 |
 |
Dear Emmanuel, >Dear SPM list > >I'm trying to compute a conjunction analysis using SPM99b including three >studies. > >When I use an UNCORRECTED p value of 0.001, the corresponding height >threshold is T = 3.23. In the "Statitistics: volume summary table" it is >stated that both uncorrected AND corrected p values are thresholded to p < >0.000. In this case, no p value (corrected and uncorrected) exceed 0.000: >very few voxels survived to this analysis. > In the results section you are, after having picked a contrast, asked for a threshold, corrected or uncorrected, which is used to decide which voxels should be displayed in the MIPs. If you ask for an uncorrected threshold of p<0.001 voxels with a z-score above 3.09 (t-values slightly higher depending on your design) will be displayed. A conjunction is presently implemented as a minimum t-field, i.e. a single spm(t) is created from all the spm(t) maps of the conjunction, which will at each location (voxel) contain the smallest t-value among the individual spm(t) maps at that specific voxel. When you set the "uncorrected" p-level to 0.001, only voxels of the minimimum t-field with a t-value >~ 3.09 will be displayed. >When I use a CORRECTED p value of 0.05, the corresponding height threshold >is T = 2.36. In the Statitistics: volume summary table it is stated that >the corrected p value is 0.05. As a matter of fact, in this case the lower >corrected p value (as given is the volume summary) is 0.017 and much more >voxel are listed than in the former uncorrected threshold. > The corrected p-value for the minimum t-field takes into account that your statistic reflects the fact that for n independent (almost) comparisons the statistic was equal to or higher than this value every time. Think of it like this: lets say you want to test the effect of some "treatment", and instead of testing it on only one group you decide to test the "treatment" on five different groups. Lets further say that you get e.g. the t-values 2.3, 1.9, 3.1, 2.6 and 1.5. Now, a t-value of 1.5 is not terribly exciting since the probability of getting it by chance is pretty high. However, the probability of obtaining p-values of 1.5 or greater in five different tests is very small. >Thus, it seems that in the conjunction analysis the use of a corrected >threshold is more permissive than the use of an uncorrected threshold. >Note that in a single contrast analysis the p<0.05 CORRECTED threshold >leads to a T value of 4.95 while the p<0.001 UNCORRECTED value leads to a T >value of 3.23. In this latter case, the use of a corrected p value is >obviously more stringent. > >How do you explain this discrepancy ? > I hope it is explained by now. For practical purposes you should specify a relatively low p-threshold for display of your MIPs. A simple role of thumb is to use a p-value pt, such that pt^n (where n is the number of contrasts in the conjunction) is approximately the threshold you should have used for a single contrast. For example, lets say your conjunction consist of 3 contrasts and you would like to view your data at p<0.001 uncorrected, then pt=0.001^0.33=0.1. Good luck Jesper %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%