I was going over this thread and realized to use this information to correlate florbetapir PET brain images from the ADNI data with to ADAS. I see a good correlation in SUVr values from the same image with ADAS but I don't see that when I convert the T-values from the SPMT map using the r2=t2/(t2+df) formula. I see more uptake in the WM areas. Not sure what could be wrong in the set-up.

I am using MATLAD R2012B and SPM12.

 

Let me know where I may be wrong.

Abhinay

On Thu, May 29, 2014 at 8:20 AM, Colin Hawco <[log in to unmask]> wrote:

I think you can and should show the regression plot as well, which can be done with the plot function in SPM (plot against an explanatory variable). the plot is basically the data on which the statistics was run. What you shouldn't do is extract that data and run a correlation outside SPM calculating an R^2 value. 

One important aspect of plotting the data is letting readers not only see the actual data, but checking for outliers. Regressions, in my experience, can be a bit prone to inflation from outliers, especially at smaller sample sized. I have run regressions in SPM in which removing one or two participants substantially reduced the observed effects. Again, more of an issue at small sample sizes (N < 20). 

On 28 May 2014 21:42, Andy Yeung <[log in to unmask]> wrote:

Thanks for clarifications from Donald.

Do you mean when I report regression results, it's better to show 'plot -> contrast estimates and 90% CI' as that bar graph shows beta value which isn't given from the t-statistic table?

If extracting input values from the peak voxel/ cluster to make a plot cannot give readers new information, what's the point of doing that?

And am I correct to say that reporting t-statistic map from regression needs p FWE <.05 as well?

Best,

Andy

On Thu, May 29, 2014 at 12:56 AM, MCLAREN, Donald <[log in to unmask]> wrote:

See responses below.

On Tue, May 27, 2014 at 11:40 PM, Andy Yeung <[log in to unmask]> wrote:

Thanks Colin for further guidance. I've read Vul et al about your point of circularity.

Sorry I don't understand it. When we look at the t-statistic map for regression (all the glass brain and stat table from that results window), what does those stat mean? I thought:

1) t-value means how far the beta of that voxel deviates from zero.

2) Beta is the slope of regression.

3) Simply reporting t-value can only indicate how steep the slope is, but CANNOT show how well the data are fitted/ how well behavioral score predicts brain signal.

>> The steepness of the slope is the beta. The t-value indicates how well the data fits the slope. The t-statistic is what you use to make inferences about your data. You can have a high beta and low t, or a low beta and a very high t-statistic. The higher the t-statistic, the closer the points are to the line through all points (e.g. the slope line).

 

4) That's why we may extract peak voxel beta to plot correlation graph with behavioral score (from each individual) to show the fitness/ predictability in a post hoc fashion.

>> No. People extract the input values (beta/con from each subject) from the peak voxel/cluster to make a plot. Researchers incorrectly use this plot to make inferences about their data. What researchers should do is make the plot as I have described, but make all of their inferences based on the statistical maps. Then you avoid the issues of Vul. One of the points in the paper is that in small samples, you need a very high R^2 value to reach statistical significance. When you add multiple comparison correction to that issue, the R^2 of the peak will be very high. If you make inferences on these isolated values, then you may mislead the reader if you use the R^2 values; hence all inferences should be made on the statistical maps. 

Another aspect is that the t-statistic can be converted to an R^2, so if  a reviewer asks for the R^2 value, you can simply convert you peak t-statistic. There is no need to extract the values and compute the R^2 value outside of SPM.

Hope this helps.

 

Is this logic flawed?

Best,

Andy

On Wed, May 28, 2014 at 9:25 AM, Colin Hawco <[log in to unmask]> wrote:

Convert the t value from the second level spmT file from a regression. 

To get one point per person, you need to extract one value. typically it would be a Beta value from the CON images. you can use marsbar or the eigenvariate functions in results for this, but I'm not much help there since I don't use them. I generally open a con file in matlab and take out any values I want. 

best,

Colin

On 27 May 2014 20:03, Andy Yeung <[log in to unmask]> wrote:

Dear Donald and Colin,

Thanks a lot for clearing my mind. Should I use the conversion formula for con_image or for spmT_image?

And if I want to make a correlation plot for a certain ROI/ peak voxel, how can I get one dot for each individual?

Best,

Andy

On Tue, May 27, 2014 at 11:33 PM, Colin Hawco <[log in to unmask]> wrote:

it is my understanding that you can calculate an r^2 value based on a t value.

r^2 = t² / (t² + DF)

(DF is of course degress of freedom). 

I am assuming this is applicable to SPM t-scores. In tests, anyways, I have found this calculated R^2 value is close to what I get when running a correlation on the results of a regression (which is highly circular statistics and should not be done). 

Aside from running a regression in SPM, you can also post-hoc run correlations with behavioral scores using significant regions as ROIs. If you, for example, contrast A with B and find a significant cluster, you can extract beta values and see if the difference between A and B is modulated by factor X (a behavioral score or test result or symptom score or whatever). 

Just be sure that factor X is independent from A and B. It is very easy for such analysis to become circular. 

On 27 May 2014 00:33, Andy Yeung <[log in to unmask]> wrote:

Dear all,

I've noticed recent fmri papers often report correlation of Beta against behavioral scores (eg VAS). Is there a common way to perform this correlation?

Also, multiple regression needs to check interaction effects among independent variables, can it be checked by SPM interface? Is it possible to get a reading of R^2 (coefficient of determination)?

Thanks for all your help.

Regards,

Andy

--

Abhinay D. Joshi
Wayne, PA
Cellular:-817-995-3962