 |
 |
Dear SPMers, I'm hoping someone can educate me a bit on something I've noticed when playing around with a recent analysis; I'm sure there's a good reason for it but my understanding seems to be a bit hazy in this area.... I have a group of 36 subjects, which for the sake of the current question we will regard as a single group. In the first level design matrix for each subject we have condition 1, condition 2 (two levels of a single factor) and condition 3 (button press no-interest regressor), plus movement regressors and a baseline. I'm comparing three different ways of analysing them in SPM8:- (1) Method 1 - One-sample t-test: first-level contrast 'Condition 1 > Condition 2', which gives me one image per subject taken forward to the second level. This is entered into one-sample t-test, which gives me a t-map. (2) Method 2 - Paired-sample t-test: first-level contrasts 'Condition 1' and 'Condition 2' separately for each subject, which gives me two images per subject taken forward to the second level. These are entered into a paired-samples t-test, which gives me a t-map. (3) Method 3 - One-way within-subjects anova: first-level contrasts 'Condition 1' and 'Condition 2' separately for each subject, which gives me two images per subject taken forward to the second level. These are entered into a one-way within-subjects ANOVA, which also includes subject regressors/covariates in the way recommended in the Henson and Penny paper (and the equivalent chapter in the SPM book). Now as it happens, there is no activation for the case of 'Condition 2 > Condition 1', so my F-map for the Main Effect of Factor 1, and the t-map for Positive Effect of Factor 1, are identical. As far as I can see, and premised on there being no deactivation in Method 3, these three methods are essentially equivalent, although I've attached the design matrices for all three in case I've made a mistake that someone can tell me about. Sure enough, the t-maps obtained from all three methods are absolutely identical (I've done a voxel-wise comparison of all three to confirm this). So far so good; now here comes the fun part. If I extract the first eigenvariate statistics for a given cluster, these are completely different for all three methods, in spite of the t-images being identical. Obviously the one-sample t-test at least needs to be different because it has just 36 values fo xY.u (because of 36 images) while the paired t-test and the anova methods have 72, but I can find no obvious relationship between any of them. I've attached maps of xY.y from the three different methods for the same cluster for information purposes. So, what is the relationship between the cluster statistics obtained from these three seemingly equivalent methods? Which should I be using? I realise that Method 1 is the most common approach, but I'll be adding some further (between and within subjects) factors into the design shortly, so I'd prefer an anova-based approach if possible. Thanks, Simon ------------- Simon Durrant, PhD Neuroscience and Aphasia Research Unit, University of Manchester, Zochonis Building, Brunswick Street, Manchester. M13 9PL.