Durbin Watson related with outliers?

#1
I am trying to run a multiple regression with a sample data belongs to a cohort of 500 subjects. My dependant variable is a psychological score (scale), independant variables are age(scale), sex(nominal) and education(scale). I run the test and Durbin-Watson was 1.99 which is good but I had 2 outliers. When I selected and removed outliers and re-run the test, Durbin-Watson became .002. Can you help me with this issue please. I dont know the reasons and how to go on. Everytime I remove an outlier shown by casewise diagnostics, DW becomes .002 or something like that. :( Anybody can help me with this please?
Thank you,
Rezzak
 

noetsi

Fortran must die
#2
Why are you running Durbin Watson to start with? It looks like you have cross-sectional data and normally autoregression (which is what Durbin Watson captures) won't be present in cross sectional data unless they way you gathered your data somehow creates it (which it should not have).

It is very unusual to run Durbin-Watson with non-time series data. But perhaps you are using time series and I missed this [I don't see any indication time is built into your model].

It is very difficult to imagine why removing two cases of five hundred would introduce or remove autocorrelation from the data. How does he diagnostics you mention determine what an outliers is (there are many ways to do this including standardized residuals, DFBETA, leverage, etc).
 
#3
Thank you very much for your help. I am new in stats, so the website I am paying to learn (Laerd statistics) told me to run DW in order to meet the assumption of independance of errors. Yes, it is cross-sectional data. And so my takeaway lesson was we do not need DW if the data has no time series?
 

noetsi

Fortran must die
#4
In rare cases you can violate the assumption of independence without time series data. This almost always will involve gathering your data in such a way that you force your responses to be related (for instance you gather your data from husbands and wives in carrying out analysis not directly related to family). You don't need Durbin-Watson for this, you know how you gathered your data. Unless you did this cross-sectional data is normally not in violation of independence so you don't normally do Durbin-Watson with cross sectional data.

Moreover, Durbin-Watson only captures first order autocorrelation and thus is less than ideal as a measure of this phenomenon. There are better test such as Breusch-Pagan, Durbin's t, etc. Durbin-Watson is brought up a lot because it was among the first and thus is well known. Even though there are far better test now.