[R] error in vcovNW

[Saba Sehrish]

Hi
Thanks for the reminder.
Actually I want to analyse whether present value of variable A is Granger caused by lag values of B and test linear hypothesis "B1,B2,B3,B4,B5=0".
Therefore, to get robust standard error NeweyWest estimates are applied.
Saba 

    On Saturday, 19 December 2015, 23:26, Achim Zeileis <Achim.Zeileis at uibk.ac.at> wrote:
 

 On Sat, 19 Dec 2015, Saba Sehrish wrote:

> Thank you. The issue is resolved by scaling the data in millions.

That solves the numerical problem but the second issue (inappropriateness 
of the Newey-West estimator for an autoregressive model) persists.

> Saba
> 
> 
> On Saturday, 19 December 2015, 15:06, Achim Zeileis
> <Achim.Zeileis at uibk.ac.at> wrote:
> 
> 
> On Sat, 19 Dec 2015, Saba Sehrish via R-help wrote:
> 
> > Hi I am using NeweyWest standard errors to correct lm( ) output. For
> example:
> > lm(A~A1+A2+A3+A4+A5+B1+B2+B3+B4+B5)
> > vcovNW<-NeweyWest(lm(A~A1+A2+A3+A4+A5+B1+B2+B3+B4+B5))
> >
> > I am using package(sandwich) for NeweyWest. Now when I run this command,
> it gives following error:
> > Error in solve.default(diag(ncol(umat)) - apply(var.fit$ar, 2:3, sum))
> :system is computationally singular: reciprocal condition number =
> 7.49468e-18
> >
> > Attached herewith is data for A&B, A1,A2,A3,A4,A5,B1,B2,B3,B4,B5 are
> > simply lag variables. Can you help me removing this error please?
> 
> Without trying to replicate the error, there are at least two issues:
> 
> (1) You should scale your data to use more reasonable orders of magnitude,
> e.g., in millions. This will help avoiding numerical problems.
> 
> (2) More importantly, you should not employ HAC/Newey-West standard errors
> in autoregressive models. If you use an autoregressive specification, you
> should capture all relevant autocorrelations - and then no HAC estimator
> is necessary. Alternatively, one may treat autocorrelation as a nuisance
> parameter and not model it - but instead capture it in HAC standard
> errors. Naturally, the former strategy will typically perform better if
> the autocorrelations are more substantial.
> 
> > Saba
> 
> 
> 
>

  
	[[alternative HTML version deleted]]