Re: compare datasubsets linear regression

On Dec 4, 11:28 am, msandow <msandow...@xxxxxxxx> wrote:
On 30 Nov., 19:08, Ray Koopman <koop...@xxxxxx> wrote:



On Nov 30, 7:57 am, msandow <msandow...@xxxxxxxx> wrote:

On 30 Nov., 15:44, richardsta...@xxxxxxxxxxx wrote:
On Fri, 30 Nov 2007 03:14:35 -0800 (PST), msandow

<msandow...@xxxxxxxx> wrote:
Dear all!

I am seraching for a method two compare the linear fit slope of the
whole dataset and a subset of these data.
Because of theoretical reasons I would like:

1) to split the x axis 1/4 to 3/4 and want to compare the linear fit
slope of the 3/4 section to the whole dataset slope.

2) split the x- axis into ten equal classes, extract the maximum y-
value for each class and compare the resulting maximum slope to the
whole dataset (including the max. values) slope.

do you have any suggestions?

Thanks

cheers

Marcel

If you are willing to slightly redefine the problem to comparing the
first section to the second section, then (1) is easy. Define D to be
1.0 if an observation falls in the first group and zero if in the
second group. Then run the regression

y = a +b*x + c*D*x

and do a t-test on c.

Question (2) is much harder, because the max value is itself a random
variable.
-*** Startz

Thanks for the suggestion. In my case the variability in the small
subfraction is that high, that the computed regression of this
subfraction does not make any sense.

Marcel

The model should allow for a different intercept,
as well as a different slope:

y = a + b*x + c*D + d*D*X + error

The significance test of d will be the same whether
D picks out the larger group or the smaller group.

Sorry but if I understood this model correctly you mention to split
these two groups. Maybe I am wrong, but cutting of the "special" group
from the whole dataset does not make any sense in my case. Therefore
the two datasets are not independent. Maybe you can give me some
details.

Thanks

Marcel

Did you see the addendum Ray Koopman posted? Here it is again:

<quote>
Addendum: while there is nothing wrong with thinking about the
difference between a subgroup and the whole group, the test of
whether that difference is zero or not is identical to the test
of whether the subgroup differs from the remainder of the group.
</quote>

--
Bruce Weaver
bweaver@xxxxxxxxxxxx
www.angelfire.com/wv/bwhomedir
"When all else fails, RTFM."
.