What to do if my durbin watson is found to be inconclusive?
Do I have to change my model?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
There are cases error serial correlation is a disaster. For example, if your model is $y_t = \beta_0 + \beta_1 y_{t-1} + u_t$, then serial correlation in $u_t$ means correlation of $y_{t-1}$ and $u_t$ (in general) and your OLS estimator is biased and inconsistent. In other cases, serial correlation does not cause endogeneity and OLS is still consistent. An example is the case the right-hand side variable is strictly exogenous. Then you can use OLS and heteroskedasticity and auto-correlation consistent (HAC) inferences, e.g., using the Newey-West standard errors.
So my answer is "it depends". If some regressors become endogenous due to the error serial correlation, you will probably want to change your model. Otherwise, you can use OLS and HAC standard errors.
If your model is dynamic ($y_{t-1}$ on the right-hand side), the chance is high that you need to seriously worry about serial correlation in $u_t$. This happens in many time-series and financial applications.
By the way, Farebrother (1980, JRSS Series C) proposed Pan's procedure for $p$ values for DW tests. Inconclusiveness can be avoided.
-
$\begingroup$ Thank you so much @chan1142. I think I'll just have to remodel it. 🤗 $\endgroup$– Ry GCommented Apr 10, 2019 at 6:55