3

Let's say in linear regression, I got a fit and I can plot residuals to see whether there is any systematic trend in such a plot. How to quantitatively determine whether the residues are really random? Is Durbin-Watson test used for this purpose? How to interpret such test if so?

Please provide an example, preferably in R.

littleEinstein
  • 523
  • 1
  • 5
  • 7
  • 2
    **Related**: http://stats.stackexchange.com/questions/14914/how-to-test-the-autocorrelation-of-the-residuals – cardinal Dec 17 '11 at 21:53
  • 1
    See also the [Breusch-Pagan test](http://en.wikipedia.org/wiki/Breusch%E2%80%93Pagan_test), where a reference to an $R$ implementation is also given. – cardinal Dec 17 '11 at 21:53

1 Answers1

2

This also depends on the type of data you have. Do you have a time series? Then looking at a ACF/PACF plot of the residuals may be one way of determining if these are white noise:

require(tseries)
tsdiag(rnorm(100))

you may find this PDF helpful.

If you have cross-section data an inspection of the residual plots yields very valuable information in my opinion:

data(cars)
plot(lm(speed~dist, data=cars))

There exist tests (see comments) but I very often find graphs more useful.

Seb
  • 571
  • 3
  • 15