Dear SPSS Users, I am conducting moderation analysis with binary moderator (M) and continuous observed (not latent) DV and IV. In order to address a multicollinearity I use mean centering on the IV, then I I used hierarchical regression with the following steps: 1. The IV was mean centered ( I called it IVcentered). 2. The IVCentered was multiplied with the M to get the product term (IVCentered*M) 3. A three step Hierarchical Linear Regression was conducted. 3.1 STEP1: Regression model of DV on IVcentered 3.2 STEP2: Regression model of DV on IVCentered + M 3.3 STEP3: Regression model of DV on IVCentered + M + IVCentered*M The problem arises was that the VIF are very high on IVcentered and IVCentered*M. These VIF are greater than 13. Can you suggest a solution on the multicollinearity? Or alternative approach? Thank you. Eins |
I expect that you have the binary moderator (M) coded as 0/1, which
=====================
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
would give the greatest size for the VIF; coding as 1/2 will be almost as bad. Is VIF okay if you code M as 1/ -1 ? If not, then you have a strongly skewed moderator (mostly one value), and you are probably stuck with those results. -- Rich Ulrich Date: Sun, 14 Dec 2014 04:43:01 +0000 From: [hidden email] Subject: Multicollinearity in Moderation Analysis with binary Moderator To: [hidden email] Dear SPSS Users, I am conducting moderation analysis with binary moderator (M) and continuous observed (not latent) DV and IV. In order to address a multicollinearity I use mean centering on the IV, then I I used hierarchical regression with the following steps: 1. The IV was mean centered ( I called it IVcentered). 2. The IVCentered was multiplied with the M to get the product term (IVCentered*M) 3. A three step Hierarchical Linear Regression was conducted. 3.1 STEP1: Regression model of DV on IVcentered 3.2 STEP2: Regression model of DV on IVCentered + M 3.3 STEP3: Regression model of DV on IVCentered + M + IVCentered*M The problem arises was that the VIF are very high on IVcentered and IVCentered*M. These VIF are greater than 13. Can you suggest a solution on the multicollinearity? Or alternative approach? Thank you. Eins |
Dear Rich, You are right that using 1/2 code is bad. The VIF reduces (down to its acceptable level) when I changed the codes from 1/2 to 0/1. So, coding in binary moderator case matters a lot. Thank you for your suggestion. I did not use anymore the +1/-1 codes. Eins On Sunday, December 14, 2014 3:12 PM, Rich Ulrich <[hidden email]> wrote: I expect that you have the binary moderator (M) coded as 0/1, which would give the greatest size for the VIF; coding as 1/2 will be almost as bad. Is VIF okay if you code M as 1/ -1 ? If not, then you have a strongly skewed moderator (mostly one value), and you are probably stuck with those results. -- Rich Ulrich Date: Sun, 14 Dec 2014 04:43:01 +0000 From: [hidden email] Subject: Multicollinearity in Moderation Analysis with binary Moderator To: [hidden email] Dear SPSS Users, I am conducting moderation analysis with binary moderator (M) and continuous observed (not latent) DV and IV. In order to address a multicollinearity I use mean centering on the IV, then I I used hierarchical regression with the following steps: 1. The IV was mean centered ( I called it IVcentered). 2. The IVCentered was multiplied with the M to get the product term (IVCentered*M) 3. A three step Hierarchical Linear Regression was conducted. 3.1 STEP1: Regression model of DV on IVcentered 3.2 STEP2: Regression model of DV on IVCentered + M 3.3 STEP3: Regression model of DV on IVCentered + M + IVCentered*M The problem arises was that the VIF are very high on IVcentered and IVCentered*M. These VIF are greater than 13. Can you suggest a solution on the multicollinearity? Or alternative approach? Thank you. Eins =====================
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 |
Coding with binary predictors especially matters when you compute
=====================
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
your own interactions: And you still have not paid enough attention to it. Potentially (and what I stated), using 0/1 is worse than using 1/2: Please do pay attention to the values that you can get when you do the computation by hand. And consider what happens when 0/1 is skewed, rather than being near 50%. Using 0/1 has to drop information: The other values are multiplied by zero, after all. If you have a small number of zero, then the interaction term computed this way has identical to the other variable, except for the few cases that are set to zero. On the other extreme -- which you may have, since it seemed an improvement -- a large number of zeros means that you are only looking at a range of non-zero scores in the interaction for the few people who were scored "1". This score will have a low VIF mainly because it contains such a small amount of information. The VIF between an interaction and its two component variables will go to zero when you use mean-adjusted-to-zero scores for both variables. That is, if 90% of the sores are scored ones, average for 0 and 1 of 0.9, you might score those as "+0.1" and "-0.9" by subtracting off the mean. Looking at the actual scores is also the way to help consider effects of other transformations, including (especially) rank-order transformations when most scores are ties. -- Rich Ulrich Date: Sun, 14 Dec 2014 15:59:43 +0000 From: [hidden email] Subject: Re: Multicollinearity in Moderation Analysis with binary Moderator To: [hidden email] Dear Rich, You are right that using 1/2 code is bad. The VIF reduces (down to its acceptable level) when I changed the codes from 1/2 to 0/1. So, coding in binary moderator case matters a lot. Thank you for your suggestion. I did not use anymore the +1/-1 codes. Eins On Sunday, December 14, 2014 3:12 PM, Rich Ulrich <[hidden email]> wrote: I expect that you have the binary moderator (M) coded as 0/1, which would give the greatest size for the VIF; coding as 1/2 will be almost as bad. Is VIF okay if you code M as 1/ -1 ? If not, then you have a strongly skewed moderator (mostly one value), and you are probably stuck with those results. -- Rich Ulrich Date: Sun, 14 Dec 2014 04:43:01 +0000 From: [hidden email] Subject: Multicollinearity in Moderation Analysis with binary Moderator To: [hidden email] Dear SPSS Users, I am conducting moderation analysis with binary moderator (M) and continuous observed (not latent) DV and IV. In order to address a multicollinearity I use mean centering on the IV, then I I used hierarchical regression with the following steps: 1. The IV was mean centered ( I called it IVcentered). 2. The IVCentered was multiplied with the M to get the product term (IVCentered*M) 3. A three step Hierarchical Linear Regression was conducted. 3.1 STEP1: Regression model of DV on IVcentered 3.2 STEP2: Regression model of DV on IVCentered + M 3.3 STEP3: Regression model of DV on IVCentered + M + IVCentered*M The problem arises was that the VIF are very high on IVcentered and IVCentered*M. These VIF are greater than 13. Can you suggest a solution on the multicollinearity? Or alternative approach? Thank you. Eins =====================
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 |
My moderator has equal respondents. On Monday, December 15, 2014 6:13 AM, Rich Ulrich <[hidden email]> wrote: Coding with binary predictors especially matters when you compute your own interactions: And you still have not paid enough attention to it. Potentially (and what I stated), using 0/1 is worse than using 1/2: Please do pay attention to the values that you can get when you do the computation by hand. And consider what happens when 0/1 is skewed, rather than being near 50%. Using 0/1 has to drop information: The other values are multiplied by zero, after all. If you have a small number of zero, then the interaction term computed this way has identical to the other variable, except for the few cases that are set to zero. On the other extreme -- which you may have, since it seemed an improvement -- a large number of zeros means that you are only looking at a range of non-zero scores in the interaction for the few people who were scored "1". This score will have a low VIF mainly because it contains such a small amount of information. The VIF between an interaction and its two component variables will go to zero when you use mean-adjusted-to-zero scores for both variables. That is, if 90% of the sores are scored ones, average for 0 and 1 of 0.9, you might score those as "+0.1" and "-0.9" by subtracting off the mean. Looking at the actual scores is also the way to help consider effects of other transformations, including (especially) rank-order transformations when most scores are ties. -- Rich Ulrich Date: Sun, 14 Dec 2014 15:59:43 +0000 From: [hidden email] Subject: Re: Multicollinearity in Moderation Analysis with binary Moderator To: [hidden email] Dear Rich, You are right that using 1/2 code is bad. The VIF reduces (down to its acceptable level) when I changed the codes from 1/2 to 0/1. So, coding in binary moderator case matters a lot. Thank you for your suggestion. I did not use anymore the +1/-1 codes. Eins On Sunday, December 14, 2014 3:12 PM, Rich Ulrich <[hidden email]> wrote: I expect that you have the binary moderator (M) coded as 0/1, which would give the greatest size for the VIF; coding as 1/2 will be almost as bad. Is VIF okay if you code M as 1/ -1 ? If not, then you have a strongly skewed moderator (mostly one value), and you are probably stuck with those results. -- Rich Ulrich Date: Sun, 14 Dec 2014 04:43:01 +0000 From: [hidden email] Subject: Multicollinearity in Moderation Analysis with binary Moderator To: [hidden email] Dear SPSS Users, I am conducting moderation analysis with binary moderator (M) and continuous observed (not latent) DV and IV. In order to address a multicollinearity I use mean centering on the IV, then I I used hierarchical regression with the following steps: 1. The IV was mean centered ( I called it IVcentered). 2. The IVCentered was multiplied with the M to get the product term (IVCentered*M) 3. A three step Hierarchical Linear Regression was conducted. 3.1 STEP1: Regression model of DV on IVcentered 3.2 STEP2: Regression model of DV on IVCentered + M 3.3 STEP3: Regression model of DV on IVCentered + M + IVCentered*M The problem arises was that the VIF are very high on IVcentered and IVCentered*M. These VIF are greater than 13. Can you suggest a solution on the multicollinearity? Or alternative approach? Thank you. Eins =====================
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 |
Administrator
|
In reply to this post by E. Bernardo
Take a look at this:
http://www.statisticalhorizons.com/multicollinearity
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
In reply to this post by E. Bernardo
"equal respondents" ?
what is the construct behind the moderator? Does the moderator represent a random assignment to treatment? By any chance is the moderator a coarsening of a variable via a median split?
Art Kendall
Social Research Consultants |
Art Do you know: has Marta stopped contributing to the list? Regards, Martin Martin P. Holt [hidden email] Persistence and Determination Alone are Omnipotent ! If you can't explain it simply, you don't understand it well enough.....Einstein Linked In: https://www.linkedin.com/profile/edit?trk=nav_responsive_sub_nav_edit_profile
|
Administrator
|
Word has it that Marta has been seduced/abducted by the Dark Side (Stata).
Something regarding abominable SPSS licensing fees at her institution from some time ago ;-( She does pop in now and then (when she feels her ears burning). I suspect she may even pop in today sometime ;-) ---
Please reply to the list and not to my personal email.
Those desiring my consulting or training services please feel free to email me. --- "Nolite dare sanctum canibus neque mittatis margaritas vestras ante porcos ne forte conculcent eas pedibus suis." Cum es damnatorum possederunt porcos iens ut salire off sanguinum cliff in abyssum?" |
Free forum by Nabble | Edit this page |