Suppose I run OLS regression and find my residuals to be autocorrelated. When should I use a procedure like Newey-West and when should I use GLS modelling, ie. specifying some ARMA structure to the residuals and refitting the coefficients?
Asked
Active
Viewed 1,303 times
3

Richard Hardy
- 54,375
- 10
- 95
- 219

badmax
- 1,659
- 7
- 19
-
1Possible duplicate of [HAC standard error or missing ARMA terms](https://stats.stackexchange.com/questions/191819/hac-standard-error-or-missing-arma-terms) or [**this**](https://stats.stackexchange.com/questions/181257/correcting-for-autocorrelation-in-simple-linear-regressions-in-r/181297#181297) or [**this**](https://stats.stackexchange.com/questions/226279/how-can-i-handle-autocorrelated-residuals/226441#226441) (itself a duplicate). – Richard Hardy Jan 29 '19 at 20:00
-
1Also related to [**this**](https://stats.stackexchange.com/questions/226120/seasonal-and-autocorrelated-regression-residuals-difference-raw-data-or-residua/226173#226173) and [**this**](https://stats.stackexchange.com/questions/246559/ols-hac-std-err-vs-conditional-mean-equation-from-garch/246620#246620). – Richard Hardy Jan 29 '19 at 20:02
1 Answers
1
I would consider them alternatives for autocorrelated and/or heteroscedastic errors. It's not like you either use one or another. Also, when you don't know the error covariance matrix, you have to estimate it in GLS using Feasible GLS or something like that. Maybe in the case where you somehow know the error covariance matrix GLS will be better, e.g. weighted least squares.
Note, also that GLS will generally produce different coefficients than OLS, while Newey West will not change coefficients.

Aksakal
- 55,939
- 5
- 90
- 176