Hi Jesper
So you think the best model in this case is to use
3 column EVs with onset time, fixed duration and then a demeaned
RT( total RT- trial RT) as a modulator,
and do one EV1 for valid and another EV2 for invalid, is that right?
however, one 'problem' I see with this option is that it seems to
assume that the response across the brain
may be related to the actual RT, and although this may be the case for
more high;level decision stages
it may not be the case for more automatic visual processing.
So fat I understand that one way of leading with the above is to
implement the other option you suggested
(although you are not quite convince about it)) with 2 EVs, one for
valid and invalid with constant duration and height
and a third EV for RT where I include ALL the demeaned RTs for valid
and invalid onset times...
I could play with the two options and see whether it makes any
difference.
Thank you very much for all your help and time-- if you have ay final
recommendation before I start doing the analyses
I'll be happy to hear.
Best wishes!
On 20 Oct 2009, at 09:07, Jesper Andersson wrote:
> Hi again,
>
>> I am thinking on this option with 6 EVS on 3 Column format
>>
>> EV1 -Search onset valid cue trial
>> Trial Onset time, Fixed Duration (i.e. 1 s), Fixed Input (i.e. 1s)
>>
>> EV2- Search onset invalid cue trial
>> Trial Onset time, Fixed Duration (i.e. 1s), Fixed Input (i.e. 1s)
>>
>> subsequently I specify the RT information for search and memory
>> tasks separately as follow
>>
>> EV3 - Search Reaction time (RT) for valid cue search
>>
>> Trial Onset Time, Fixed Duration (i.e. 1 s), search RT
>>
>> NOTE A: I understand that RT should be demeaned right ? so any RT
>> value in column 3
>> should be equal to the individual RT - mean RT on each particular
>> event (i.e, mean valid for EV3, mean invalid for EV4 below)
>>
>> Also from your comments I understand that EV3 should be
>> orthogonalised wrt EV1
>>
>> NOTE B: somewhere in FSL forum I see suggestions to NOT
>> orthogonalise any RT regressor wrt any of the other ones -
>> it is argued that in doing so, if the EVs are positively correlated
>> with RTs then we end up boosting the EVs for valid and invalid
>> trials because by
>> orthogonalising RTs any amount of variance which could be explained
>> either by A/B or RT ends up being attributed to A/B only.
>>
>> EV4- Search RT for Invalid cue search
>> Trial Onset Time, Fixed Duration (i.e 1s),search RT
>> This is orthogonalised wrt EV2
>
> I am not sure I think it is a great idea orthogonalising 3 and 4 w.r.t
> 1 and 2. My previous suggestion with a single modulated EV which you
> mean corrected with the mean of ALL reaction times would not have been
> orthogonal (unless the means for the two event types were identical).
> That mean that you potentially lose some sensitivity, but also means
> that you can be more certain of the interpretation of anything you
> find in the [1 -1] contrast.
>
> Let's say e.g. that you have means 0.3 and 0.6 seconds RT for valid
> and invalid cues respectively and lets further say that you observe a
> difference (in brain activation) between valid and invalid cues. The
> problem now is the interpretation. Is this difference due to the cues
> per se, or is it just a consequence of the reaction times. I.e. if you
> were to e.g. repeat the experiment with only valid cues and then
> divided up your events based on reaction times, would you then see the
> same activation in the short-long contrast?
>
> The way to avoid this ambiguity is by NOT orthogonalising the
> categorical effect w.r.t. reaction times, but instead let the reaction
> time "claim" any activation it possibly can and then leave the
> leftovers for the categorical.
>
>
>> EV5 - Memory RT Valid Trials
>> Trial Onset Time, Fixed Duration (i.e 1s), memory RT
>>
>> EV6 - Memory RT Invalid Trials
>> Trial Onset Time, Fixed Duration (i.e 1s), memory RT
>>
>> I GUESS EV5 and EV6 might also ortogonalised also wrt to EV1 and
>> EV2?
>
> The same comment applies here.
>
>> Remember that I am interested in looking at only
>>
>> EV1 EV2
>>
>> 1 -1
>>
>> -1 1
>
> In conclusion I would say that you should always be very careful with
> explicit orthogonalisation in your design matrices. They are a way of
> removing ambiguities within "numerical analysis", but if the
> ambiguities were already there in the design (as opposed to "design
> matrix") then that have to be reflected by the design matrix.
> Otherwise you risk assigning observed activations to the wrong cause.
> The time to avoid these ambiguities is at the design stage, i.e. when
> deciding on the exact tasks to use. But quite often one ends up in a
> situation where it is just not possible to make sure all confounds are
> orthogonal to the effects of interest, and then one just have to take
> a hit in sensitivity.
>
> Good luck Jesper
|