Need help with Standardized Beta Coefficients...

[Robyn Berkley]

I am running a regression analysis with a continuous, scale score dependent variable (Perceptions of Sexual Harassment), 2 independent variables that are dichotomous/categorical ("Amount of Harassment (low or high)" & "Email or Face to Face"), and 1 IV that is continuous (Sex Role Identity).  As I enter the IVs in the second step (after the CVs), the results are fine and within acceptable statistical ranges...and significant which is nice.  In the third step, when I enter the interaction variables, one of the dichotomous variables, which has had a strong effect all along, enters with a standardized beta coefficient that is OVER 1.00 (1.03 to be specific).  It SHOULDN'T be doing that as far as I've been taught but I don't know what to do to diagnose the problem.  The variable had been re-coded from 3 categorical variables (A, B, and C) to just A and B/C which were collapsed because B & C were not perceived differently.  It is an unbalanced design and when I do a univariate
 ANOVA, my results are still significant to the .001 level of analysis.  I want to be able to specify the model which is why I am running the regression.

  A colleague suggested there may be collinearity going on when I enter the interaction terms and he suggested I "center" the variable that is causing problems to avoid that.  I'm curious if there are any other ways to handle the situation.  Please advise!

  Thanks much for your attention and help.

  Robyn


****************************************************************
...Peace...it does not mean to be in a place where there is no noise, trouble or hard work, it means to be in the midst of those things and still be calm in your heart...(author unknown)

Robyn A. Berkley, Ph.D.
Southern Illinois University Edwardsville
****************************************************************
 __________________________________________________
Do You Yahoo!?
Tired of spam?  Yahoo! Mail has the best spam protection around
http://mail.yahoo.com

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[David Greenberg]

It is not impossible for standardized coefficients to be larger than 1 in magnitude, but it doesn't happen too often because most social-science effects are not that big. A common reason for them is multi-collinearity. A common source of multi-collinearity is a strong relationship between product terms and the factors that make up the product. You should check for that and take appropriate steps, such as centering the variables before multiplying them to create your interaction term. David Greenberg, Sociology Department, New York University

----- Original Message -----
From: Robyn Berkley <[hidden email]>
Date: Wednesday, October 31, 2007 3:59 pm
Subject: Need help with Standardized Beta Coefficients...
To: [hidden email]

> I am running a regression analysis with a continuous, scale score
> dependent variable (Perceptions of Sexual Harassment), 2 independent
> variables that are dichotomous/categorical ("Amount of Harassment (low
> or high)" & "Email or Face to Face"), and 1 IV that is continuous (Sex
> Role Identity).  As I enter the IVs in the second step (after the
> CVs), the results are fine and within acceptable statistical
> ranges...and significant which is nice.  In the third step, when I
> enter the interaction variables, one of the dichotomous variables,
> which has had a strong effect all along, enters with a standardized
> beta coefficient that is OVER 1.00 (1.03 to be specific).  It
> SHOULDN'T be doing that as far as I've been taught but I don't know
> what to do to diagnose the problem.  The variable had been re-coded
> from 3 categorical variables (A, B, and C) to just A and B/C which
> were collapsed because B & C were not perceived differently.  It is an
> unbalanced design and when I do a univariate
>  ANOVA, my results are still significant to the .001 level of
> analysis.  I want to be able to specify the model which is why I am
> running the regression.
>
>   A colleague suggested there may be collinearity going on when I
> enter the interaction terms and he suggested I "center" the variable
> that is causing problems to avoid that.  I'm curious if there are any
> other ways to handle the situation.  Please advise!
>
>   Thanks much for your attention and help.
>
>   Robyn
>
>
> ****************************************************************
> ...Peace...it does not mean to be in a place where there is no noise,
> trouble or hard work, it means to be in the midst of those things and
> still be calm in your heart...(author unknown)
>
> Robyn A. Berkley, Ph.D.
> Southern Illinois University Edwardsville
> ****************************************************************
>  __________________________________________________
> Do You Yahoo!?
> Tired of spam?  Yahoo! Mail has the best spam protection around
> http://mail.yahoo.com
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Robyn Berkley]

Thanks David...and in fact that is what a colleague here suggested and it fixed the problem.  Glad to have triangulation with your recommendation as it gives me confidence the fix is statistically acceptable!

  I really appreciate you weighing in on the situation.  So glad to have helpful colleagues!

  Best Regards...Robyn Berkley

David Greenberg <[hidden email]> wrote:
  It is not impossible for standardized coefficients to be larger than 1 in magnitude, but it doesn't happen too often because most social-science effects are not that big. A common reason for them is multi-collinearity. A common source of multi-collinearity is a strong relationship between product terms and the factors that make up the product. You should check for that and take appropriate steps, such as centering the variables before multiplying them to create your interaction term. David Greenberg, Sociology Department, New York University

----- Original Message -----
From: Robyn Berkley
Date: Wednesday, October 31, 2007 3:59 pm
Subject: Need help with Standardized Beta Coefficients...
To: [hidden email]

> I am running a regression analysis with a continuous, scale score
> dependent variable (Perceptions of Sexual Harassment), 2 independent
> variables that are dichotomous/categorical ("Amount of Harassment (low
> or high)" & "Email or Face to Face"), and 1 IV that is continuous (Sex
> Role Identity). As I enter the IVs in the second step (after the
> CVs), the results are fine and within acceptable statistical
> ranges...and significant which is nice. In the third step, when I
> enter the interaction variables, one of the dichotomous variables,
> which has had a strong effect all along, enters with a standardized
> beta coefficient that is OVER 1.00 (1.03 to be specific). It
> SHOULDN'T be doing that as far as I've been taught but I don't know
> what to do to diagnose the problem. The variable had been re-coded
> from 3 categorical variables (A, B, and C) to just A and B/C which
> were collapsed because B & C were not perceived differently. It is an
> unbalanced design and when I do a univariate
> ANOVA, my results are still significant to the .001 level of
> analysis. I want to be able to specify the model which is why I am
> running the regression.
>
> A colleague suggested there may be collinearity going on when I
> enter the interaction terms and he suggested I "center" the variable
> that is causing problems to avoid that. I'm curious if there are any
> other ways to handle the situation. Please advise!
>
> Thanks much for your attention and help.
>
> Robyn
>
>
> ****************************************************************
> ...Peace...it does not mean to be in a place where there is no noise,
> trouble or hard work, it means to be in the midst of those things and
> still be calm in your heart...(author unknown)
>
> Robyn A. Berkley, Ph.D.
> Southern Illinois University Edwardsville
> ****************************************************************
> __________________________________________________
> Do You Yahoo!?
> Tired of spam? Yahoo! Mail has the best spam protection around
> http://mail.yahoo.com
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD

****************************************************************
...Peace...it does not mean to be in a place where there is no noise, trouble or hard work, it means to be in the midst of those things and still be calm in your heart...(author unknown)

Robyn A. Berkley, Ph.D.
Southern Illinois University Edwardsville
****************************************************************
 __________________________________________________
Do You Yahoo!?
Tired of spam?  Yahoo! Mail has the best spam protection around
http://mail.yahoo.com

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD