Hi BoKyung
> Hi! I have a question about my 1st-level modeling. I have two
> conditions, one is the experimental condition, in which the
> participants were asked to choose one of two different items presented
> on the monitor, and the other was the baseline condition, in which
> they performed another simililarly-matched task.
> In experimental condition, the participants made lots of correct
> answers and a few wrong answers--I want to see the difference between
> these 'correct' versus 'incorrect' trials, just in the experimental
> condition, but it's hard to decide which 1st level matrix to make
> between the followings.
> 1. Should I extract the stimuli onset time of both 'correct' and
> 'incorrect' trials, and then consider these two types of trials as two
> different conditions? If this is the case, I should put the onset
> times of these two conditions in the 'Condition-Onsets' of the gui and
> make a contrast image by contrasting these two conditions '+1, -1',
> right?
> or
> 2. Should I consider these two types of trials as a parametric
> modulation? If this is the case, I should put the 'experimental' and
> 'baseline' condition onset times in the 'Condition-Onsets' of the spm
> gui, and then make parametric modulations for both conditions. In
> 'experimental' condition, for example, the value of the parametric
> modulation for each trial will be +1 for the 'correct' trials and -1
> for the 'incorrect' trials; in 'baseline' condition--this is a little
> complicated for me because I just want to see two different types of
> trials in only 'experimental' condition and not 'baseline'
> condition--all parametric values should probably be 0, right? Then
> I'll make a contrast image by putting '0, 1, 0, 0', each refers to
> 'experimental condition, parametric modulations of experimental
> condition, baseline condition, parametric modulations of baseline
> condition.
> Which of these two approaches will be appropriate for my analysis?
> Could you guys help me with this? Further, most of the participants
> made lots of correct responses and just a few of incorrect
> responses--is this imbalance of the number of trials ok?
For coding, I can see 3 options
- separate good and bad trials (probably for both conditions) ie make 4
regressors, this will allow you to test what you want (eg [1 -1 0 0]),
although the error related to the error beta will be different due to
the different number of good / bad trials (not such a big deal) --> one
issue with this approach is however that it could lead to spurious
results. Imagine that in some trials there is some weird activity going
on in area X such as high activity increases the likelihood to make an
error but only for your experimental task, testing good trials
experimental task vs good trials baseline task could show you this area
X even though it is not related to your experiment, that is contrasting
conditions based on the behaviour (ie good trials) could be a problem.
(see R VanRullen paper in Front. in Psychology for an example with ERPs)
- code all trials + bad trials (again I think it's better to do it for
both conditions) ie make again for 4 regressors but this time 2
regressors code for all trials and 2 only for bad trials. Of course,
regressors will be correlated to some degree and the shared variance
will go into the error term (that is the strength of the effect for good
trial might be reduced). However, since you have only a few errors that
should not be a problem. You can still test good vs. bad ([1 -1 0 0])
and you get ride of the possible spurious effect mentioned above. Also
the assumptions are different: before you assume that a brain area
involved in your task 'switch off' when there is an error, whereas here
you assume that a brain area is always on for the task and the same area
or another one is also activated by the error (I personally favour this
approach when there are only few errors)
- code all trials + parametric regressors ; here your assumption is that
there are some brain regions always on for the task and others (or the
same) that changes as a function of the answer - it makes sense to use
+1 and -1 (again I would just do the same for experimental and baseline
tasks). However using +1/-1 means that you expect an increase of
activity when right and a decrease when wrong. If you were to code 0 for
good and +1 for error that would give you the same model as option 2. To
test this effect a contrast for the parametric regressor is what you
want ([0 1 0 0 ]).
Now the million dollar answer to your main question 'Which of these two
approaches will be appropriate for my analysis?' --> in GLM the M is for
modelling and that's where your assumptions (priors related to the
experiment) come in ; I think this is a matter of you thinking about
what is happening in the brain when there are errors ... if there are
only few errors I prefer option 2 but I can't advice on what to choose
Good luck
Cyril
--
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
|