 |
 |
Hi James, There are multiple ways in which different scanners can affect the data -- signal and contrast to noise ratio, geometric distortion, field inhomogeneity, etc. Some of these differences will affect more or less certain steps of the processing, and often are difficult to predict. Some imaging modalities are more robust than others to scanner differences, and this also has an impact. When it comes to the statistical analyses using the GLM, the options are essentially: 1) Add extra EVs to account for scanner(s). 2) Assume different variances for the different scanners. Regarding (1), for two scanners, just one extra EV, coded for instance as 0 and 1, and mean-centered (make sure that the intercept is already included in the model), will take care of biases on global mean differences in signal intensity. This works for FEAT, randomise, and PALM. Regarding (2), define one variance group for each scanner. In FEAT, this means that the design will need to be "separable" (see the documentation on how to accomplish this). In PALM, this means that permutations will happen among subjects acquired within scanner only, and a statistic that is robust to different variances will be computed. For randomise, also permutations will be within scanner, but in this case, unless the number of subjects in both scanners is the same, the results will tend to be conservative. Hope this helps. All the best, Anderson On 24 June 2015 at 22:31, James Melrose <[log in to unmask]> wrote: > Hello, > > I am currently in the process of analyzing some data we just finished > collecting. During the process of our 5 year grant there was a scanner > upgrade so about 2/3 of my data is collected on one 3T scanner (24 > subjects, 3 separate days for each subject), and the other 1/3 is collected > on a different 3T scanner (12 subjects, again 3 days for each). The tasks > and sequences themselves are identical across the scanners, but my question > regards if it is recommended to attempt to statistically control for having > multiple scanners? I reasoned that including two extra EVs in my second > level analysis (or perhaps my 3rd level?) one that spanned all subjects on > one scanner and the other spanning all subjects on the other scanner this > would perhaps pull out any scanner specific variance. Is it recommended to > do something of this nature if an analysis with multiple scanners is > unavoidable? > > Your guidance would be greatly appreciated! > > Best, > James >