GEE contrast marginal mean

[julia333]

Hi,
I have checked out the generalized estimating equations in spss. I have two questions:
- I want to calculate a special contrast in the marginal means (EMMEANS). one of my main factor has 6 categories and i wish to contrast category 5&6 to 2&3&4. I did not find it possible to implement this in the syntax. is there an easy way to do this or to get this? My idea was to get the covariance-matrix and parameter estimates and do a generalized wald test by matrix calculation. The contrast vector would be 0 for all other parameters, -0.3 for 5 and 6 and 2,3,4 for 0.2. what do you think? Any smarter ideas are welcome :-) (is my idea correct anyway?)

And finally, just to be sure about the different interpretation of the parameter estimates and the marginal mean. I think that the parameter estimates are the estimates of the marginal mean, so in case of no interaction, both are the same. in case of interactions , parameter estimates are depending on coding (at least for logit link function), but the marginal means are not and thus main effects are still interpretable.

Thank you so much for helping me. I garantee that I researched the web before posting here. bye

[Ryan]

What's your dependent variable? Also, what link function do you plan on using?

Ryan

On Nov 10, 2011, at 10:13 AM, julia333 <[hidden email]> wrote:

> Hi,
> I have checked out the generalized estimating equations in spss. I have two
> questions:
> - I want to calculate a special contrast in the marginal means (EMMEANS).
> one of my main factor has 6 categories and i wish to contrast category 5&6
> to 2&3&4. I did not find it possible to implement this in the syntax. is
> there an easy way to do this or to get this? My idea was to get the
> covariance-matrix and parameter estimates and do a generalized wald test by
> matrix calculation. The contrast vector would be 0 for all other parameters,
> -0.3 for 5 and 6 and 2,3,4 for 0.2. what do you think? Any smarter ideas are
> welcome :-) (is my idea correct anyway?)
>
> And finally, just to be sure about the different interpretation of the
> parameter estimates and the marginal mean. I think that the parameter
> estimates are the estimates of the marginal mean, so in case of no
> interaction, both are the same. in case of interactions , parameter
> estimates are depending on coding (at least for logit link function), but
> the marginal means are not and thus main effects are still interpretable.
>
> Thank you so much for helping me. I garantee that I researched the web
> before posting here. bye
>
> --
> View this message in context: http://spssx-discussion.1045642.n5.nabble.com/GEE-contrast-marginal-mean-tp4981370p4981370.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

[julia333]

Thanks, my dependent variable is binary (1-0), and I use a logit link function. the dependent variable indicates if a participant has responded to a task or not. the experiment hast several repeated measures, therefore ordinary logistic regression is not appropriate. I tried also to implement "fake variables" which are redundant, but describte the comparisons I need and then order the marginal mean, but the results of these marginal mean contrast make no sense at all. Tanks

[Ryan]

My guess is that this type of contrast on the probability scale derived from a binary logistic regression is not possible in GENLIN.

Ryan

On Nov 11, 2011, at 2:16 PM, julia333 <[hidden email]> wrote:

> Thanks, my dependent variable is binary (1-0), and I use a logit link
> function. the dependent variable indicates if a participant has responded to
> a task or not. the experiment hast several repeated measures, therefore
> ordinary logistic regression is not appropriate. I tried also to implement
> "fake variables" which are redundant, but describte the comparisons I need
> and then order the marginal mean, but the results of these marginal mean
> contrast make no sense at all. Tanks
>
> --
> View this message in context: http://spssx-discussion.1045642.n5.nabble.com/GEE-contrast-marginal-mean-tp4981370p4985274.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