P-values

Determine p-values.

fmri.pvalue(spm, mode="basic", na.rm=FALSE, minimum.signal = 0, alpha= 0.05)

Arguments

spm	`fmrispm` object
mode	type of pvalue definition
na.rm	`na.rm` specifies how NA's in the SPM are handled. NA's may occur in voxel where the time series information did not allow for estimating parameters and their variances or where the time series information where constant over time. A high (1e19) value of the variance and a parameter of 0 are used to characterize NA's. If `na.rm=TRUE` the pvalue for the corresponding voxels is set to 1. Otehrwise pvalues are assigned according to the information found in the SPM at the voxel.
minimum.signal	allows to specify a (positive) minimum value for detected signals. If minimum.signal >0 the thresholds are to conservative, this case needs further improvements.
alpha	Significance level in case of `mode="FDR"`

Details

If only a contrast is given in spm, we simply use a t-statistic and define p-values according to random field theory for the resulting gaussian field (sufficiently large number of df - see ref.). If spm is a vector of length larger than one for each voxel, a chisq field is calculated and evaluated (see Worsley and Taylor (2006)). If delta is given, a cone statistics is used.

The parameter mode allows for different kinds of p-value calculation. mode="voxelwise" refers to voxelwise tests while mode="Bonferroni" adjusts the significance level for multiple testing. An alternative is mode="FDR" specifying signal detection by False Discovery Rate (FDR) with proportion of false positives level specified by alpha. The other choices apply results on excursion sets of random fields (Worsley 1994, Adler 2003) for smoothed SPM's. "basic" corresponds to a global definition of the resel counts based on the amount of smoothness achieved by an equivalent Gaussian filter. The propagation condition ensures, that under the hypothesis $$\hat{\Theta} = 0$$ adaptive smoothing performs like a non adaptive filter with the same kernel function which justifies this approach. "local" corresponds to a more conservative setting, where the p-value is derived from the estimated local resel counts that has been achieved by adaptive smoothing. In contrast to "basic", "global" takes a global median to adjust for the randomness of the weighting scheme generated by adaptive smoothing. "global" and "local" are more conservative than "basic", that is, they generate sligthly larger p-values.

Value

Object with class attributes "fmripvalue" and "fmridata"

pvalue

p-value. use with plot for thresholding.

weights

voxelsize ratio

dim

data dimension

hrf

expected BOLD response for contrast (single stimulus only)

alpha

maximal pvalue as scale information

thresh

actual threshold used

References

Polzehl, J. and Tabelow, K. (2007) fmri: A Package for Analyzing fmri Data, R News, 7:13-17 .

Tabelow, K., Polzehl, J., Voss, H.U., and Spokoiny, V. (2006). Analysing fMRI experiments with structure adaptive smoothing procedures, NeuroImage, 33:55-62.

Worsley, K.J., and Taylor, J.E., Detecting fMRI activation allowing for unknown latency of the hemodynamic response, NeuroImage 29:649-654 (2006).

Author

Karsten Tabelow tabelow@wias-berlin.de

Note

Unexpected side effects may occur if spm does not meet the requirements, especially if a parameter estimate vector of length greater than 2 through argument vvector in fmri.lm has beeen produced for every voxel.

Examples

if (FALSE) fmri.pvalue(smoothresult)