5
$\begingroup$

How to interpret the labels with the outlier numbers when you plot the following in R (QQplot)

set.seed(1)
y <- rnorm(100)
x <- rnorm(100)
plot(lm(y ~ x), which=2)   # which = 2 gives the plot

It gives a number 61 on the top. What is it?

I figured it might be the index of the outlier couple. It appears to be connected to a score of around y = 3 and x = 3. But when:

cbind(y,x)[61,]

>  y         x 
2.4016178 0.4251004 

How to read these numbers in R's QQplot?

$\endgroup$
2
  • $\begingroup$ What I still don't get is why the number 61 is mentioned here; the residual does not seem heterogeneous? $\endgroup$ Commented Jun 6, 2013 at 15:10
  • 2
    $\begingroup$ because the 61st standardized residual is ~2.5, not the 61st observation. Check the axis labels on your plot, and try rstandard(lm(y~x))[61]. The x-axis is not x but the (approximate) expected largest order statistic of a standard normal. $\endgroup$
    – Glen_b
    Commented Jun 7, 2013 at 1:25

1 Answer 1

13
$\begingroup$

The number in the plot corresponds to the indices of the standardized residuals and the original data. By default, R labels the three most extreme residuals, even if they don't deviate much from the QQ-line. So the fact that the points are labelled doesn't mean that the fit is bad or anything. This behaviour can be changed by specifying the option id.n. Let me illustrate this with your example

set.seed(1)
y <- rnorm(100)
x <- rnorm(100)
lm.mod <- lm(y ~ x) # linear regression model
plot(lm.mod, which=2) # QQ-Plot
lm.resid <- residuals(lm(y ~ x)) # save the residuals
sort(abs(lm.resid), decreasing=TRUE) # sort the absolute values of the residals
        14         61         24
2.32415869 2.29316200 2.09837122

The first three most extreme residuals are number 14, 61 and 24. These are the numbers in the plot. These indices correspond to the indices of the original data. So the data points 14, 24 and 26 are the ones that cause the most extreme residuals. We can also mark them in a scatterplot (the blue points). Note that because you generated your y and x independently, the regression line is simply the mean of y without any slope:

# The original data points corresponding to the 3 most extreme residuals

cbind(x,y)[c(14, 24, 61), ]
             x         y
[1,] -0.6506964 -2.214700
[2,] -0.1795565 -1.989352
[3,]  0.4251004  2.401618

# Make a scatterplot of the original data and mark the three points
# and add the residuals

par(bg="white", cex=1.6)
plot(y~x, pch=16, las=1)
abline(lm.mod, lwd=2) # add regression line
pre <- predict(lm.mod)

# Add the residual lines
segments(x[c(14, 24, 61)], y[c(14, 24, 61)], x[c(14, 24, 61)], 
         pre[c(14, 24, 61)], col="red", lwd=2)

# Add the points
points(x[c(14, 24, 61)], y[c(14, 24, 61)], pch=16, cex=1.1, col="steelblue", las=1)

Residuals

$\endgroup$
1
  • $\begingroup$ nice explanation and plot $\endgroup$
    – Glen_b
    Commented Jun 7, 2013 at 1:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.