Statistical analysis (SPM5)

Although there are many potential design offered in the 2nd-level analysis I recommend to use the “Full factorial” design because it covers most statistical designs. For cross-sectional VBM data you have usually 1..n samples and optionally covariates and nuisance parameters:

Number of factor levels   Number of covariates   Statistical Model
1 0 one-sample t-test
1 1 single regression
1 >1 multiple regression
2 0 two-sample t-test
>2 0 Anova
>1 >0 Ancova (for nuisance parameters) or Interaction (for covariates)

Example for cross-sectional design

The following suggestions are only valid if you have unmodulated data or data modulated for non-linear effects only (indicated by ‘m0’) and you would like remove the confounding effects due to different brain size. Please check the section on modulation for more information.

SPM interface Comment
-Full factorial Also covers regression and one/two-sample t-test
-Factors For VBM usually one factor to define your samples
Name: sample Give name like “sample”
Levels: n Enter number of samples
Independence: Yes Always ‘yes’ for cross-sectional designs
Variance: Unequal/Equal Is variance equal/unequal between samples?
Grand mean scaling: No Do not use
ANCOVA: No Do not use
-Specify Cells
Levels: 1 Level=sample (start with 1)
Scans: Select scans for sample 1 (swc* or sm0wc*)
Define for each sample
Levels: n n=number of samples
Scans: Select scans for sample n (swc* or sm0wc*)
-Covariates Define covariates and nuisance parameters
Vector Enter values for each scan/subject
Interactions Interaction sample x covariate ?
Centering: No centering You can also choose “Overall mean”
-Threshold Masking Always use abs. threshold to restrict analysis to one tissue type
Threshold: 0.1 Start with 0.1 and increase up to 0.25 to decrease search volume
Implicit Mask: Yes Redundant if you are using absolute thresholds
Explicit Mask: None Usually not neccessary if you are using absolute thresholds
-Global calculation: Omit Do not use
-Global normalisation Do not use
-Overall grand mean scaling: No Do not use
Normalisation: None Do not use