On Mon, Jul 9, 2012 at 9:40 PM, Zhenhao SHI <[log in to unmask]> wrote:
> Just to join in on this discussion. According to my experience, the incorrect error term, which was caused by putting contrasts in an ANOVA-like model, could give weird conjunction results. For example, many regions not activated in contrast A may light up in conjunction of A&B.
There are two issues here:
(1) If you have a repeated-measures ANOVA, the error term represents
the within-subject error, but when you are evaluating A, you need to
use the between-subject error term that is not available in SPM. It is
available in GLM Flex or in one-sample t-tests (the results will be
slightly different because GLM Flex would use the average of all
conditions, where as you only have 1 condition in each one-sample
t-test). Both are acceptable methods.
(2) If you use the global null, a test that states all tests are 0,
then you might get the conjunction A&B in the absence of A. Under the
conjunction null, the test of A in the current model MUST be
significant, for the conjunction null to exist. If you test A in the
one-sample test and then use the test of A in the repeated-measures,
you might find the conjunction null to be significant, but only
because the test of A has become significant.
Global Null: "For those people who have used the global null for inferences
about cognitive conjunctions, and simply want to qualify their
inference. An appropriate passage might be:
It should be noted that our significant conjunction does not mean
all the contrasts were individually significant (i.e., a conjunction of
significance). It simply means that the contrasts were consistently
high and jointly significant. This is equivalent to inferring one or
more effects were present." Friston 2005.
The global null simply means that there is at least one effect amongst
your tests. It also assumes the tests are independent (Nichols et al.
2005).
The conjunction null can test that all effects are significant.
>
> So, dear Donald, is there a way to retain the correct error term while being able to perform conjunction using SPM "conjunction null"?
Because the correct error term is not available in SPM, the
conjunction null cannot be performed in SPM.
Or could you elaborate a bit, as you mentioned, how to "recode the
conjunctions outside of the models"?
I simply meant that someone needs to provide the code (probably an
update of Tom's SPM99/SPM2 code) that can compute the conjunction
statistics for SPM8. Alternatively, someone could write a script to
feed two statistical maps into SPM to produce the conjunction null.
>
> Thank you very much!
>
> Best,
> Zhenhao
>
> -----
>
> Zhenhao SHI 石振昊
> Culture and Social Cognitive Neuroscience Lab
> Department of Psychology
> Peking University
> 5 Yiheyuan Road
> Beijing 100871, P.R.China
> Phone: 86 134 6655 0474
> Email: [log in to unmask]
> http://labs.psy.pku.edu.cn/CSCN_lab
>
> -----Original Message-----
> From: "MCLAREN, Donald" <[log in to unmask]>
> Sender: "SPM (Statistical Parametric Mapping)" <[log in to unmask]>
> Date: Mon, 9 Jul 2012 14:47:39
> To: <[log in to unmask]>
> Reply-To: "MCLAREN, Donald" <[log in to unmask]>
> Subject: Re: [SPM] Group conjuction analysis across three experiments
>
> SPM provided the global null and conjunction null; however, they
> require that all conditions be in the same model. Because all
> conditions are in the same model, the individual effects of each
> contrast are inflated because the error term is incorrect.
>
> If you want to use this approach accurately, you would probably need
> recode the conjunctions outside of the models.
>
> Best Regards, Donald McLaren
> =================
> D.G. McLaren, Ph.D.
> Research Fellow, Department of Neurology, Massachusetts General Hospital and
> Harvard Medical School
> Postdoctoral Research Fellow, GRECC, Bedford VA
> Website: http://www.martinos.org/~mclaren
> Office: (773) 406-2464
> =====================
> This e-mail contains CONFIDENTIAL INFORMATION which may contain PROTECTED
> HEALTHCARE INFORMATION and may also be LEGALLY PRIVILEGED and which is
> intended only for the use of the individual or entity named above. If the
> reader of the e-mail is not the intended recipient or the employee or agent
> responsible for delivering it to the intended recipient, you are hereby
> notified that you are in possession of confidential and privileged
> information. Any unauthorized use, disclosure, copying or the taking of any
> action in reliance on the contents of this information is strictly
> prohibited and may be unlawful. If you have received this e-mail
> unintentionally, please immediately notify the sender via telephone at (773)
> 406-2464 or email.
>
>
> On Mon, Jul 9, 2012 at 2:55 PM, <[log in to unmask]> wrote:
>>
>> Thanks for your reply. The conjunction method that you have recommended
>> seems to be the most accepted in the literature, but my question is
>> regarding the principle in this technique.
>>
>> In this method, we are not doing any statistical analysis to estimate the
>> degree of overlap and merely looking at the common regions. Is there any
>> method by which we can get a statistical heat map representing the degree
>> of overlap between the different conditions.
>>
>> Thanks in advance,
>>
>> Atesh
>>
>>
>>
>>
>>> (1) I would run 3 one-sample t-tests, one for each condition.
>>> (2) I would threshold each condition and save the map of significant
>>> voxels.
>>> (3) I would convert each of those to a binary image using imcalc (i1>0).
>>> (4) I would then create a combined image using imcalc (i1+2.*i2+4.*i3)
>>> from
>>> the binary value images.
>>>
>>> Voxels with a 7 are sig. in all 3 conditions, 6 are sig. in conditions
>>> 2&3,
>>> 5 are sig. in conditions 1&3, 4 is sig. in condition 3, 3 are sig. in
>>> conditions 1 and 2, 2 is sig. in condition 2, and 1 is sig. in condition
>>> 1.
>>>
>>>
>>>
>>> Best Regards, Donald McLaren
>>> =================
>>> D.G. McLaren, Ph.D.
>>> Research Fellow, Department of Neurology, Massachusetts General Hospital
>>> and
>>> Harvard Medical School
>>> Postdoctoral Research Fellow, GRECC, Bedford VA
>>> Website: http://www.martinos.org/~mclaren
>>> Office: (773) 406-2464
>>> =====================
>>> This e-mail contains CONFIDENTIAL INFORMATION which may contain PROTECTED
>>> HEALTHCARE INFORMATION and may also be LEGALLY PRIVILEGED and which is
>>> intended only for the use of the individual or entity named above. If the
>>> reader of the e-mail is not the intended recipient or the employee or
>>> agent
>>> responsible for delivering it to the intended recipient, you are hereby
>>> notified that you are in possession of confidential and privileged
>>> information. Any unauthorized use, disclosure, copying or the taking of
>>> any
>>> action in reliance on the contents of this information is strictly
>>> prohibited and may be unlawful. If you have received this e-mail
>>> unintentionally, please immediately notify the sender via telephone at
>>> (773)
>>> 406-2464 or email.
>>>
>>>
>>>
>>> On Sun, Jul 8, 2012 at 2:16 AM, Atesh Koul <[log in to unmask]> wrote:
>>>
>>>> Dear all,
>>>>
>>>> I have a fairly simple question but have got a bit confused with the
>>>> different options people have suggested on the mailing list. I want to
>>>> do
>>>> a conjunction analysis and find regions which are common in three
>>>> experiments. In my results, I want a heat map that represents the level
>>>> of
>>>> overlap between the three experiments. However, reading through some of
>>>> the approaches that people have suggested on the mailing list, I have
>>>> come
>>>> up with more than one ways to do this. I would like to know which
>>>> approach
>>>> is better and would give me correct results:
>>>>
>>>> 1. Inclusive masking: I select Results from one experiment, then use
>>>> inclusive masking using a thresholded t-map and see the regions common
>>>> in
>>>> the two experiments. Then use this t-map to mask my result from third
>>>> experiment. (In this case however, I have found that the regions depend
>>>> slightly on which experiments' results you use first and the heat map is
>>>> not an indication of extent of overlap)
>>>>
>>>> 2. Use Imcalc to mask two t-maps and get the results. (In this case as
>>>> well, heat map is not an indication of extent of overlap).
>>>>
>>>> 3. Using single contrasts: I take only the single condition contrasts
>>>> (use
>>>> a contrast vector 1 0 0 etc.) for all participants, take it to group
>>>> level
>>>> for all conditions and experiments, then run a random effects analysis
>>>> on
>>>> them (an approach I am not familiar with).
>>>>
>>>> I would highly appreciate any help in this regard.
>>>>
>>>>
>>>> Atesh Koul
>>>> Graduate student,
>>>> National Brain Research Centre, India
>>>>
>>>
>>
|