The RATS Software Forum

Hi all,

I have used the following simple regression with NW robust errors:

Linreg(robust, lags=1,lwindow=newey) y
# constant $
x{ 1} $
dummy1{ 0} $
dummy2{ 0}

where dummy1 and dummy2 are time series dummies. It is 1 for a particular period and 0 for others.

I can estimate the model with the robust NW standard errors presented without any problem using RATS.

However, when I use SAS to reproduce this simple model, SAS program result shows some funny standard error numbers and indicates that the t-test and p-values for all parameters are all biased (and can't be estimated). I have checked the SAS program setting, and it is correct. The interesting thing is that, if I remove the two dummies, the SAS program can estimate the NW errors, but not with two dummies included in the model.

This is a bit odd to me. I was wondering who knows where the problem is. Is my RATS program a problem, and the estimated NW errors using RATS are biased, but I was not warned, or something else?

Thank you very much for your feedback

Are they single observation dummies? Then the HAC standard errors on the dummies aren't valid. (The mcov(x,u) matrix is singular because the residuals are zero at the location of the dummies, and the x's are zero elsewhere). The standard errors on the other two variables are fine so I don' know why they would look odd. You would need to check with SAS about that.

Hi Tom

Thank you very much for your confirmation.

The SAS program uses Nonlinear GMM estimation procedure to estimate HAC errors. I guess that maybe the kernel based SAS optimization procedure is coded in such a way that is suited for nonlinear models. For my simple model with dummies, the program can estimate variable parameters alright, but produce 0 or extremely large standard errors for these parameters.

I was wondering how RATS can produce both valid and invalid HAC S.Es at the same time, with invalid S.Es for one period dummies (it is OK, as these variables are not usually the focus of testing), but valid ones for other variables in the model (this is important as people usually want to do some hypothesis testing)? Wouldn't the co-variance matrix and its adjustments be estimated jointly?

Many thanks

anozman wrote:Hi Tom

Thank you very much for your confirmation.

The SAS program uses Nonlinear GMM estimation procedure to estimate HAC errors. I guess that maybe the kernel based SAS optimization procedure is coded in such a way that is suited for nonlinear models. For my simple model with dummies, the program can estimate variable parameters alright, but produce 0 or extremely large standard errors for these parameters.

There shouldn't be any difference due to non-linear vs linear.

anozman wrote: I was wondering how RATS can produce both valid and invalid HAC S.Es at the same time, with invalid S.Es for one period dummies (it is OK, as these variables are not usually the focus of testing), but valid ones for other variables in the model (this is important as people usually want to do some hypothesis testing)? Wouldn't the co-variance matrix and its adjustments be estimated jointly?

The rank two matrix never gets inverted, so it doesn't contaminate the other estimates of the covariance matrix. Even with standard least squares asymptotics, the limit matrix is rank two.