Subject: | | Re: Unequal sample size and power loss |
From: | | Roberto Viviani <[log in to unmask]> |
Reply-To: | | [log in to unmask][log in to unmask]> wrote:
> Dear Jose, > > thank you for your help. it worked! > > cheers, > HL > > > On 5 July 2010 16:15, jmanjon <[log in to unmask]> wrote: > >> Dear HweeLing, >> >> The problem is that the noise estimation function "cg_noise_estimation.m" >> requieres a lot of memory for the 3D wavelet decomposition. >> >> Obviously, the best solution is to increase your RAM memory but it this is >> not an alternative for you I can suggest you a small trick that can do the >> job. >> >> as the noise is supposed to be stationary you can estimate the noise from >> a portion of you data instead of the full dataset. >> >> you can add this to lines to line 28 in "cg_noise_estimation.m" to process >> only the 50% of your data >> >> ima=ima(:,:,round(0.25*s(3)):round(0.75*s(3))); >> s = size(ima); >> >> problably this will fix your problem >> >> hope this helps >> >> Jose Manjon >> >> >> HweeLing Lee escribió: >> >> Dear all, >>> >>> I'm encountering this error message whenever I tried to run the vbm8 >>> toolbox "estimate and write": >>> >>> Running 'VBM8: Estimate & Write' >>> Initial Coarse Affine Registration.. >>> Fine Affine Registration.. >>> VBM8 Revision 343 >>> Failed 'VBM8: Estimate & Write' >>> Out of memory. Type HELP MEMORY for your options. >>> In file "C:\MATLAB\spm8\toolbox\vbm8\cshift3D.m" (???), function >>> "cshift3D" at line 23. >>> In file "C:\MATLAB\spm8\toolbox\vbm8\afb3D.m" (???), function "afb3D_A" >>> at line 95. >>> In file "C:\MATLAB\spm8\toolbox\vbm8\afb3D.m" (???), function "afb3D" at >>> line 40. >>> In file "C:\MATLAB\spm8\toolbox\vbm8\dwt3D.m" (???), function "dwt3D" at >>> line 27. >>> In file "C:\MATLAB\spm8\toolbox\vbm8\cg_noise_estimation.m" (???), >>> function "cg_noise_estimation" at line 39. >>> In file "C:\MATLAB\spm8\toolbox\vbm8\cg_vbm8_write.m" (v342), function >>> "cg_vbm8_write" at line 247. >>> In file "C:\MATLAB\spm8\toolbox\vbm8\cg_vbm8_run.m" (v331), function >>> "run_job" at line 228. >>> In file "C:\MATLAB\spm8\toolbox\vbm8\cg_vbm8_run.m" (v331), function >>> "cg_vbm8_run" at line 89. >>> >>> The following modules did not run: >>> Failed: VBM8: Estimate & Write >>> >>> Could someone please advise? Thank you. >>> >>> -- >>> Best wishes, >>> HweeLing >>> >> >> >> -- >> ________________________________________________________ >> >> Dr. Jose V. Manjón Herrera ------- [log in to unmask] >> home: http://personales.upv.es/jmanjon >> [IBIME - Biomedical Imaging Area] >> BET - Bioengineering, Electronics and Telemedicine Group >> UPV - Politechnical University of Valencia - Spain >> ________________________________________________________ >> >> > > > -- > Best wishes, > HweeLing >
-- Best wishes, [log in to unmask] |
Date: | | Sat, 24 Jul 2010 09:53:28 +0200 |
Content-Type: | | text/plain |
Parts/Attachments: |
|
|
|
|
Dear Steven,
the contribution of the design X to the power of your contrast c
depends on c*inv(X'X)*c', and that's not hard to look up. You can
check exactly the effect of unbalancedness by setting up a
hypothetical balanced experiment.
Apart from the algebra, I would guess that the main reason for your
test to fail to reach significance is the need to test an interaction,
not the unbalancedness or the addition of covariates. Your degree of
unbalancedness is modest, and covariates only a cost in df's. Besides,
I cannot imagine how you could change things such as the number of
subjects you have or the need to control for sex. Rather, if you have
an unsurmountable problem here, it is that the interaction is
inherently estimated with less precision than a main effect.
Having been involved in the analysis of a pre-post treatment
neuroimaging study recently, I came to realize how nuanced is the
issue of control for error rates. In certain studies, the only chance
we have to get an effect that survives strong control is for this
effect to be an artefact. This is because the effect size that would
be required to survive correction in an interaction (for example,
interaction task-control X pre-post X patient-healthy) is
physiologically unplausible for sample sizes that can be collected in
practice.
(There is an urgent need for methodologists to develop measures of
power in the multiple testing situation to clarify this issue -- not
necessarily to prescribe sample sizes in grants)
Hence, I believe that in a situation like this there are two choices:
either report the data with uncorrected levels, or figure out a
strategy to localize ROIs a priori (for example, as a replication of
previous studies, or through the identification of clusters in which
the main effect can be demonstrated).
The message I would like to send is the following: down with all these
studies with inference at uncorrected levels that could be easily be
conducted with larger sample sizes -- typically, cognitive
neuroimaging on healthy subjects, now a healthy industry getting away
with providing no inference at all; up with those studies who report
uncorrected levels that cannot possibly be made to survive correction.
Roberto Viviani
Dept. of Psychiatry
University of Ulm, Germany
Quoting Steven Berman <[log in to unmask]>:
> Dear SPM List,
> In a post vs pre treatment FDG study, I find highly significant
> relative activity decreases in all the expected structures in the
> experimental group at p<.001 using voxel or cluster criteria, with no
> voxels decreasing at all in the placebo group. Sounds lovely, right?
> However, the interaction of group with scan did not attain
> significance anywhere. I suspect this is due to reduced power from
> three design problems I inherited. First, there were two scanners used
> so I had to include scanner as a covariate. Second, the 2 groups were
> not well age matched, so I had to include age as a covariate. Third,
> there were 15 usable subjects in the experimental group, but only 10
> in the placebo group. I think that I need to modify my contrasts to
> account for the unequal sample size, but I’m not sure how to do so.
> Any other suggestions?
> Thanks,
>
> Steven Berman, Ph.D.
> UCLA Dept. of Psychiatry & Biobehavioral Sciences,
> Brain Research Institute, and Center for Addictive Behaviors
> Tel. (310) 825-0616 Fax: (310) 825-0812
>
> http://ibs.med.ucla.edu
> http://uclamindbody.org
> http://www.bri.ucla.edu[1]
> http://www.semel.ucla.edu/cab/[2]
>
> Links:
> ------
> [1] http://www.bri.ucla.edu/
> [2] http://www.semel.ucla.edu/cab/
>
|
|
|
|