Dear Ian and Rachel,
This is a rather late comment on the mailbase thread below, pertaining
to the design of single subject design matrices for random effects
analyses in SPM.
*********************************
Dear Rachel,
Rachel Mitchell <[log in to unmask] >said:
> Is it true that for a random effects analysis of fMRI data that it is
> preferable that you present subjects with the set of experiments in
> the same i.e. fixed order rather than counterbalance them?
This is a question relating to experimental design and repeated
measures analysis of variance. The simple answer is 'no, it is not
true'. The longer answer is that designs which use counterbalancing or
randomising orders of conditions may perhaps allow better estimation of
condition effects in the context where there may be carry-over effects
from preceding conditions.
Best wishes,
Ian
********************************
fMRI random-effects (RFX) analyses as currently implemented by SPM99
rely on two assumptions: first-level, single session/subject design
matrices must be both 'balanced' and 'separable'. If the first-level
design matrices have i) different numbers of covariates or ii) contain
the same numbers of covariates but change the order of the elements in
the covariates (by randomising/counter-balancing), then these
assumptions may fall down.
However, it is not currently clear how robust SPM's RFX analysis is to
violations of these assumptions. This would be useful to know, because
for design reasons one usually wants to randomise or counter-balance the
presentation of stimuli to subjects (as Dr. Nimmo-Smith says above).
One's choices therefore seem to be limited to: i) randomised
experimental design which estimates subject and condition effects well
BUT may not be suitable for RFX analyses or ii) present each subject
with same design, perform 'correct' RFX analysis BUT fall victim to
order effects.
There is undoubtedly a 'third way' that will allow you to have your cake
and eat it - the implementation of a weighted-regression type approach
to
2nd-level analyses, for example. And, while acknowledging these
difficulties, it may be that in most cases the use of randomised designs
in SPM RFX analyses will not present major problems. However, as I
haven't seen much discussion on this topic, I wanted to see if anyone
was exploring this area.
Best ,
Dave McG.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|