My data han N=90 observation, I plot my data and I don't see any peculiar plot indicating heteroscedaticity, so I run BP and other test like Goldfeld-Quandt and Harrison-McCabe, but BP return significantly and other test is not with enourmous gap of P-Value (BP=0.040 while other 0.9). Can any one help what should I do with this result?
$\begingroup$
$\endgroup$
3
-
$\begingroup$ Depends on your data and your theoretical model. Is there good reason to expect serial correlation between observations? If so, model it and see whether it makes a difference. If not, use HAC standard errors. $\endgroup$– DurdenCommented Jan 14 at 15:13
-
$\begingroup$ I use linear regression to model comparrison between theorical result vs real time result, so there are two output from same predictor $\endgroup$– tolakCommented Jan 14 at 15:37
-
$\begingroup$ There are some people who would argue that these tests are useless, and would basically cite was is happening here as evidence. Actually, maybe I'll be one of those people. You are right to trust what the resids vs fitted plot is showing you instead of the point-null test. $\endgroup$– John MaddenCommented Jan 14 at 19:16
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Your graph does not have the classical "funnel" shape that one would expect to see in the presence of heteroscedasticity where there is clearly an uneven variance across residuals.
The Breusch-Pagan test is a go-to test for heteroscedasticity, and a p-value of 0.040 warrants further investigation.
If you are using R, one potential course of action would be to implement a Box-Cox transformation on the data and then generate a QQ plot as well as a plot of residuals vs. fitted values to determine whether variance has become constant as a result of the transformation. Of course, you could also run the Breusch-Pagan test once again to determine whether the p-value becomes insignificant across the transformed data.
I suggest referring to the following for further information: R-bloggers: How to detect heteroscedasticity and rectify it?
As a final caveat, the source of the heteroscedasticity will also be significantly dependent on your data and model, as a previous comment alluded to - this should factor in to your decision as to what steps you should take following heteroscedasticity testing.
answered Jan 14 at 18:50