Date:         Thu, 10 Feb 2000 12:20:54 -0800
Reply-To:     "Parise, Carol A." <PariseC@SUTTERHEALTH.ORG>
Sender:       "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From:         "Parise, Carol A." <PariseC@SUTTERHEALTH.ORG>
Subject:      Re: Interactions between categorical variables in logistic regres
              sion
Comments: To: Sheila Underhill <sheilau@NYAHSA.ORG>
Content-Type: text/plain

Hi,

Before making any decisions on this, try a couple of things...

1. run a*b crosstabs tables  for each level of your DV and check the cell
sizes.
2. construct the confidence intervals around the odds ratios (expB) for each
of the levels that are significant.

It's possible that the reason for the inconsistency is that the when getting
down to the interaction level, the cell sizes get too small. when this
happens you can end up with a statistically significant interaction at one
level but it's actually meaningless because of small numbers or missing data
in the cells.

carol parise



On Tuesday, February 08, 2000 6:47 AM, Sheila Underhill
[SMTP:sheilau@NYAHSA.ORG] wrote:
> Hi listers,
>
> Can anyone shed some light on this? I have several categorical independent
> variables in a logistic regression model (binary). I want to investigate
> interaction effects between a few of the independents, each of which have
5
> or 6 levels. Results show the overall interaction variable (a*b) is not
> significant, but a few of the combinations are. Am I to conclude that the
> interactions between the levels that show as significant are, or do I
> conclude that there is no interaction because the overall variable is not
> significant?
>
> Any insights will be greatly appreciated. Thanks.
>
> Sheila Underhill, M.S.
> Research Associate
> NYAHSA/FLTC
> 150 State Street, Suite 301
> Albany, NY  12207
> Tel: 518-449-2707 x148
> Fax: 518-455-8908