Date: Tue, 16 May 2006 14:07:27 -0300
Reply-To: Hector Maletta <email@example.com>
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: Hector Maletta <firstname.lastname@example.org>
Subject: Re: CHAID with small sample
Content-Type: text/plain; charset="us-ascii"
Chaid is based on chi square, and therefore the number of cases and the
number of categories involved IN EACH LEVEL is important for the
significance of results. If a simple chi square test fails with your sample,
then Chaid will give you non significant results.
De: SPSSX(r) Discussion [mailto:SPSSX-L@LISTSERV.UGA.EDU] En nombre de
Enviado el: Tuesday, May 16, 2006 1:28 PM
Asunto: Re: CHAID with small sample
When I worked with CHAID and CART a few years ago with about 100-150
cases, I ended up feeling like I was reaching ---- and we were just
looking for major predictors not a precise prediction tool. Remember
that at each split level, CHAID is attempting to determine a splitting
predictor and the groupings of that predictor's categories that have the
largest between group differences (via an F-test like with a one-way
anova). With 35 cases, you "might" find a top-level split predictor if
there are big differences between categories. Because CHAID helps
identify localized, rather than globalized, interactions via sequential
partitioning, each subgroup goes through the same process of looking at
potentially good splitting predictors. Since each subgroup will have
less cases than the original group of 35, further splits will only occur
if there are very big differences between the group categories. You
certainly can run the CHAID procedure as an exploratory tool to see what
happens (even with a higher alpha of 0.10); however, I would suggest
being careful about how you use any results. With such a small sample
you have no data for validating any resulting model. I'm not aware of
an Exact Test option for CHAID.
Hope that helps.
Carol B. Thompson
181 N. Arroyo Grande Blvd, Ste 105
Henderson, NV 89074-1624
Ph: (702) 731-5550 x 111
Fax: (702) 731-4027
From: SPSSX(r) Discussion [mailto:SPSSX-L@LISTSERV.UGA.EDU] On Behalf Of
Sent: Friday, May 12, 2006 6:57 AM
Subject: CHAID with small sample
Has anyone had experience doing CHAID with a small sample size N = about
35. Not worth it? Better to use Exact Tests? My client is reaching
for some way to milk this small bit of data. The data fits for CHAID,
as do the questions, but it's just such a small amount. Is there an
exact test option in CHAID?
Nico Peruzzi, Ph.D.