Hi,
On 20 Oct 2006, at 20:46, Jonathan Hakun wrote:
> I looked further into the tripled t-test design example on the
> site. I was
> wondering if you wouldn't mind helping me understand the
> application of this
> model to my dataset. 8 BOIs-Ntl brought to the group level; I want
> to group
> 4 of the BOIs together and compare that group of BOIs to the other
> group of
> 4 BOIs (my 4 high saliency videos less my 4 low saliency ones --
> and vice
> versa).
>
> The site says that the triple t-test is an extension of the paired
> t-test,
> and I'm a bit confused about the application of a paired model to a
> dataset
> that has 1 group with 1 scan per subject -- granted that one scan
> can be
> divided into 8 conditions (each separate BOI), so maybe that's
> where the
> "paired" piece comes in.
That's right.
> You mentioned that the degrees of freedom in this
> model would be correct. That makes sense. And it makes sense to
> me that we
> would want to remove the global mean (as in the example online) by
> making
> EV's 1-8 be the BOI group value, and the next 18 EV's account for
> variations
> in global mean among the 18 subjects. Does the global mean issue
> serve as
> one of the reasons to use this tripled model?
Indeed - though I wouldn't use the word "global" as that could be
confusing - you're removing each subject-specific mean. By including
that in the paired t-test case you convert the test from unpaired to
paired t-test.
> Any other suggestions about using this model...i.e. Tips on how to
> draw the
> contrasts, I'm a little shaky on the "2 1 0 0 0 0 0" connotation of
> c1 in
> the example model.
Sure - go through the derivation covered in the manual for the
tripled t-test case, namely:
We now want to form the 3 contrasts A-B, A-C and B-C. Note that,
somewhat surprisingly, A-B is not given by [1 0 0...]!
We define PE1=a and PE2=b. Then, we can see by looking at the three
condition blocks, that the mean (on top of the global mean taken out
by EVs 3-7) of condition A is modelled by A=a+b. Likewise, B=-a, and
C=-b (look at the values in EVs 1 and 2).
Therefore we have A-B = 2a+b = contrast [ 2 1 0....],
and A-C = a+2b = contrast [ 1 2 0....],
and B-C = -a+b = contrast [ -1 1 0....].
So first write down the equation for each _row_ of the design,
starting with:
A (input 1) = a + b (i.e. EVs 1 and 2 have a "1")
etc.
Then re-arrange to give the desired tests (e.g. A-C).
Cheers.
>
> ~Jonathan
>
>
>
>
>
> On 10/19/06 11:58 AM, "Steve Smith" <[log in to unmask]> wrote:
>
>> Hi,
>>
>> So this makes sense, and it looks like you're right to just bring
>> copes 10-17 up to second-level.
>>
>> I think your second-level model should be more like an extension of
>> the "tripled t-test" shown at
>> http://www.fmrib.ox.ac.uk/fsl/feat5/
>> detail.html#TripledTwoGroupDifference
>> Otherwise, for example, your degrees-of-freedom will be wrong at the
>> second level.
>>
>> Good luck! Cheers, Steve.
>>
>>
>>
>> On 18 Oct 2006, at 21:20, Jonathan Hakun wrote:
>>
>>> Glad to hear we can adequately model all 8 Neutrals in 1 EV and the
>>> others
>>> individually. I felt like it would be fine, but wasn't sure.
>>>
>>> Regarding neutral, rest, etc. We do have a rest, we did not
>>> model it
>>> however. We chose to not do so, as to not "overmodel."
>>>
>>> Further, we want to use the neutral as a control for our activation
>>> blocks.
>>> All blocks are videos...however the activation blocks have
>>> components that
>>> we wanted to control for by comparing them to the neutral condition
>>> at some
>>> point. This is why I brought COPEs 10-17 (BOI-Ntl) up to the 2nd
>>> level.
>>> The effect of "seeing a video" would be controlled for in the 1st
>>> level
>>> contrast per subject in this way.
>>>
>>> My goal for the 2nd level is to see the group effect of each
>>> activation
>>> block less the effect of the neutral; as well as combinations of
>>> the BOIs
>>> less the neutral (i.e. The 4 high saliency BOIs-Ntl). My decision
>>> was to
>>> contrast the activation blocks minus the neutral on the 1st level
>>> on the
>>> basis that the variability in response to videos in general from
>>> subject to
>>> subject, be they neutral or "activation" videos, would be more
>>> similar per
>>> subject than on the group. Does this make sense? Or would it be
>>> best to
>>> model group BOIs minus group Neutral later?
>>>
>>> That seems like it for the 1st level. On to the 2nd? OR should I
>>> revisit
>>> something in the 1st level?
>>>
>>> ~Jonathan
>>>
>>> On 10/18/06 9:11 AM, "Steve Smith" <[log in to unmask]> wrote:
>>>
>>>> Hi,
>>>>
>>>> On 17 Oct 2006, at 22:29, Jonathan Hakun wrote:
>>>>
>>>>> Hello,
>>>>>
>>>>> I am working with a sample composed of 18 subjects, each with 1
>>>>> scan
>>>>> session. The paradigm I've been modeling is a block design
>>>>> with 16
>>>>> total
>>>>> blocks: 8 individual activation blocks being modeled separately (8
>>>>> individual activation blocks that we want to later weigh, and run
>>>>> some
>>>>> covariates against; so each is being modeled as a separate EV),
>>>>> and
>>>>> another
>>>>> 8 blocks (the neutral condition) which are being modeled as 1 EV,
>>>>> as we can
>>>>> see no need to break these into 8 individual EVs.
>>>>>
>>>>> Question #1: Is it safe to assume that modeling all 8 Neutrals
>>>>> together
>>>>> (denoted in 1 stick file -- unlike the 8 individual activation
>>>>> blocks
>>>>> denoted in 8 different stick files) is similar enough to the
>>>>> activation
>>>>> blocks which are each model separately, such that we can
>>>>> adequately
>>>>> define a
>>>>> contrast of each activation block minus neutral in the 1st level
>>>>> contrast.
>>>>> Or is the EV of Neutral being modeled much more profoundly b/c it
>>>>> consists
>>>>> of 8 blocks versus the activation blocks being 1 block each?
>>>>
>>>> It certainly doesn't matter that the "neutral" condition includes
>>>> more timepoints than the activation blocks - the differential
>>>> contrasts are not biased by this.
>>>>
>>>> Though I do have a question - do you also have separate "rest"
>>>> conditions to the neutral? If not, then in effect your neutral
>>>> is the
>>>> "rest", in which case, as the data and model are demeaned at first
>>>> level, then you should simply exclude the neutral covariate.
>>>> Then you
>>>> would simply discard contrasts 10-17; their questions would be
>>>> covered by 1-8.
>>>>
>>>>> This design, in a nutshell looks like this:
>>>>> (note: "Block_of_interest" = "activation block")
>>>>>
>>>>> 9 EVs in the 1st level:
>>>>> Block_of_interest1 (BOI1_stick)
>>>>> Block_of_interest2 (BOI2_stick)
>>>>> Block_of_interest3 (BOI3_stick)
>>>>> Block_of_interest4 (BOI4_stick)
>>>>> Block_of_interest5 (BOI5_stick)
>>>>> Block_of_interest6 (BOI6_stick)
>>>>> Block_of_interest7 (BOI7_stick)
>>>>> Block_of_interest8 (BOI8_stick)
>>>>> Neutral(8blocksin1)(NTL_stick)
>>>>>
>>>>> BOI1 through 4 are the "low saliency activation blocks" and BOI5
>>>>> through 8
>>>>> are the "high saliency activation blocks."
>>>>>
>>>>> My Matrix for Contrasts at the 1st level looks like this:
>>>>> EV1 EV2 EV3 EV4 EV5 EV6 EV7 EV8 EV9
>>>>> C1_BOI1 1 0 0 0 0 0 0 0 0
>>>>> C2_BOI2 0 1 0 0 0 0 0 0 0
>>>>> C3_BOI3 0 0 1 0 0 0 0 0 0
>>>>> C4_BOI4 0 0 0 1 0 0 0 0 0
>>>>> C5_BOI5 0 0 0 0 1 0 0 0 0
>>>>> C6_BOI6 0 0 0 0 0 1 0 0 0
>>>>> C7_BOI7 0 0 0 0 0 0 1 0 0
>>>>> C8_BOI8 0 0 0 0 0 0 0 1 0
>>>>> C9_NTL 0 0 0 0 0 0 0 0 1
>>>>> C10_BOI1-Ntl 1 0 0 0 0 0 0 0 -1
>>>>> C11_BOI2-Ntl 0 1 0 0 0 0 0 0 -1
>>>>> C12_BOI3-Ntl 0 0 1 0 0 0 0 0 -1
>>>>> C13_BOI4-Ntl 0 0 0 1 0 0 0 0 -1
>>>>> C14_BOI5-Ntl 0 0 0 0 1 0 0 0 -1
>>>>> C15_BOI6-Ntl 0 0 0 0 0 1 0 0 -1
>>>>> C16_BOI7-Ntl 0 0 0 0 0 0 1 0 -1
>>>>> C17_BOI8-Ntl 0 0 0 0 0 0 0 1 -1
>>>>> C18_AllLow 1 1 1 1 0 0 0 0 0
>>>>> C19_AllHigh 0 0 0 0 1 1 1 1 0
>>>>> C20_AllLow-AllHigh 1 1 1 1 -1 -1 -1 -1 0
>>>>> C21_AllHigh-AllLow -1 -1 -1 -1 1 1 1 1 0
>>>>>
>>>>> Question #2 for the 1st level: Now, another question on the 1st
>>>>> level was
>>>>> whether it was statistically legal to model the "all low" and the
>>>>> "all high"
>>>>> contrasts the way I did. Is it reasonable to add 4 betas together
>>>>> in a
>>>>> contrast like this? Or should the "all low" line have looked like
>>>>> this:
>>>>>
>>>>> All Low .25 .25 .25 .25 0 0 0 0 0
>>>>>
>>>>> I was having a hard time understanding whether a contrast had to
>>>>> add up to 1
>>>>> or 0, or not, and in which situations it did not have to.
>>>>
>>>> Indeed - it doesn't matter in terms of the 1st-level tstats that
>>>> you
>>>> will get out from this - scaling a contrast doesn't affect the
>>>> resulting t-stat. However, if you are going to then contrast this
>>>> contrast with another at a higher-level, you need to make sure that
>>>> their absolute scaling is comparable - the easiest way to be
>>>> safe is
>>>> to follow your suggest of using 0.25 instead of 1 as this then
>>>> means
>>>> that the contrast is giving the "mean" effect.
>>>>
>>>> I think given my previous comments above it would be simpler to get
>>>> the first-level queries sorted first, before I embark on the
>>>> second-
>>>> level questions - so maybe you can re-submit the 2nd levels
>>>> questions
>>>> in the light of my answers above?
>>>>
>>>> Cheers, Steve.
>>>>
>>>>
>>>>> 2nd Level:
>>>>>
>>>>> When I bring this data to the 2nd level, in order to model for
>>>>> "BOI1-Ntl",
>>>>> etc, for the group, I had brought up the COPEs for every subject
>>>>> (all 18
>>>>> subjects) for each BOI-Ntl contrast (8 contrasts-C10 through C17:
>>>>> BOI1-Ntl,
>>>>> BOI2-Ntl, ..., BoI8-Ntl). I did this in order to evaluate the
>>>>> group
>>>>> contrast of a given BOI versus neutral, i.e. to see activation on
>>>>> the group
>>>>> level corresponding directly to each activation block less the
>>>>> value of the
>>>>> neutral stimulus. We'll forgoe looking at the data from the 1st
>>>>> level that
>>>>> I had a question about (i.e. C19_AllHigh or C18_AllLow) as I'm not
>>>>> certain I
>>>>> adequately modeled this correctly, and I did not want to bring it
>>>>> up to a
>>>>> group level until I was certain I had it right on the 1st level).
>>>>> So for
>>>>> now just the 8 copes Copes10-17.
>>>>>
>>>>> The model on the 2nd level looks like this:
>>>>>
>>>>> EV's (copes from 1st level 8 per subject by 18 subjects):
>>>>> EV1 EV2 EV3 EV4 EV5 EV6 EV7 EV8
>>>>> Subj1_Cope10 1 0 0 0 0 0 0 0
>>>>> Subj1_Cope11 0 1 0 0 0 0 0 0
>>>>> Subj1_Cope12 0 0 1 0 0 0 0 0
>>>>> Subj1_Cope13 0 0 0 1 0 0 0 0
>>>>> Subj1_Cope14 0 0 0 0 1 0 0 0
>>>>> Subj1_Cope15 0 0 0 0 0 1 0 0
>>>>> Subj1_Cope16 0 0 0 0 0 0 1 0
>>>>> Subj1_Cope17 0 0 0 0 0 0 0 0
>>>>> Subj2_Cope10 1 0 0 0 0 0 0 0
>>>>> Subj2_Cope11 0 1 0 0 0 0 0 0
>>>>> Subj2_Cope12 0 0 1 0 0 0 0 0
>>>>> Subj2_Cope13 0 0 0 1 0 0 0 0
>>>>> Subj2_Cope14 0 0 0 0 1 0 0 0
>>>>> Subj2_Cope15 0 0 0 0 0 1 0 0
>>>>> Subj2_Cope16 0 0 0 0 0 0 1 0
>>>>> Subj2_Cope17 0 0 0 0 0 0 0 1
>>>>> Subj3_Cope10 1 0 0 0 0 0 0 0
>>>>> Subj3_Cope11 0 1 0 0 0 0 0 0
>>>>> ...
>>>>> ...
>>>>> ...and so on
>>>>>
>>>>> I do this in order to bring all the copes for BOI1-Ntl up to the
>>>>> group level
>>>>> where I can make a contrast of "groupmean_BOI1-Ntl"
>>>>> "groupmean_BOI2-
>>>>> Ntl" and
>>>>> so on.
>>>>>
>>>>> So in my contrast page I define things like:
>>>>> EV1 EV2 EV3 EV4 EV5 EV6 EV7 EV8
>>>>> C1-gm_BOI1 1 0 0 0 0 0 0 0
>>>>> C2-gm_BOI2 0 1 0 0 0 0 0 0
>>>>> C3-gm_BOI3 0 0 1 0 0 0 0 0
>>>>> C4-gm_BOI4 0 0 0 1 0 0 0 0
>>>>> C5-gm_BOI5 0 0 0 0 1 0 0 0
>>>>> C6-gm_BOI6 0 0 0 0 0 1 0 0
>>>>> C7-gm_BOI7 0 0 0 0 0 0 1 0
>>>>> C8-gm_BOI8 0 0 0 0 0 0 0 1
>>>>> C9-gm_AllLow-Ntl 1 1 1 1 0 0 0 0
>>>>> C10-gm_AllHigh-Ntl 0 0 0 0 1 1 1 1
>>>>> C11-gm_(Low-Ntl)-(High-Ntl) 1 1 1 1 -1 -1 -1 -1
>>>>> C12-gm_(High-Ntl)-(Low-Ntl) -1 -1 -1 -1 1 1 1 1
>>>>> C13-gm_ALLBOI-Ntl 1 1 1 1 1 1 1 1
>>>>>
>>>>> Question 3: Now, same quesion about C9 and C10. Was I allowed to
>>>>> define a
>>>>> contrast that added up to 4? I didn't feel comforted by this.
>>>>>
>>>>> Question 4: When evaluating C13, I asked whether I could have a
>>>>> contrast add
>>>>> up to 8. I want to see the value of ALL the BOI's minus Neutral.
>>>>> Since
>>>>> each BOI minus neutral was carried out as COPE10-17 on the 1st
>>>>> level, I was
>>>>> hoping I could somehow make a contrast on the 2nd level that would
>>>>> depict
>>>>> ALL BOI-neutral by bring them up and averaging them somehow. I
>>>>> also tried
>>>>> this in a 2nd pass at this data as C14:
>>>>>
>>>>> C14-gm_ALLBOI-Ntl .125 .125 .125 .125 .125 .125 .125 .125
>>>>>
>>>>> This way the contrast adds up to one, giving 1/8 weight to each
>>>>> contrast.
>>>>> Peculiar thing, is that the results are identical (at least they
>>>>> appear to
>>>>> have the same range of Z-scores, and appear to be identical z-
>>>>> maps.
>>>>>
>>>>> So again the major questions here have to do w/ generating
>>>>> contrasts. 1)
>>>>> Can the contrast add up to something other than 1 and why or why
>>>>> not? 2)
>>>>> Does this 2nd level model I've drawn up seem to be a reasonable
>>>>> way to
>>>>> generate a group map for BOI's individually minus neutral, and
>>>>> then
>>>>> again as
>>>>> ALL BOI minus neutral. Is weighing the contrast as 1/8 each the
>>>>> correct
>>>>> modification to the contrast to see ALL_BOI-Ntl? 3) Also, might I
>>>>> instead
>>>>> carry out a 3rd level analysis, where each of C1-8 from the 2nd
>>>>> level is
>>>>> defined as 1 single EV where a 3rd level contrast shows the
>>>>> positive 1 value
>>>>> of that EV for the group...would that look exactly the same as my
>>>>> C13 or C14?
>>>>>
>>>>> Thank you in advance for glancing at these models. I'm not new so
>>>>> much to
>>>>> modeling fMRI data, as a I am to FSL. I've typically worked w/
>>>>> SPM, and the
>>>>> conversion is stumping me from time to time.
>>>>>
>>>>> After someone digests this, I have a question about modeling
>>>>> correlations
>>>>> based on the contrasts I've generated at the 2nd level. I'll ask
>>>>> this in
>>>>> another string.
>>>>>
>>>>> ~Jonathan
>>>>
>>>>
>>>> -------------------------------------------------------------------
>>>> --
>>>> ---
>>>> ---
>>>> Stephen M. Smith, Professor of Biomedical Engineering
>>>> Associate Director, Oxford University FMRIB Centre
>>>>
>>>> FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
>>>> +44 (0) 1865 222726 (fax 222717)
>>>> [log in to unmask] http://www.fmrib.ox.ac.uk/~steve
>>>> -------------------------------------------------------------------
>>>> --
>>>> ---
>>>> ---
>>>>
>>
>>
>> ---------------------------------------------------------------------
>> ---
>> ---
>> Stephen M. Smith, Professor of Biomedical Engineering
>> Associate Director, Oxford University FMRIB Centre
>>
>> FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
>> +44 (0) 1865 222726 (fax 222717)
>> [log in to unmask] http://www.fmrib.ox.ac.uk/~steve
>> ---------------------------------------------------------------------
>> ---
>> ---
>>
------------------------------------------------------------------------
---
Stephen M. Smith, Professor of Biomedical Engineering
Associate Director, Oxford University FMRIB Centre
FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
+44 (0) 1865 222726 (fax 222717)
[log in to unmask] http://www.fmrib.ox.ac.uk/~steve
------------------------------------------------------------------------
---
|