9:01 a.m.
Hello.
1. Can GRETL test whether a given sample is normally distributed?
2. I haven't found a Q-Q plot test in GRETL, (like in SAS e.g.). I think it
would be a good thing to add it.
3. Is there a portable version of GRETL?
Thank you
Constantin
10:57 a.m.
New subject: Goodness of Fit test to normal distribution and QQ plot
Costia schrieb:
Hello.
1. Can GRETL test whether a given sample is normally distributed?
Yes, for example right-click on a variable, choose frequency plot (or
somesuch), select show normal distribution (instead of only data), and
in the plot there will be the test result. Or more straightforwardly,
menu Variable -> normality test.
2. I haven't found a Q-Q plot test in GRETL, (like in SAS e.g.).
I think
it would be a good thing to add it.
I agree. It's not difficult to use a quick-and-dirty script to get it
done, but properly handling missing values etc. needs more care.
3. Is there a portable version of GRETL?
I don't know -- on Windows and Mac I guess everything is bundled and so
I can imagine it just works portably (never tried it though). On linux I
doubt it, because it depends on other software. However I agree that it
would be a nice feature, for example for teaching. It could mean
a(nother) competitive advantage over Eviews and other closed-source
software.
-sven
11:25 a.m.
New subject: Goodness of Fit test to normal distribution and QQ plot
Sven Schreiber schrieb:
Costia schrieb:
> Hello.
> 1. Can GRETL test whether a given sample is normally distributed?
BTW, a related question to Allin: I just noticed the relatively new
stuff in the Tools menu, namely the CDF-plot entry, and I'm wondering
why only the standard normal and logistic functions are available and
whether that warrants a separate menu entry. Instead I would suggest to
incorporate it into the density plotting dialog by adding a tick box
like "CDF instead of density".
thanks,
sven
11:59 a.m.
Em 30/11/2009, às 09:25, Sven Schreiber <svetosch(a)gmx.net> escreveu:
Sven Schreiber schrieb:
> Costia schrieb:
>> Hello.
>> 1. Can GRETL test whether a given sample is normally distributed?
>
BTW, a related question to Allin: I just noticed the relatively new
stuff in the Tools menu, namely the CDF-plot entry, and I'm wondering
why only the standard normal and logistic functions are available and
whether that warrants a separate menu entry. Instead I would suggest
to
incorporate it into the density plotting dialog by adding a tick box
like "CDF instead of density".
I think Sven's suggestions are very good!
Best,
Henrique
Enviado de meu celular
3:34 p.m.
On Mon, 30 Nov 2009, Sven Schreiber wrote:
BTW, a related question to Allin: I just noticed the relatively new
stuff in the Tools menu, namely the CDF-plot entry, and I'm wondering
why only the standard normal and logistic functions are available and
whether that warrants a separate menu entry. Instead I would suggest to
incorporate it into the density plotting dialog by adding a tick box
like "CDF instead of density".
Agreed, it doesn't really warrant a separate entry; I've taken
your suggestion and added a checkbox for the Normal tab.
I'm not in a hurry to add more CDFs. The normal and logistic are
there because they're pedagogically useful when talking about
logit and probit. I don't see so much point in adding others; if
anyone wants to see all the CDFs known to man, run the gnuplot
demo file "prob.dem".
Allin.
2:53 a.m.
On Mon, 30 Nov 2009, Sven Schreiber wrote:
Costia schrieb:
> 2. I haven't found a Q-Q plot test in GRETL, (like in SAS e.g.). I think
> it would be a good thing to add it.
I agree. It's not difficult to use a quick-and-dirty script to get it
done, but properly handling missing values etc. needs more care.
I haven't messed much with Q-Q plots. Can anyone confirm if the
following script is correct? If so, then I guess we could do it
as a built-in at some point.
<script>
nulldata 500
scalar n = $nobs
# artificial series to test
genr y = sort(normal())
series Qemp, Qnorm
scalar p, N, nl, nh
loop i=1..n -q
p = i/(n+1)
N = (n + 1) * p - 1;
nl = floor(N);
nh = ceil(N);
if (nh == 0 || nh == n)
Qemp[i] = NA
else
Qemp[i] = quantile(y, p)
endif
Qnorm[i] = critical(N, 1-p)
endloop
store @dotdir/gpdata.dat Qemp Qnorm --traditional
set echo off
outfile @dotdir/qq.gp --write
print "set datafile missing '-999'"
print "set nokey"
print "set xlabel 'Normal quantiles'"
print "set ylabel 'Empirical quantiles'"
print "plot '(a)dotdir/gpdata.dat', x"
outfile --close
if WIN32
launch @gnuplot @dotdir/qq.gp
else
launch @gnuplot -persist @dotdir/qq.gp
endif
</script>
Allin.
11:22 p.m.
Allin Cottrell schrieb:
On Mon, 30 Nov 2009, Sven Schreiber wrote:
> Costia schrieb:
>> 2. I haven't found a Q-Q plot test in GRETL, (like in SAS e.g.). I think
>> it would be a good thing to add it.
> I agree. It's not difficult to use a quick-and-dirty script to get it
> done, but properly handling missing values etc. needs more care.
I haven't messed much with Q-Q plots. Can anyone confirm if the
following script is correct? If so, then I guess we could do it
as a built-in at some point.
sorry for not answering earlier...
<script>
nulldata 500
scalar n = $nobs
# artificial series to test
genr y = sort(normal())
series Qemp, Qnorm
scalar p, N, nl, nh
loop i=1..n -q
p = i/(n+1)
N = (n + 1) * p - 1;
given your def. of p, wouldn't that be equiv. to N = i-1 ?
nl = floor(N);
accordingly redundant? (nl = i-1)
nh = ceil(N);
ditto?
if (nh == 0 || nh == n)
Qemp[i] = NA
don't understand the reason for this: here it seems the first item in
the empirical quantiles will always be set to NA?
else
Qemp[i] = quantile(y, p)
endif
Qnorm[i] = critical(N, 1-p)
or rather Qnorm[i] = 1-critical(N,p)? (could make a difference for
discrete distributions, not for continuous as here)
endloop
and the 45-degree line is needed for comparison
altogether I think something like this also would do the trick (bugs
probably included, even if it's only two lines...):
matrix mp = transp(seq(1,n))/n
# (col vector of cumulative relative frequencies for n obs)
matrix mQtheory = invcdf(N,mp)
# BTW, the help about invcdf() says P(X<x), but shouldn't it be P(X<=x)?
and your y (the sorted input series) can already be plotted against
mQtheory to give the QQ plot (if I'm not missing something right now
which may well be the case since I'm a bit tired). You might argue that
this doesn't really deal with ties, but from the various possible
definitions I've seen of quantiles per se and how to deal with ties I
think it's actually legitimate.
Why not just copy the convention and maybe the code from R?
good night,
sven
5254
days inactive
5260
days old
6 comments
4 participants
participants (4)
-
Allin Cottrell -
Carlos Henrique Co?lho de Andrade -
Costia -
Sven Schreiber