Reporting for transparency; bug will be fixed intrinsically in #1693, where the investigation of new command fixelfilter highlighted an issue with master but not the development code.

Issue was introduced in 9248b5b; specifically this line.

Normalisation of the smoothing matrix by the sum of smoothing kernel weights is looped by value rather than reference, so smoothing matrix data remain un-normalised. As a result, input data are scaled during smoothing, with the scaling factor being the sum of smoothing matrix weights for each individual fixel. As this scaling will be identical for all subjects, results of statistical inference should not be affected. It will however influence the interpretation of any direct utilisation of the beta coefficients, which will all be correspondingly scaled by this per-fixel factor. Percentage effect calculations should not be affected if beta coefficients are used for both the group difference and the control population mean, but would be erroneously scaled if mrmath were used to obtain the mean but a beta coefficient used for the difference.