quantile regression, different sparsity

[maxchen]

ver: EViews7Patch_060211.exe

try the following code

Code: Select all

%ex7 = "\Example Files\EV7 Manual Data\" %ch28 = "Chapter 28 - Quantile Regression" %wf = @evpath + %ex7 +%ch28 +"\engel.wf1" wfopen %wf equation eq12.qreg(cov=iid) y c x equation eq13.qreg(cov=sandwich) y c x

then eq12 and eq13 produce the same coef estimation, save Objective and Restri Objective.
And report the same method for sparsity
Sparsity method: Kernel (Epanechnikov) using residuals
Bandwidth method: Hall-Sheather, bw=0.15744
however, the value of sparsity is different, and thus produce different QLR.

[EViews Glenn]

Thanks, we'll take a look.

[EViews Glenn]

It's not documented clearly in the manual, but the Welsh method of estimating the kernel sparsity is used when you are not computing sandwich covariances, while the Powell method is used when you are computing the kernel using the sandwich covariances. The Powell is specifically designed to handle the non-iid data under the latter assumption. Both are described in the documentation, though we are somewhat vague about when each is used.

I'll see about putting additional messaging in the output to distinguish between the two cases and for adding to the manual description.

Thanks for pointing this out.

[maxchen]

Thanks for the replying.
one more question, when there is inid, each i has a sparsity, what is the sparsity reported in the output of qreg then?

[EViews Glenn]

As I said, it's the Powell estimator; it's described in technical documentation. You should think of it as an analogue of the White estimator.