If EV2 is empty, then usually a contrast of EV1-EV2 is not going to be
meaningful. (Having no responses available to estimate a response does
not mean that the "true" response was zero).
So yes, not only do you need to exclude single empty EVs from your higher
level analyses, but generally you'll want to exclude ANY contrasts that
involve an empty EV. This takes a little bookkeeping to keep track of.
cheers,
-MH
--
Michael Harms, Ph.D.
-----------------------------------------------------------
Conte Center for the Neuroscience of Mental Disorders
Washington University School of Medicine
Department of Psychiatry, Box 8134
660 South Euclid Ave. Tel: 314-747-6173
St. Louis, MO 63110 Email: [log in to unmask]
On 3/1/14 11:43 PM, "Rita Elena Loiotile" <[log in to unmask]> wrote:
>Hi,
>
>Sorry, I just realized that what I said in my first email is not quite
>right...
>In a fixed effects analyses (combining over a single subject's runs) I
>do get 0's for a cope that includes any 0 copes. However, in my case,
>this only happens when the original cope is a single EV (and that EV
>is empty for at least one run).
>Therefore my revised question is: Are my non-zero copes that includes
>empty EVs kosher?
>For example, if I have 2 EV's (EV1 & EV2) and EV2 is empty on some runs.
>So for my first level analyses I have the following copes:
>cope 1 (EV1): 1 0
>cope 2 (EV2): 0 1
>cope 3 (EV1>EV2): 1 -1
>cope 4: EV2>EV3): -1 1
>Since EV2 is empty on some runs, a fixed effects analysis over all
>runs will yield a zero cope-2. Therefore, I should do what was
>suggested in the previous emails.
>However, cope-3 and cope-4 will be non-zero even for runs in which EV2
>is 0. I can also do a fixed effects analysis over these 2 copes and
>get non-zero values (I'm assuming this will give me a fixed effects
>over paired,within-run differences as Michael suggested above). But
>are my results actually usable since several of the run pairs
>contained an empty EV2? Or should I do what Jeanette described above:
>a fixed effects analyses on all cope 1's and non-zero cope 2s and then
>do a 1 -1 contrast?
>
>Sorry again for the multiple emails. Just want to make sure I'm not
>unintentionally messing up some of the statistical assumptions!
>
>Thanks,
>Rita
>
>
>On Wed, Feb 26, 2014 at 6:57 PM, Jeanette Mumford
><[log in to unmask]> wrote:
>> Ah thanks, good point.
>>
>> Jeanette
>>
>>
>> On Wed, Feb 26, 2014 at 1:34 PM, Harms, Michael <[log in to unmask]>
>>wrote:
>>>
>>>
>>> Hi Jeanette,
>>> But just to be clear, in that case you're making an assumption of
>>> "stationarity" of response if you first average the runs with a usable
>>>r2,
>>> and then differencing that with the average of the runs that have a
>>>usable
>>> r1. i.e,. if the r2 response (e.g., its magnitude) is say changing
>>>across
>>> runs, then you're not really getting a true paired, within-run
>>>treatment of
>>> the r1-r2 difference. That may, or may not, be a reasonable assumption
>>> depending on the precise nature of the task.
>>>
>>> cheers,
>>> -MH
>>>
>>> --
>>> Michael Harms, Ph.D.
>>> -----------------------------------------------------------
>>> Conte Center for the Neuroscience of Mental Disorders
>>> Washington University School of Medicine
>>> Department of Psychiatry, Box 8134
>>> 660 South Euclid Ave. Tel: 314-747-6173
>>> St. Louis, MO 63110 Email: [log in to unmask]
>>>
>>> From: Jeanette Mumford <[log in to unmask]>
>>> Reply-To: FSL - FMRIB's Software Library <[log in to unmask]>
>>> Date: Wednesday, February 26, 2014 12:57 PM
>>> To: FSL - FMRIB's Software Library <[log in to unmask]>
>>> Subject: Re: [FSL] Empty EV's on Second Level Analyses
>>>
>>> Dear Rita,
>>>
>>> I wouldn't omit a run completely if only 1 EV was affected. The
>>>solution
>>> is to use the "lower level copes" option in your level 2 analysis and
>>> average *all* good lower level copes for each contrast. So, if cope1
>>>was
>>> only present for runs 1-3 and cope 2 was present for all 6 runs, you'd
>>>have
>>> 2 separate models, one with 3 inputs and one with 6. Note, this
>>>example I'm
>>> assuming you're simply interested in averaging within a copes,
>>>although it
>>> seems you have something else in mine.
>>>
>>> If you're really interested in r1-r2 at level 2, I would use the "use
>>> lower level copes" option and select 9 inputs: 3 good level 1 copes
>>>for r1
>>> and the 6 good level 1 copes for r2. Then your model would be
>>>(assuming the
>>> r1's are first, followed by r2's)
>>>
>>> 1 0
>>> 1 0
>>> 1 0
>>> 0 1
>>> 0 1
>>> 0 1
>>> 0 1
>>> 0 1
>>> 0 1
>>>
>>> and the [1 -1] contrast would yield r1-r2. Use fixed effects.
>>>
>>> There is no hit in power doing it this way. Don't concatenate your
>>>runs,
>>> as it messes up FSL's temporal autocorrelation model. This level 2
>>>model
>>> I've suggested will solve that problem.
>>>
>>> Hope that helps!
>>> Jeanette
>>>
>>>
>>>
>>> On Wed, Feb 26, 2014 at 11:55 AM, Rita Elena Loiotile
>>> <[log in to unmask]> wrote:
>>>>
>>>> Hi,
>>>> Thanks for the response. I have 2 related questions. I asked the
>>>> first question a while ago but no one responded so I'm really hoping
>>>> someone can help me! (I'll try to be clearer too.)
>>>>
>>>> 1. Even though the fixed effects analysis on all runs for one subject
>>>> is recommended over run concatenation, it seems that the current
>>>> implementation tends to leave out more information than it needs to.
>>>> For example, let's say I have 2 regressors -- r1 and r2. r1 is an
>>>> empty EV in 3 of my 6 runs for a given subject. I am interested in
>>>> the contrast r1 - r2. If I run each run separately I will have empty
>>>> COPES (r1-r2) for 3 of the 6 runs. According to the recommendations
>>>> on the listserve, I should then do a fixed effects analysis on these 3
>>>> runs to get my overall COPE (r1-r2). However, this fixed effects
>>>> analysis seems to leave out 3 perfectly valid PEs of r2 (which is
>>>> never empty) from the final COPE. That is, the resulting cope is the
>>>> average of {(r1_1 - r2_1), (r1_2 - r2_2), (r1_3 - r2_3)} where _x
>>>> indicates the xth run. Don't I really want the final COPE to be given
>>>> by: average(r1_1, r1_2, r1_3) - average (r2_1, r2_2, r2_3, r2_4, r2_5,
>>>> r2_6)?
>>>>
>>>> 2. If I do decide to concatenate runs (because of the above problem
>>>> giving me too little power) is there a preferred stage at which to
>>>> concatenate? I realize that concatenation is not recommended for a
>>>> variety of reasons but I'm wondering if those reasons tend to be less
>>>> problematic if one concatenates after some stage of preprocessing...
>>>>
>>>> Thank you,
>>>> Rita
>>>>
>>>> On Sun, Feb 23, 2014 at 5:03 PM, Harms, Michael <[log in to unmask]>
>>>>wrote:
>>>> > Hi,
>>>> > Sorry, but you've identified the work-around.
>>>> >
>>>> > cheers,
>>>> > -MH
>>>> >
>>>> > --
>>>> > Michael Harms, Ph.D.
>>>> >
>>>> > -----------------------------------------------------------
>>>> > Conte Center for the Neuroscience of Mental Disorders
>>>> > Washington University School of Medicine
>>>> > Department of Psychiatry, Box 8134
>>>> > 660 South Euclid Ave. Tel: 314-747-6173
>>>> > St. Louis, MO 63110 Email: [log in to unmask]
>>>> >
>>>> >
>>>> >
>>>> >
>>>> > On 2/23/14 2:39 PM, "Rita Elena Loiotile" <[log in to unmask]>
>>>>wrote:
>>>> >
>>>> >>Hi,
>>>> >>I'd like advice on the best workaround for this situation: I am
>>>>trying
>>>> >>to do a second level analyses--i.e., fixed effects on all the runs
>>>>for
>>>> >>a single subject. Some of these runs have a few empty EVs. If I
>>>>run
>>>> >>the second level analyses submitting all the runs, I get 0's in my
>>>> >>copes that include at least one run with one empty EV. Ideally,
>>>>what
>>>> >>I would like to do is run a second level analyses on all runs such
>>>> >>that only non-empty EVs are considered for each individual cope.
>>>> >>However, the only way I can think of doing this is by running the
>>>> >>second level analyses for each cope separately, only including the
>>>> >>runs with non-empty EVs that comprise that cope. Seems really time
>>>> >>consuming...
>>>> >>Any suggestions?
>>>> >>Thanks,
>>>> >>Rita
>>>> >
>>>> >
>>>> > ________________________________
>>>> > The materials in this message are private and may contain Protected
>>>> > Healthcare Information or other information of a sensitive nature.
>>>>If you
>>>> > are not the intended recipient, be advised that any unauthorized
>>>>use,
>>>> > disclosure, copying or the taking of any action in reliance on the
>>>>contents
>>>> > of this information is strictly prohibited. If you have received
>>>>this email
>>>> > in error, please immediately notify the sender via telephone or
>>>>return mail.
>>>
>>>
>>>
>>>
>>> ________________________________
>>>
>>> The materials in this message are private and may contain Protected
>>> Healthcare Information or other information of a sensitive nature. If
>>>you
>>> are not the intended recipient, be advised that any unauthorized use,
>>> disclosure, copying or the taking of any action in reliance on the
>>>contents
>>> of this information is strictly prohibited. If you have received this
>>>email
>>> in error, please immediately notify the sender via telephone or return
>>>mail.
>>
>>
________________________________
The materials in this message are private and may contain Protected Healthcare Information or other information of a sensitive nature. If you are not the intended recipient, be advised that any unauthorized use, disclosure, copying or the taking of any action in reliance on the contents of this information is strictly prohibited. If you have received this email in error, please immediately notify the sender via telephone or return mail.
|