Re: [Gretl-users] Mean compare
On Fri, 30 May 2014, Frank Benford wrote:
I have a question about gretl's procedure here. Why is the
p-value
calculated "in relation to the normal distribution"? Shouldn't it be
calculated in relation to a t random variable with number of degrees of
freedom computed by the Smith-Satterthwaite formula? Of course, if n1 and
n2 are even moderately large, this refinement makes no difference as the t
random variable is effectively a standard normal random variable. Could
this be the explanation for why R and gretl return different p-values in
Rosen Iliev's case?
Probably it is the reason for the difference from R. Up till now
we've used the asymptotic test (with a warning if the sample sizes
are too small). It doesn't cost much to use the Satterthwaite
approximation to the t distribution, so maybe we'll do that. Within
reason, anything that cuts down on "Why do your results differ from
program P?" is good.
Allin Cottrell
-----Original Message-----
From: gretl-users-bounces(a)lists.wfu.edu
[mailto:gretl-users-bounces@lists.wfu.edu] On Behalf Of Allin Cottrell
Sent: Friday, May 30, 2014 12:54 PM
To: Gretl list
Subject: Re: [Gretl-users] Mean compare
On Fri, 30 May 2014, Rosen Iliev wrote:
> Hi,
> first sorry for the stupid question. I am performing datamining with R
> from one year and recently found Gretl program. It is wonderful for
> fast descriptive analysis, mean compare and graphics for novices! Can
> You tell me what type analysis is calculated from "statistic
> calculator" when checkbox "assume common population standard
> population" is not tick? The reason for asking is different "p"
> results, compared to R, when there is no thick.
If a common variance is not assumed we use the asymptotic standard error of
the difference in means, namely
sqrt(vm1 + vm2)
where vm1 = [(variance of variable 1) / n1] and similarly for vm2.
We then find the p-value for the difference divided by its standard error in
relation to the normal distribution.
Allin Cottrell