Newey West bartlett Kernel automatic bandwidth
Posted: Tue Feb 05, 2013 5:41 am
Hello
I' m working with autocorrelation and experiencing the effect of different bandwidths in the results of a model.With fixed bandwidth I have no problems but when it comes to automatic bandwidth,I get a difference in the result.
Particularly I'm dealing with Newey-West estimator with kernel function Bartlett and automatic bandwidth. According to technical notes (page 780 users guide II) before determining BT(bandwidth) I need to estimate alfa(q),this case p=1 because of Bartlett function.
For alfa(1), I'm estimating AR(1) models (without intercept) for each element of X(t)*e(t) that is the product of original matrix X of independent variables (except for intercept) times the estimated error.
So if I have k parameters(including intercept) I need k-1 AR(1) models.And for each one I pick both the coefficient and standard deviation estimated to plug in the
the formula for the calculation of alfa(1) provided by Eviews (formula D.11 page 180)
If I consider only a parameter rather than the intercept, the result for the bandwidth is ok but if there is more than one independent variable the result no longer matchs.
Note:The expressions for alfa(1) still depends on the weighting vector Ws that is considered to be Ws=1 so it cancels out.
I don´t know if there is any weighting we must consider but the formula seems very clear
To make things clear I attach 2 simple models (with one and two independent variables) only for help purpose.
If somebody can help I'll appreciate
Tks
Joao Pereira
I' m working with autocorrelation and experiencing the effect of different bandwidths in the results of a model.With fixed bandwidth I have no problems but when it comes to automatic bandwidth,I get a difference in the result.
Particularly I'm dealing with Newey-West estimator with kernel function Bartlett and automatic bandwidth. According to technical notes (page 780 users guide II) before determining BT(bandwidth) I need to estimate alfa(q),this case p=1 because of Bartlett function.
For alfa(1), I'm estimating AR(1) models (without intercept) for each element of X(t)*e(t) that is the product of original matrix X of independent variables (except for intercept) times the estimated error.
So if I have k parameters(including intercept) I need k-1 AR(1) models.And for each one I pick both the coefficient and standard deviation estimated to plug in the
the formula for the calculation of alfa(1) provided by Eviews (formula D.11 page 180)
If I consider only a parameter rather than the intercept, the result for the bandwidth is ok but if there is more than one independent variable the result no longer matchs.
Note:The expressions for alfa(1) still depends on the weighting vector Ws that is considered to be Ws=1 so it cancels out.
I don´t know if there is any weighting we must consider but the formula seems very clear
To make things clear I attach 2 simple models (with one and two independent variables) only for help purpose.
If somebody can help I'll appreciate
Tks
Joao Pereira