Advice on interaction between within-subjects factor and between-subjects covariate in GLM

[Hepper E.G.D.]

Dear Listserve experts,

 

I have run into problems breaking down an interaction between within-subjects factors and a between-subjects continuous predictor in GLM, and would really appreciate any advice you can offer.

 

Briefly, the study assessed reaction times to words after presenting a face stimulus.  The design has 2 within-subjects factors: type of face (3 levels: f1, f2, f3) and type of word (2 levels: w1, w2).  It was hypothesised that the effects of face and word would be moderated by a personality variable (assessed on a continuous scale) and by gender.

 

Thus, I ran SPSS GLM entering the centered personality variable as a covariate as follows:

 

GLM f1w1 f1w2 f2w1 f2w2 f3w1 f3w2 by gender WITH c.personality

/WSFACTOR=face 3 SPECIAL (1 1 1 1 -.5 -.5 0 1 -1) word 2 Simple

/PRINT=PARAMETER

/WSDESIGN=face word face*word

/DESIGN=c.personality gender c.personality*gender.

 

As predicted, there is a significant 3-way Face X Word X Personality interaction.  Now I'm stuck!  What is the best way to probe this interaction?  I'm used to testing simple slopes in regression by entering "high" and "low" scores of the continuous predictor instead of the centered variable; is this also possible in GLM?  Any suggestions will be very gratefully received.

 

(PS I also tried running the model using SPSS MIXED, but equally don’t know a way to test simple slopes for a level 1 X level 1 X level 2 interaction – if anyone does then this would also be great.)

 

Many thanks in advance,

Erica

 

_______________________

 

Erica Hepper, Ph.D.

Research & Teaching Fellow

School of Psychology

University of Southampton, UK

SO17 1BJ

[hidden email]

 

[Bruce Weaver]

Hepper E.G.D. wrote

Dear Listserve experts,

I have run into problems breaking down an interaction between within-subjects factors and a between-subjects continuous predictor in GLM, and would really appreciate any advice you can offer.

Briefly, the study assessed reaction times to words after presenting a face stimulus.  The design has 2 within-subjects factors: type of face (3 levels: f1, f2, f3) and type of word (2 levels: w1, w2).  It was hypothesised that the effects of face and word would be moderated by a personality variable (assessed on a continuous scale) and by gender.

Thus, I ran SPSS GLM entering the centered personality variable as a covariate as follows:

GLM f1w1 f1w2 f2w1 f2w2 f3w1 f3w2 by gender WITH c.personality
/WSFACTOR=face 3 SPECIAL (1 1 1 1 -.5 -.5 0 1 -1) word 2 Simple
/PRINT=PARAMETER
/WSDESIGN=face word face*word
/DESIGN=c.personality gender c.personality*gender.

As predicted, there is a significant 3-way Face X Word X Personality interaction.  Now I'm stuck!  What is the best way to probe this interaction?  I'm used to testing simple slopes in regression by entering "high" and "low" scores of the continuous predictor instead of the centered variable; is this also possible in GLM?  Any suggestions will be very gratefully received.

(PS I also tried running the model using SPSS MIXED, but equally don't know a way to test simple slopes for a level 1 X level 1 X level 2 interaction - if anyone does then this would also be great.)

Many thanks in advance,
Erica

I think you've answered your own question.  I would try running the model again with the personality variable centered on some different values.

Also, remember that on the /EMMEANS sub-command, the default is to show estimated marginal means with any covariates set to their mean values.  But you can set the covariates to whatever value you want by adding the WITH option.  Something like:

 /EMMEANS table(face*word) WITH (Personality = 5)
 /EMMEANS table(face*word) WITH (Personality = 10)
 /EMMEANS table(face*word) WITH (Personality = 15)

HTH.

[Hepper E.G.D.]

Thank you so much for getting back to me, Bruce. It's good to have some support for the plan to run with differently centered covariates, as I thought of this based on my experiences with regression but could not find it recommended anywhere for GLM. Thanks for the EMMEANS tip too - will be very helpful.

Could you clarify for me whether GLM tests the effects of other variables holding the covariate at 0? If so, I'd re-centre the personality variable so that a score of 0 represented the desired level (e.g., one SD above/below the mean). Would that work?

Thanks again for your help,
Erica


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Bruce Weaver
Sent: 17 August 2010 14:58
To: [hidden email]
Subject: Re: Advice on interaction between within-subjects factor and between-subjects covariate in GLM

Hepper E.G.D. wrote:

>
> Dear Listserve experts,
>
> I have run into problems breaking down an interaction between
> within-subjects factors and a between-subjects continuous predictor in
> GLM, and would really appreciate any advice you can offer.
>
> Briefly, the study assessed reaction times to words after presenting a
> face stimulus.  The design has 2 within-subjects factors: type of face (3
> levels: f1, f2, f3) and type of word (2 levels: w1, w2).  It was
> hypothesised that the effects of face and word would be moderated by a
> personality variable (assessed on a continuous scale) and by gender.
>
> Thus, I ran SPSS GLM entering the centered personality variable as a
> covariate as follows:
>
> GLM f1w1 f1w2 f2w1 f2w2 f3w1 f3w2 by gender WITH c.personality
> /WSFACTOR=face 3 SPECIAL (1 1 1 1 -.5 -.5 0 1 -1) word 2 Simple
> /PRINT=PARAMETER
> /WSDESIGN=face word face*word
> /DESIGN=c.personality gender c.personality*gender.
>
> As predicted, there is a significant 3-way Face X Word X Personality
> interaction.  Now I'm stuck!  What is the best way to probe this
> interaction?  I'm used to testing simple slopes in regression by entering
> "high" and "low" scores of the continuous predictor instead of the
> centered variable; is this also possible in GLM?  Any suggestions will be
> very gratefully received.
>
> (PS I also tried running the model using SPSS MIXED, but equally don't
> know a way to test simple slopes for a level 1 X level 1 X level 2
> interaction - if anyone does then this would also be great.)
>
> Many thanks in advance,
> Erica
>
>

I think you've answered your own question.  I would try running the model
again with the personality variable centered on some different values.

Also, remember that on the /EMMEANS sub-command, the default is to show
estimated marginal means with any covariates set to their mean values.  But
you can set the covariates to whatever value you want by adding the WITH
option.  Something like:

 /EMMEANS table(face*word) WITH (Personality = 5)
 /EMMEANS table(face*word) WITH (Personality = 10)
 /EMMEANS table(face*word) WITH (Personality = 15)

HTH.


-----
--
Bruce Weaver
[hidden email]
http://sites.google.com/a/lakeheadu.ca/bweaver/

"When all else fails, RTFM."

NOTE: My Hotmail account is not monitored regularly.
To send me an e-mail, please use the address shown above.

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Advice-on-interaction-between-within-subjects-factor-and-between-subjects-covariate-in-GLM-tp2637845p2638178.html
Sent from the SPSSX Discussion mailing list archive at Nabble.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

[Bruce Weaver]

Hepper E.G.D. wrote

Thank you so much for getting back to me, Bruce. It's good to have some support for the plan to run with differently centered covariates, as I thought of this based on my experiences with regression but could not find it recommended anywhere for GLM. Thanks for the EMMEANS tip too - will be very helpful.

Could you clarify for me whether GLM tests the effects of other variables holding the covariate at 0? If so, I'd re-centre the personality variable so that a score of 0 represented the desired level (e.g., one SD above/below the mean). Would that work?

Thanks again for your help,
Erica

Hi Erica.  In order to answer your question, I would have to try it and see what happens.  So you could do that yourself and report back to the list.  You might also find this letter to the editor useful:

   http://journals.cambridge.org/action/displayAbstract?fromPage=online&aid=1296348

If you don't have access to the journal, write me at the e-mail address shown in my signature file, and I'll send you a copy.

HTH.

[Hepper E.G.D.]

For anyone who is interested, my GLM analyses did work while adjusting covariates so that a score of zero represented either a high or low level of the variable, as recommended for multiple regression by Aiken and West (1991). Bruce, thanks for your advice.

Interestingly, though, there were slight differences between the estimated means and simple comparisons provided by:
(a) entering the high/low adjusted covariate and requesting EMMEANS with adjusted-covariate=mean, vs.
(b) entering the centered covariate and requesting EMMEANS with covariate=high/low.

If you have thoughts on which of the above is more valid, I'd be interested to hear them.
Best wishes,
Erica


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Bruce Weaver
Sent: 18 August 2010 12:41
To: [hidden email]
Subject: Re: Advice on interaction between within-subjects factor and between-subjects covariate in GLM

Hepper E.G.D. wrote:

>
> Thank you so much for getting back to me, Bruce. It's good to have some
> support for the plan to run with differently centered covariates, as I
> thought of this based on my experiences with regression but could not find
> it recommended anywhere for GLM. Thanks for the EMMEANS tip too - will be
> very helpful.
>
> Could you clarify for me whether GLM tests the effects of other variables
> holding the covariate at 0? If so, I'd re-centre the personality variable
> so that a score of 0 represented the desired level (e.g., one SD
> above/below the mean). Would that work?
>
> Thanks again for your help,
> Erica
>
>

Hi Erica.  In order to answer your question, I would have to try it and see
what happens.  So you could do that yourself and report back to the list.
You might also find this letter to the editor useful:


http://journals.cambridge.org/action/displayAbstract?fromPage=online&aid=1296348

If you don't have access to the journal, write me at the e-mail address
shown in my signature file, and I'll send you a copy.

HTH.


-----
--
Bruce Weaver
[hidden email]
http://sites.google.com/a/lakeheadu.ca/bweaver/

"When all else fails, RTFM."

NOTE: My Hotmail account is not monitored regularly.
To send me an e-mail, please use the address shown above.

-----------------------

I think you've answered your own question.  I would try running the model again with the personality variable centered on some different values.

Also, remember that on the /EMMEANS sub-command, the default is to show estimated marginal means with any covariates set to their mean values.  But you can set the covariates to whatever value you want by adding the WITH option.  Something like:

 /EMMEANS table(face*word) WITH (Personality = 5)  /EMMEANS table(face*word) WITH (Personality = 10)  /EMMEANS table(face*word) WITH (Personality = 15)

HTH.

Bruce Weaver

> On 17 August 2010 09:33, Hepper E.G.D. <[hidden email]> wrote:
>> Dear Listserve experts,
>>
>> I have run into problems breaking down an interaction between
>> within-subjects factors and a between-subjects continuous predictor
>> in GLM, and would really appreciate any advice you can offer.
>>
>> Briefly, the study assessed reaction times to words after presenting
>> a face stimulus.  The design has 2 within-subjects factors: type of face (3 levels:
>> f1, f2, f3) and type of word (2 levels: w1, w2).  It was hypothesised
>> that the effects of face and word would be moderated by a personality
>> variable (assessed on a continuous scale) and by gender.
>>
>> Thus, I ran SPSS GLM entering the centered personality variable as a
>> covariate as follows:
>>
>> GLM f1w1 f1w2 f2w1 f2w2 f3w1 f3w2 by gender WITH c.personality
>> /WSFACTOR=face 3 SPECIAL (1 1 1 1 -.5 -.5 0 1 -1) word 2 Simple
>> /PRINT=PARAMETER
>> /WSDESIGN=face word face*word
>> /DESIGN=c.personality gender c.personality*gender.
>>
>> As predicted, there is a significant 3-way Face X Word X Personality
>> interaction.  Now I'm stuck!  What is the best way to probe this
>> interaction?  I'm used to testing simple slopes in regression by
>> entering "high" and "low" scores of the continuous predictor instead
>> of the centered variable; is this also possible in GLM?  Any
>> suggestions will be very gratefully received.
>>
>> (PS I also tried running the model using SPSS MIXED, but equally
>> don't know a way to test simple slopes for a level 1 X level 1 X
>> level 2 interaction - if anyone does then this would also be great.)
>>
>> Many thanks in advance,
>> Erica

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Advice-on-interaction-between-within-subjects-factor-and-between-subjects-covariate-in-GLM-tp2637845p2639247.html
Sent from the SPSSX Discussion mailing list archive at Nabble.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

[Bruce Weaver]

Hepper E.G.D. wrote

For anyone who is interested, my GLM analyses did work while adjusting covariates so that a score of zero represented either a high or low level of the variable, as recommended for multiple regression by Aiken and West (1991). Bruce, thanks for your advice.

Interestingly, though, there were slight differences between the estimated means and simple comparisons provided by:
(a) entering the high/low adjusted covariate and requesting EMMEANS with adjusted-covariate=mean, vs.
(b) entering the centered covariate and requesting EMMEANS with covariate=high/low.

If you have thoughts on which of the above is more valid, I'd be interested to hear them.
Best wishes,
Erica

Hi Erica.  Good to hear it worked out reasonably well.  How slight were the differences?  Can you post the data?