Dear Mark,
Thank you for explaining this to me! I really appreciate it. I set up my F-test as you recommended and am looking forward to seeing how it comes it.
Sincerely,
Jake
On Jan 13, 2013, at 3:47 AM, Mark Jenkinson wrote:
> Dear Jake,
>
> There are a few things that are not quite right in your email.
>
> 1 - The second design does not need different (variance) groups, and a lot of the time it is better not to have different group values unless the groups are large. Furthermore, the group designation is only something that is used in the parametric stats (GLM within FEAT) and is not used by randomise, as randomise is non-parametric and does not model variances and so does not need to know about variance groups.
>
> [Aside: the design.grp file is sometimes used as an input to randomise, and this requires the user to do it explicitly, but is not used to denote different variance groups in that case, it is used to denote exchangeability, which is not required here or for most uses of randomise]
>
> 2 - It is not first_utils that prevents the use of different groups, it is a result of using randomise.
>
> 3 - When using first_utils the quantity that is created is a scalar (at each boundary voxel) that represents the projection of that boundary coordinate onto the normal vector of the averaged surface (the signed, perpendicular distance from the average surface, where a positive sign is outside and a negative sign is inside). Hence the mean value of this quantity contains no information and all designs should reflect this by not explicitly modelling the mean. That is why the first design is more correct (as it does not include the mean) whereas the second design you include does allow the mean to be modelled (a [1 1] contrast would give the mean). This is also the reason that the -D option is randomise is used.
>
>
> If you want to know about the directionality of the change, then it is easiest to look at the 4D input file. I've updated the documentation to include a paragraph describing this:
>
> "The --fonly option is included so that only the F-statistic is calculated and used by randomise for inference. This is then equally sensitive to changes in either direction (e.g., growth or atrophy). To determine which direction the change is in there are two possible approaches: (i) omit this option and look at the t-contrast results (carefully taking into account the signs used in both the design matrix EVs and the contrast) or; (ii) more simply, just look at the signs of the individual data points in the 4D input file. The second option is generally easier and is quite simply done with FSLView, by loading the corrected p-value output from randomise (the f-statistic result, having kept the --fonly option) then adding the 4D input file and using the "timeseries" option in FSLView to see the values from the 4D input file at boundary locations where there was a statistically significant change."
>
> The reason I discourage using the t-contrast results is because (i) randomise will create quite a few different output files in that case which can be confusing, and (ii) it is difficult to keep track of the signs correctly, as the signs used in both the EVs and in the contrasts make a difference, and so it is much safer to just look at the input values.
>
> I hope this answers your questions.
> All the best,
> Mark
>
>
>
> On 11 Jan 2013, at 16:55, John Kuster <[log in to unmask]> wrote:
>
>> Thanks Chris.
>>
>> Another question, I am assuming that in both cases I found the single EV, with -1=ctrl and 1=pat and using --fonly was just set up as an example. (both in practical and user guide)
>> ie:
>> group EV
>> 1 -1
>> 1 -1
>> 1 1
>> 1 1
>> for randomise:
>> tcontrast 1= 1
>> ftest selected on tcontrast 1
>>
>> I guess normally for the randomise step to create a contrast, you'd want to run EV1=ctrl as 1, EV2 = ctrl as 1.
>> ie:
>> group EV1 EV2
>> 1 1 0
>> 1 1 0
>> 2 0 1
>> 2 0 1
>> tcontrast 1 = 1 -1
>> tcontrast 2 = -1 1
>> ftest selected on contrast1 vs contrast2
>>
>> However, I realize that for first_utils, it does not recognize groups so the design must be set up the original way. Do people recommend just using a separate design file for the randomise step should we be interested in which group is changing? Or, is the only way to see group directionality by going back to the "old" method, the surface_fdr tool. I apologize if I am not writing this clearly as I am new to this tool, in essence I want to see know if there is a difference, and also for which group and which direction.
>>
>> Thanks again,
>> Jake
>>
>>
>> On Jan 11, 2013, at 10:27 AM, Chris Watson wrote:
>>
>>> I wouldn't say the F test is one directional, but rather non directional. It will just tell you if there is a difference between groups. It won't tell you if A > B or B > A.
>>>
>>> On 01/10/2013 11:43 PM, Jake Kuster wrote:
>>>> PS, to clarify, I know F test is one directional and shows if a difference
>>>> exists between groups. I was just wondering if there might have been
>>>> something else going on due to the wording of the quote below, and because
>>>> the --fonly flag was used in the example. Maybe it only was reiterating
>>>> the sensitivity to change in either direction.
>>>>
>>>> In this case wouldn't there be benefit in setting up the contrast
>>>> differently, so that a t-test could be performed at the same time?
>>>>
>>>> Thanks!
>>>>
>>>> Jake
>>>>
>>>>
>>>>
>>>> Dear FSLers,
>>>>> Reading the website, I noticed it says
>>>>>
>>>>> "This F-test will be the main contrast of interest for our vertex analysis
>>>>> as it allows us to test for differences in either direction."
>>>>> http://www.fmrib.ox.ac.uk/fslcourse/lectures/practicals/seg_struc/index.html
>>>>>
>>>>> So I was wondering if someone could elaborate on how the F-test shows
>>>>> these differences in either direction. Is directionality maintained at
>>>>> this step because of the way the EV is set up, ctrl=-1 and patients=1? I
>>>>> wasn't able to figure it out yet by reading the archives and was confused
>>>>> by the nature of an F-test.
>>>>>
>>>>> Thanks in advance,
>>>>>
>>>>> Jake
>>>>>
>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>> The information in this e-mail is intended only for the person to whom it is
>>>> addressed. If you believe this e-mail was sent to you in error and the e-mail
>>>> contains patient information, please contact the Partners Compliance HelpLine at
>>>> http://www.partners.org/complianceline . If the e-mail was sent to you in error
>>>> but does not contain patient information, please contact the sender and properly
>>>> dispose of the e-mail.
>>>>
>>>
>>>
>
>
|