Linear mixed model/ time-dependent covariate

[pomperu]

Hi,
I have a question regarding the statistical analysis of an EEG experiment.
In my dataset I have a between subject factor ‘group’ (patients, controls), a within subject factor ‘condition’ (A, B) and a covariate ‘performance’ (perf_A, perf_B, one for each of the two conditions).The dependent variable is EEG amplitude.

In order to take the condition-dependence of the covariate into account, I figured the correct way would be to conduct a linear mixed-model analysis.

My first question is whether it is legitimate to use a full factorial model, or whether including interactions between the factors ‘condition’ and ‘performance’ is incorrect, since the values of the covariate ‘performance’ are necessarily associated with either condition A or B.

My second question is which kind of post-hoc test I should use, given that I find a significant interaction between the factors ‘group’ and ‘condition’.

Any help on this issue is appreciated very much!!
Thanks, Ulrich

[Poes, Matthew Joseph]

Is your measure of performance constant between conditions, or does it reflect something different in each condition?

If the measure of performance is constant, means the same thing, then you are fine to use the interaction, in fact, it would be desirable.  This interaction is what will tell you the impact of performance in a given condition, i.e. the moderating effect of performance between condition type and eeg amplitude.

Typically condition is dummy coded, and so performance is not a main effect, but rather the effect of performance when your dummy indicator is 0, i.e. your reference group.  You would need to effect code it if you want a main effect for performance.  You actually need the interaction in order to get the effect of performance within each condition, instead of only within your referent condition.

If performance is different, you have a different set of problems and the whole model would be wrong.  You would need to come back to us with this information for further advice.

Matthew J Poes
Research Data Specialist
Center for Prevention Research and Development
University of Illinois
510 Devonshire Dr.
Champaign, IL 61820
Phone:� 217-265-4576
email: [hidden email]


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of pomperu
Sent: Tuesday, July 03, 2012 10:16 AM
To: [hidden email]
Subject: Linear mixed model/ time-dependent covariate

Hi,
I have a question regarding the statistical analysis of an EEG experiment.
In my dataset I have a between subject factor ‘group’ (patients, controls), a within subject factor ‘condition’ (A, B) and a covariate ‘performance’
(perf_A, perf_B, one for each of the two conditions).The dependent variable is EEG amplitude.

In order to take the condition-dependence of the covariate into account, I figured the correct way would be to conduct a linear mixed-model analysis.

My first question is whether it is legitimate to use a full factorial model, or whether including interactions between the factors ‘condition’ and ‘performance’ is incorrect, since the values of the covariate ‘performance’
are necessarily associated with either condition A or B.

My second question is which kind of post-hoc test I should use, given that I find a significant interaction between the factors ‘group’ and ‘condition’.

Any help on this issue is appreciated very much!!
Thanks, Ulrich


--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Linear-mixed-model-time-dependent-covariate-tp5713986.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

[pomperu]

Hi Matthew,
Thanks a lot for your quick reply!

In fact, performance IS different for each condition. One condition involves a spatial discrimination task, the other a semantic discrimination task. It might very well be that the two groups (patients and controls) perform different in the two conditions (i.e there is an interaction between group and performance).

This condition-dependence of the covariate  is the reason why I figured the linear-mixed model would be the correct method. Am I wrong? Would there be some regression based approach to this analysis?

Thanks a lot,
Ulrich

[Poes, Matthew Joseph]

I think you are a bit confused on what I meant by difference in the performance measure.  I didn't mean that there could be a difference in the level, that is the precise reason to test the variable.  Rather, I meant that the variable needed to be measuring the same thing.  In other words, the same concept of performance was used in each scenario.  For instance, if you were to measure performance as reaction time in one scenario, and as completeness in the other, these two measure would not be comparable.  You could call each performance in a model, but they wouldn't be able to be used as a single variable in that sense.  It sounds like you do have the same measure for performance in each scenario, so what I previously wrote will work fine.  Your model approach should be more than acceptable.

Matthew J Poes
Research Data Specialist
Center for Prevention Research and Development
University of Illinois
510 Devonshire Dr.
Champaign, IL 61820
Phone: [phone]
email: [hidden email]


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of pomperu
Sent: Wednesday, July 04, 2012 3:50 AM
To: [hidden email]
Subject: Re: Linear mixed model/ time-dependent covariate

Hi Matthew,
Thanks a lot for your quick reply!

In fact, performance IS different for each condition. One condition involves a spatial discrimination task, the other a semantic discrimination task. It might very well be that the two groups (patients and controls) perform different in the two conditions (i.e there is an interaction between group and performance).

This condition-dependence of the covariate  is the reason why I figured the linear-mixed model would be the correct method. Am I wrong? Would there be some regression based approach to this analysis?

Thanks a lot,
Ulrich

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Linear-mixed-model-time-dependent-covariate-tp5713986p5714001.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

[pomperu]

Dear Matthew,
Thanks a lot for you reply! The concept of performance is the same in both conditions, so it seems I'm fine.

Could you give me a hint on the correct post-hoc procedure? I found an interaction between group (between subjects) and condition (within subjects).
My first guess was to run another linear-mixed model (including the covariate), seperately for both groups.

Again thanks a lot for your help!

[Poes, Matthew Joseph]

No post hoc is necessary, all the information you need is in that model really.  If the main effect and interaction are both significant, and the coefficients are both positive, you know that everything is significant, because as coefficients get farther from zero, they have to be significant.  If one is positive and the other negative, that does give the possibility that one scenario is not significant, while the other is.  Plotting helps with this, and yes, running separate models can help with this.  I believe there are also calculators online that you can plug in your numbers to see if everything is significant or not.  Still, this is often not strictly necessary, most people don't do this extra last step work (not that it will hurt).

Matthew J Poes
Research Data Specialist
Center for Prevention Research and Development
University of Illinois
510 Devonshire Dr.
Champaign, IL 61820
Phone: [phone]
email: [hidden email]



-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of pomperu
Sent: Friday, July 06, 2012 4:13 AM
To: [hidden email]
Subject: Re: Linear mixed model/ time-dependent covariate

Dear Matthew,
Thanks a lot for you reply! The concept of performance is the same in both conditions, so it seems I'm fine.

Could you give me a hint on the correct post-hoc procedure? I found an interaction between group (between subjects) and condition (within subjects).
My first guess was to run another linear-mixed model (including the covariate), seperately for both groups.

Again thanks a lot for your help!

--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Linear-mixed-model-time-dependent-covariate-tp5713986p5714046.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

[pomperu]

Awesome Matthew, thanks a lot!!!