Mixed model repeated measures problem

[Mike Ford]

I want to do a meta-analysis of two psycholinguistics experiments
investigating reading times to words. Both experiments have the same words
in them, but used slightly different instructions.

I would like to do a mixed model in which I investigate the influence of
five continuous IVs and the factorial IV "Experiment" (and later
interactions of Experiment and the continuous IVs)

I think I should model the fact that Word1 in Experiment 1 is the same as
Word1 in Experiment 2. I think I need some kind of REPEATED statement to
model the fact that the residual error for particular words will be
correlated across experiments. But I can't come up with a way to do this,
the things I try generate error messages.

The model I have run without the REPEATED analysis is as below with random
effects for participant and word variation.

Mixed ReadingTime by Experiment with x1 x2 x3 x4 x5
    /FIXED = Experiment x1 x2 x3 x4 x5
    /PRINT = SOLUTION TESTCOV
    /RANDOM = INTERCEPT | SUBJECT(Participant) COVTYPE(UN)
    /RANDOM = INTERCEPT | SUBJECT(Word) COVTYPE(UN) .

How should I model the repetition of items across the two experiments?

[Art Kendall]

If I understand your post. using the same conditions (e.g., same words)
in the two experiments would not create a repeated measure unless the
same subjects were used in the two experiments. In that instance,
experiment would be the repeated factor.

Art Kendall
Social Research Consultants.

Mike Ford wrote:

> I want to do a meta-analysis of two psycholinguistics experiments
> investigating reading times to words. Both experiments have the same words
> in them, but used slightly different instructions.
>
> I would like to do a mixed model in which I investigate the influence of
> five continuous IVs and the factorial IV "Experiment" (and later
> interactions of Experiment and the continuous IVs)
>
> I think I should model the fact that Word1 in Experiment 1 is the same as
> Word1 in Experiment 2. I think I need some kind of REPEATED statement to
> model the fact that the residual error for particular words will be
> correlated across experiments. But I can't come up with a way to do this,
> the things I try generate error messages.
>
> The model I have run without the REPEATED analysis is as below with random
> effects for participant and word variation.
>
> Mixed ReadingTime by Experiment with x1 x2 x3 x4 x5
>     /FIXED = Experiment x1 x2 x3 x4 x5
>     /PRINT = SOLUTION TESTCOV
>     /RANDOM = INTERCEPT | SUBJECT(Participant) COVTYPE(UN)
>     /RANDOM = INTERCEPT | SUBJECT(Word) COVTYPE(UN) .
>
> How should I model the repetition of items across the two experiments?
>
>
>

[Mike Ford]

Thanks for the advice. Feel a bit silly I didn't think about the problem
properly. Got a bit caught up in thinking about how to do mixed models.

I have had a look at some documentation but couldn't find an example similar
to what I wanted to do, they all seemed to be much more complex repeated
measures.  The syntax that seemed to work was


Mixed ReadingTime by Experiment with x1 x2 x3 x4 x5
    /FIXED = Experiment x1 x2 x3 x4 x5
    /PRINT = SOLUTION TESTCOV
    /RANDOM = INTERCEPT | SUBJECT(Participant) COVTYPE(UN)
    /RANDOM = INTERCEPT | SUBJECT(Word) COVTYPE(UN) .
/REPEATED = Exp | SUBJECT(Item*Subj) COVTYPE(AR1)  .

 /REPEATED = Exp COVTYPE(AR1)  .