Re: Longitudinal comparison partial vs. whole sample

Posted by Ryan on Mar 05, 2013; 9:32pm
URL: http://spssx-discussion.165.s1.nabble.com/Longitudinal-comparison-partial-vs-whole-sample-tp5718123p5718383.html

First off, one can employ a linear mixed model (via REML estimation) on just the paired data (removing subjects who did not provide data at time 2, which will produce the same results as those produced by a paired t-test, assuming one were to force a compound symmetric structure when parameterizing the linear mixed model. I'd also like to point out that, within a linear mixed model, unlike a paired-t test, one can allow for heterogeneous variances across time points. It is not uncommon to see variances decrease substantially over time. Allowing for heterogeneous variances could significantly improve model fit, and produce a more statistically powerful test of the question at hand.

 

As I mentioned in a previous email to the OP, it is still unclear to me how many times individuals were measured at each time point. If individuals were measured thousands of times at a given time point, then I would want to know why, and potentially build these repeated measures into the statistical model. Clearly, a paired t-test would not be a viable option under such circumstances, unless one were to aggregate the data per subject in some way (e.g., take the mean of the thousands of measurements on a given subject at each time point and treat the mean as the subject's score). I'm generally not in favor of such an approach, but could be convinced otherwise, depending on the situation. Admittedly, eye tracking is not my area of expertise, so my understanding of the study may be incorrect.

 

I agree with Rich that if the data are not missing at random the second time around, we need to try to find out WHY before employing a linear mixed model on all possible data. Frankly, if the data are not missing at random, then I would question any type of analysis on the retained data as well. This is why I posted my message yesterday asking the OP to indicate whether the data were missing at random, among other things.

 

Ryan

On Mon, Mar 4, 2013 at 7:36 PM, Rich Ulrich <[hidden email]> wrote:

Sebastion,

I think it will be a lousy paper that uses a mixed model approach
on eye-tracking data instead of using the paired t-test.  Paul's
suggestion that the non-paired data can be used for estimating
(perhaps) the variance would be somewhat legitimate if the
correlations between pairs were small, and *after*  you have
established that there is no reason to think that the people with
missing information may be different.

My comment about "here is what you would see" ...
I was saying that, (1) after you do the tests as I have described,
for the sake of doing "proper testing,"  (2) then you might resort
(for the sake of convenient presentation) to testing the 76 vs. the
20 as if they were two independent groups, even though they are not.
That should be tossed in as something extra, and not as the test
that ought to be considered primary and most useful. 

--
Rich Ulrich

> Date: Mon, 4 Mar 2013 06:15:12 -0800
> From: [hidden email]
> Subject: Re: Longitudinal comparison partial vs. whole sample
> To: [hidden email]

>
> Dear Paul, Rich and Bruce,
>
> thank you for your quick replies and sorry I kept you waiting. I was quite
> busy the past two weeks.
>
> @ Paul: I will go with the mixed models approach. I would be VERY thankful
> for a hint on a paper, which might have used a similar approach OR advice on
> how to do this on SPSS.
>
> @ Rich: What did you mean by "and (therefore) here is what you would see as
> that test"?
>
> @ Bruce: I have a few thousand data points per subject.
>
> Best regards,

> ...