# Durbin-Watson vs. Test of AutoCorrelation

I’ve realized that I missed out the time series analysis part in my reading (painfully this is a trouble part)…Got a question when doing practice question: This is about question 17. Practice Problems for Reading 13 (sorry I can’t type in here as the question is quite long & includes tables). But if you do care, please open the book. I got different thought from the book’s answer for question a (if the model is correct). 1.According to the book they only look at the autocorrelation’s t-test. As long as the t(s) are insignificant, we cannot reject null hypothesis that there’s no autocorrelation of the residual, hence the model is correct. 2.However I’m still concern about the Durbin-Watson. According to the Durbin-Watson statistics (1.8784), which is higher than the D(u), isn’t there a negative serial correlation of the errors, which violate regression assumption, hence we can’t conduct the hypothesis on coefficient. Anyone can share a thought on the (2) - Durbin-Watson concern? Thanks.

I don’t have the book with me - but I know for sure that if [DW-stat] > Du, then it’s a positive serial correlation (and not negative.)

agree with swap. most of the time if DWcritical is > Dupper, then its a positive correlation. if Dupper is not known, then 2 is a good value to use as mid-point between positive and negative serial correlation. EDIT - whoops, now im gonna fail for sure. i flipped the table in my mind. positive serial correlation is on the left hand side and negative is on the right. anything above Dupper will have either none, inconclusive, or negative serial correlation.

Oh yah. Typo Typo…:">

Btw, I’ve just firgured out why. It’s simply that DW is not appropriate to use here. Thanks guy. I found that this forum is a very good way for me to write down the mess in my thought, got the solution, and then remember it :). 22 days till the exam …

thats good. i don’t have the book with me but i’ll post if otherwise when i get home. swap, flip your DW table.

swaption, remember that DW ~ 2(1-r), where r is correlation. So when r is really negative, DW approaches 4, so when DW > Du, then it shows negative correlation. swaptiongamma Wrote: ------------------------------------------------------- > I don’t have the book with me - but I know for > sure that if > Du, then it’s a positive serial > correlation (and not negative.)

For autoregressions, you use t-stats to test if the correlation of the residuals is statistically significant. if you are just doing a normal regression, you would use DW to test for serial correlation.

Sorry, what this r stand for? Is the correlation of residual term itself? If yes, how to calculate it?

you do not use DW for time series…need to look at the autocorrelations…

You can use DW for a trend model, but not for auto-regressive model.

deriv108 Wrote: ------------------------------------------------------- > You can use DW for a trend model, but not for > auto-regressive model. yep, just happened to finish up on Log-linear trend models. From the Stalla text “However, to be sure that serial correlation is not still a problem, the DW test must be performed”. I skipped most of the discussion, but I think you get the point. If you need reference info, ping back.

I’m using schweser notes. It seems to me that autocorrelation shall get time involved. In DW, is the “r” a lag-1 correlation coef of residual?