Many thanks Ged and Christian,
Looks like good advice either way. I'll think about this one for a
while to work out which technique makes most sense for our model (we are
using the 3 level parametric N-BACK task). I'm not sure about the birth
order idea as birth order and obstetrics may have some (not well
defined) impact on our patient group (bipolar). I think it might be
better just to randomise the direction in the controls and concordant
pairs?
I'll reply to the list later on to let people know what we chose and
whether it worked!
Thanks again
Fergus
----------------------------
Christian's Reponse:
Dear Fergus,
the simplest way to analyze the twin pairs is to calculate difference
images between twin and co-twin. These intra-pair differences can be now
compared between discordant monozygotic and dizygotic twins (genetic
effect) and between discordant and control monozygotic twins (disease
effect). The concordant twins can be used to test the effect of genetic
liability. See the paper of T.D. Cannon:
http://www.pubmedcentral.nih.gov/articlerender.fcgi?tool=pubmed&pubmedid
=11867725
The advantage of using intra-pair differences is that you remove the
variance between the twin and co-twin and only incorporate variance
between twins (intra-pair vs. inter-pair variance). This is maybe the
same idea like a random effects (or second level) analysis of fMRI data
where you remove the intra-subject variance (intra-subject vs.
inter-subject variance).
One problem of using intra-pair differences is the order/direction of
the difference. For the discordant twins this is clear, for the
concordant and control twins one suggestion is to use birth order.
However, the order/direction issue is also occurring if you you don't
use intra-pair differences but the single data in a more fancy factorial
model (as Ged proposed).
Best regards,
Christian
---------------------------------
Ged's Response.
Subject: Re: [SPM] Twins in SPM - Assumptions of independence
Hi Fergus,
I think you can probably model this using SPM5's flexible factorial
set-up, with a factor for twin-pair and a factor identifying each twin
within each pair (and a factor for concordance, and/or a factor for
diagnosis -- I'm not sure from your description whether you need one
or both) in a similar way to longitudinal studies where there is a
subject factor and then other factors including time. There were a few
posts to the list recently about within-subjects designs.
You might also want the non-sphericity option for dependence over
levels of your "within-pair" (like within-subjects for a longitudinal
study) factor(s). Though my (limited) understanding of this is that:
1. It will only make a difference if you have more than two levels of
any within-subject factor(s), i.e. if you're only interested in a
contrast of concordant vs discordant then it shouldn't matter.
2. It won't give quite the same results as modelling dependency in a
conventional stats package (like STATA, SPSS or SAS), since the
variance components are estimated with data pooled over all voxels
that pass a certain "main-effects" threshold, rather than separately
for every voxel (which would be very slow, and possibly unstable).
I hope that helps,
Ged.
P.S. Please direct any replies to the list, rather than just to me;
firstly because that'll let other more knowledgeable people correct my
nonsense (!) and secondly because I might not be checking the list
quite so (over-) frequently as I usually do, for the next month or so.
Fergus Kane wrote:
> Dear All,
>
> This question has been asked once before but went unanswered. I am
> analysing twin pairs discordant (one has one has not a diagnosis) and
> concordant (both have) for a psychiatric disorder - as well as control
> pairs. The sample is - mixed identical and non identical.
>
> The problem is: most statistical tests assume independence of
observation,
> but twins are clearly not independent. Depending on the model, this
results
> in a violation of such assumptions between and within groups. I'd
like to
> know if there is any way of addressing this issue within SPM?
>
> Many Thanks
>
> Fergus
>
|