2/5/2009
Diagnostics and remedial measures
- Tests
for Constant Variance
- Brown-Forsythe Test (robust)
- Breusch-Pagan Test (large-sample, normal errors)
- F-test
for lack of fit
- Must have at least 2 repeated observations on at
least one level of x
- Start with full model
- Consider the reduced model under the appropriate
null hypothesis
- The difference between the two error sums of
squares is SSLF (see Figure 3.12 on p. 125)
- See modified ANOVA table on p. 126
- Examples
- Transformations…
- …on X for nonlinearity
- …on Y for non-normality or non-constant
variance of errors
- Box-Cox Transformations - A general family of
power transformations
- Scatterplot Smoothing Methods
- Method of moving averages
- Method of running medians
- Locally weighted scatterplot
smoothing (LOWESS)
Please prepare the following for a problem session on Tuesday:
3.3, 3.11, 3.18, 3.22