters <- function (x, y, corr = NULL, col.regions, cor.method, ...) {
  if (!is.null(corr)) 
    return()
  plot.xy(xy.coords(x, y), type = "p", ...)
  abline(lm(y ~ x))
  box(col = "lightgray")
}
corrgram(data.ratings3[,7:10], diag=panel.density, lower.panel=panel.raters, upper.panel=panel.conf)

References

corrgram documentation built on Sept. 3, 2026, 1:06 a.m.