Date: Fri, 16 May 2003 12:50:07 +1000
Reply-To: paulandpen@optusnet.com.au
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: Paul Dickson <paulandpen@optusnet.com.au>
Subject: Re: Help is urgently needed and greatly appreciated!!
Content-Type: text/plain
Hi A B,
My understanding of a suppressor variable is that it only works in the the presence
of other IVs within a model. It distorts/suppresses the effects of the other IV's in
the model. The sign (+ or -) of the beta weight for the suppressor variable is
usually the oppostive of other IV if it is actually suppressing these IVs in the
model. Run your analysis with the other IV's and exclude your hypothesised
suppressor variable and then rerun it with the same IV's and include the
suppressor variable to see the effect of it being included in the model and then
being excluded from the model. The question then becomes "Do other beta
weights in the model without the suppressor variable change significance in the
presence of the suppressor variable?"
At first glance it looks like the effects of your predictor variable in the multivariate
model (other IV's included) and the less complicated bivariate model (no other IVs
except the "suppressor variable" included) are clearly different. Given the
correlations at the initial stages of your analysis, it is not surprising that the
bivariate regression with the "suppressor variable only" included in the model is not
significant. The surprise for me is the beta weight of the "suppressor variable"
becomes significant in the more complicated model.
The positive Beta Weight in the more complex model for your variable is very weird
and may suggest that in the presense of another IV or group of IV's that is has a
different type of effect on the DV. Check that the beta weight derived from the "so
called suppressor variable" in the complex model with more IVS is not actually the
result of a mediation/moderation effect. If you want to know how to do this email
me again.
Also check that the relationship in the more complex model is actually linear, and
not some other relationship. You can do this very easily using a scatterplot of the
zpred vs zresiduals and examining these to see if there is a linear effect across
the variables. You can also check this assumption for each variable individually
using single bivariate scatterplots of the z predictors and z residuals
Regards Paul
> A B <beraidi1@yahoo.com> wrote:
>
> Hello,
>
> Sorry, I sent a message but returned unanswered. Thus, I need to
> resend it again. I might send a message in the past but this one
> includes new dimensions of the same problem. Therefore, please
> proceed it and reply to me.
>
> I’ve got a hypothesis briefly saying that:
>
> Pressure (ID) was negatively associated with creativity (DV).
>
> Note: both variables are scales (5 point-likert scale)
>
> I’ve got the following results:
>
> 1- correlation analysis showed non-significant correlation
> (-.01).
>
> 2- regression analysis was run with the inclusion of other IV
> showed
> that pressure was a positive predictor of creativity (beta was +.05).
>
> 3- after I rerun the regression analysis with the inclusion of
> only
> the IV of pressure, the result showed that pressure was not
> significant
> predictor (beta was –.01).
>
> I have heard about what is called the suppressor variable but I don’t
> know how I could explain the result for the purpose of testing my
> hypothesis. Also, I'm not sure if it is correct to rerun the
> regression analysis to check this finding in such a case.
> Any explanation is most appreciated.
>
> Regards,
>
>
>
> ---------------------------------
> Do you Yahoo!?
> The New Yahoo! Search - Faster. Easier. Bingo.
