Interpreting Durbin-Watson results

Question

I have fitted a glm to my data set and used to the Durbin-Watson test to check model fit. I have obtained the result. How can i interprete it?

lag    Autocorrelation     D-W Statistic    p-value

   1       0.7750748     0.4466024       0

Alternative hypothesis: rho != 0

score 2 · Answer 1 · edited Mar 21 '19 at 19:55

2

If the test statistic = approx 2 then then there is no autocorrelation between variables. If <2 or >2 (a lot) then you have a problem (negative or positive autocorrelation). Autocorrelation indicates the order of observations has some effect on the response. Example: If residuals are serially correlated, the order of observations affects the response.

edited Mar 21 '19 at 19:55

Elsa Braun

3
2

answered Mar 09 '14 at 18:49

Dino Abraham

439
5
11

@user40494 if you are happy with the answer by Dino Abraham please don't forget to accept his answer by ticking the green check mark on the left-hand side of his post. – Andy Mar 09 '14 at 19:28

score 1 · Answer 2 · answered Mar 09 '14 at 21:42

The DW statistic tests whethwer or not there is first order autocorrelation in the residuals. It ignores a host of other possible violations likw autocorrelation of lag 12 or any other lag in the errors. It assumwes many things such as no pulses,no seasonal pulses, no level shifts and no time trends in the residuals. It assumes that the errror variance is constant over time. It assumes that there is no omitted structure do to omitted lags of the predictor series. In short it is relatively useless in my opinion . More powerful tests are the Ljumg-Box test , Intervention Detection tests and Tsay's test for constanct of error variance and Chow's test for constancy of model parameters.

Interpreting Durbin-Watson results

2 Answers2