Hi Amit,
Assuming you had inserted null events such as rests or fixation
periods, and assuming that you did not explicitly model them, then
you could answer your questions with contrasts of [1 00] or its
negative [-1 0 0] for condition A, and [010] and its negative [0-1 0]
for condition B. These would tell you for each contrast whether the
HRF was positive going or negative going at each voxel, helping to
disambiguate the direct A vs B contrast.
>Hi Eric and Dan,
>
>Thanks, it seems that this question would be difficult to answer if
>there is no baseline.
>
>If there was a baseline in between two condtions one task and the
>other control, what would be the contrast that you would use to
>answer the original question. Can this be done with just one time
>analysis with an appropriate contrast or would you have to mask with
>other contrasts?
>
>Amit
>
> -----Original Message-----
> From: Eric Zarahn [mailto:[log in to unmask]]
> Sent: Sun 5/29/2005 6:48 PM
> To: Daniel H. Mathalon
> Cc: Anand, Amit; [log in to unmask]
> Subject: Re: [SPM] Inference fo contrast
>
>
> Yes to your statement if there are null events; the null
>events would be a third condition. Whether the third condition is
>implicitly or explicitly modeled would affect the form of the
>contrasts.
>
> No to your statement if there are not null events (or some
>other condition), in which case there is no baseline (implicit or
>otherwise) with which to compare. If there are only two total
>conditions (that's total, including both implicitly and explicitly
>modeled conditions) in the experiment, A and B (i.e., the rank of
>the design matrix is 2 == there are only 2 estimable parameters in
>the model), then a contrast of the type you suggest will give one of
>two unusable answers depending on how the design is parameterized:
>
> 1) If the 2 conditions are explicitly modeled and there is in
>addition an overall intercept term (which I believe is the default
>in SPM) such that the design matrix is not of full column rank, then
>the contrasts [x 0 0] and [0 x 0], where x is any scalar, are each
>inestimable (or so unstable that for pratical purposes they are
>inestimable).
>
> 2) If the 2 conditions are explicitly modeled and there is no
>additional overall intercept term (which might be irrelevant for
>SPM, but nevertheless) then what you'd get from [x 0] or [0 x] would
>be dominated by non-physiological T2* signal, and both would be
>quite positive, always.
>
> One needs to be careful from a modeling perspective when
>discussing "baselines", "rest condition", "null events", and
>implicit/explicit modeling of them. In your question you said
>
> Assuming that there are some time periods that are
>not modelled (i.e., null events), or even if there aren't,
>
> That last bit is not inconsequential. There is a qualitative
>difference vis a vis modeling between there being a baseline either
>implicit or explicit (which you can call a null event) or not.
>Assuming there is an overall intercept term (and I believe there
>always is one used in all SPM programs to date), what this affects
>is whether the design matrix is of full column rank (corresponding
>to my (2) above) or not (corresponding to my (1)).
>
> And finally, the null event does not need to be a "rest"
>condition per se. It needs to be whatever you want it to be to make
>your statements about the signal in each of the remaining conditions
>meaningful scientifically.
>
> Eric
>
>
>
>
>
> couldn't amit do a one sample t-test on the beta
>images for each condition separately, testing whether beta values
>are significantly different from zero? By looking at t contrasts of
>1, then -1, wouldn't he be able to assess whether beta values at a
>particular voxel location are significantly positive or
>significantly negative relative to implicit baseline?
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
> Hi Amit,
>
>
>
>
>
> ----- Original Message -----
>
> From: "Amit" <[log in to unmask]
><mailto:[log in to unmask]> >
>
> To: <[log in to unmask] <mailto:[log in to unmask]> >
>
> Sent: Sunday, May 29, 2005 4:57 PM
>
> Subject: [SPM] Inference fo contrast
>
>
>
>
> > Dear SPMers,
> >
> > The experimental contrasts (a difference
>between two conditions A and B)
> > is interpreted in usual SPM analysis as
>indicating that condition A shows
> > a larger response than condition B.
>However, difference measures,
> > as conceptualized in SPM could have three
>potential ways in which
> > conditions may differ: positive activity in
>A may be greater than positive
> > activity in B, positive activity in A may
>be greater than negative activity
> > in B, and finally, negative activity in B
>may be greater than negative
> > activity in A. All lead to a positive
>difference between conditions.
>
>
>
>
> Yes, absolutely. But implicit in your
>explanations is a third conditiion C to which you are comparing A
>and B.
>
>
>
>
>
>
>
>
>
> >
> > Is there a way of teasing out these three
>differences?
>
>
>
> If you have a meaningful C in your
>experimental design, then you can compute 2 new contrasts to get
>your answers:
>
>
>
> (1) A minus C
>
>
>
> and
>
>
>
> (2) B minus C.
>
>
>
>
>
>
> >
> > Also, how does one get a contrast for areas
>which are getting inhibited by
> > a particular task - does one just reverse
>the contrast for the active
> > versus control condition in a block-design
>experiment or is there a more
> > sophisticated way for doing this?
>
>
>
> The use of the word "inhibited" has
>connotations that I am not sure you intended to convey. But if you
>simply meant "how do see where the signal in B is greater than that
>in A?", then yes just compute the contrast B minus A.
>
>
>
>
>
> Eric
>
>
>
>
>
>
|