Screenshot instructions:
Windows
Mac
Red Hat Linux
Ubuntu
Click URL instructions:
Rightclick on ad, choose "Copy Link", then paste here →
(This may not be possible with some types of ads)
From: Hu, Sien <sien.hu@ya...>  20131010 00:51:20

Hi Matthew, Thank you for your explanations. If I want to do a pairsample ttest across a large sample of subjects (n>100), then it wouldn't matter if I use the contrast value or the t statistics because the sample size is large enough. However, if I have a small sample size (say, n<30）, then probably I want to use the t if I'm interested in the signal/noise, but I should use con if I want to compare just the signal. Is it correct? Sien ________________________________________ From: matthew.brett@... [matthew.brett@...] on behalf of Matthew Brett [matthewb@...] Sent: Wednesday, October 09, 2013 8:07 PM To: Hu, Sien Cc: SPM Mailing List Subject: Re: [SPM] [MarsBaR] marsS.con and Y from getdata Hi, I think you want the marsbar mailing list for this one rather than the SPM list. On Wed, Oct 9, 2013 at 12:07 PM, Hu, Sien <sien.hu@...> wrote: > Dear MarsBaR members, > > I have a question on the MarsBaR analysis. I got the barch script from the MarsBaR FAQ under 'How do I run a MarsBaR analysis in batch mode?'. The structure marsS from 'compute_contrasts' stores many things including con, stat, and p, and etc. If I want to use a value to represent subject's performance on a specific ROI, should I use the 'con' value or the 'stat' value? It depends what you are interested in. Remember the t statistic is the con value divided by the variability of the con value estimate. It's not normally distributed except for high degrees of freedom. It reflects something like the signal divided by the noise. You might prefer the con value if you want an estimate of the signal regardless of the noise. It should be normally distributed if the errors are normally distributed and the null hypothesis is true. The F statistic is always positive so it is not normally distributed; likewise the F statistic con value is always positive  so these can be hard to use for further testing. > Another related question is that, when I use the function 'getdata' on a contrast image of interest, the mean of Y is very similar to what is returned by marsS.con(contrast_number), but not completely identical. I thought that the rationale is the same for the way to get marsS from 'compute_contrasts' and the Y from 'getdata' so ideally marsS.con(contrast_number) should be identical to Y. Is this correct? Or I'm understanding it in a wrong way? > You mean you have done an SPM analysis of contrast N and have a con image, and you've done a marsbar analysis of the same contrast? They won't always be the same because marsbar will usually reestimate the autocorrelation estimates for the time series in this ROI, whereas SPM uses autocorrelation estimates from all the 'activated' brain voxels at the same time. See: http://marsbar.sourceforge.net/faq.html#howdoidoarandomeffectanalysisinmarsbar Cheers, Matthew 
From: Matthew Brett <matthew.brett@gm...>  20131010 02:23:30

Hi. On Wed, Oct 9, 2013 at 5:51 PM, Hu, Sien <sien.hu@...> wrote: > Hi Matthew, > > Thank you for your explanations. If I want to do a pairsample ttest across a large sample of subjects (n>100), then it wouldn't matter if I use the contrast value or the t statistics because the sample size is large enough. Well  they have different meanings, and the results would be different  but they are both more or less normally distributed on the null hypothesis. > However, if I have a small sample size (say, n<30）, then probably I want to use the t if I'm interested in the signal/noise, but I should use con if I want to compare just the signal. Is it correct? > Well  the t will not be normally distributed in this case, so normality assumptions will be broken  for the t  but the same thing applies in both cases  the t refers to the reliability of the signal but the con value does not. Cheers, Matthew 