cottrell(a)wfu.edu @ INTERNET skrev 2008-06-13 17:51:24 :
On Fri, 13 Jun 2008, andreas.rosenblad(a)ltv.se wrote:
> cottrell(a)wfu.edu @ INTERNET skrev 2008-06-11 15:55:14 :
> > > Unless anyone has any show-stopping problems to report, I plan to
> > release gretl 1.7.5 tomorrow, after doing an OS X build tonight.
> > Allin.
> > I think it looks great. Some comments:
> > 1. The normtest command should be accessible also from the GUI.
> E.g. Variable > Normality test
Yes, that is planned, but I didn't want to introduce yet another new
string for translation at this point.
> 2. The quantile regression function: When the confidence >
intervals cannot be computed (I get a "Not a number in >
Could you send an example where you get the "Not a number" message?
It may be there's nothing we can do about that, but it might be
Open the Ramanathan 4-1 data set. Set price as dependent and sgft, bedrms,
baths as independent. Set desired quantiles to 0.25 0.50 0.75. Choose
"compute confidence intervals". Press OK: "Not a Number in
> and I instead let gretl compute the standard errors for a tau
sequence, the graph with confidence intervals for the tau > sequence
is still available, giving weird CIs.
It's intended that the tau graph is still available -- the idea is
that it will be based on +/- so many standard errors. I've
experimented with this and got OK results. Could you post an
example where the CIs are "weird"?
Take the same example as above but choose "compute standard errors"
instead. The graph of tau sequence for e.g. baths looks weird (very short
CI for 0.25 compared to 0.50 and 0.75). But maybe it isn't wrong, since I
see that the standard errors are much smaller for tau=0.25 than for 0.50 or
> Further, for the quantile regression function the Gretl User?s
Guide says that for large data sets, "If you want confidence >
intervals in such cases, you are advised not to use the > --
intervals option, but to compute them using the method of > ?plus or
minus so many standard errors?." It would be great if > there was an
option to automatically compute and print the CI > calculated with
this option. E.g. a new radio button in the GUI: > "Compute
confidence intervals using normality approximation".
OK, that's a possibility for the future. If you select just one tau
value you can get such intervals by the usual means (Model window,
Analysis, Confidence intervals), but I guess you're talking about
the multiple-tau case.
I just thought that if the computations fail (got "Not a number" message)
or it takes too long time to compute the CIs, one could choose the
asymptotic method for the multiple-tau case.