Limits...
Multivariate and repeated measures (MRM): A new toolbox for dependent and multimodal group-level neuroimaging data.

McFarquhar M, McKie S, Emsley R, Suckling J, Elliott R, Williams S - Neuroimage (2016)

Bottom Line: This approach has a number of drawbacks as certain designs and comparisons of interest are either not possible or complex to implement.Comparisons with existing approaches and software packages for dependent group-level neuroimaging data are made.Follow-up of these multimodal models using linear discriminant functions (LDA) is also discussed, with applications to future studies wishing to integrate multiple scanning techniques into investigating populations of interest.

View Article: PubMed Central - PubMed

Affiliation: Neuroscience & Psychiatry Unit, Stopford Building, The University of Manchester, Oxford Road, Manchester M13 9PL, UK. Electronic address: martyn.mcfarquhar@manchester.ac.uk.

No MeSH data available.


Comparisons between the restricted and unrestricted models in each of the software packages. Voxels in pink indicate those that survive thresholding in both the restricted and unrestricted models. Voxels in orange are those that survive thresholding in the restricted model only, with voxels in green showing the same for the unrestricted model. The values beneath the images indicate the number of voxels that survive thresholding in the restricted and unrestricted models. For SPM, GLM FLEX, and SwE the unrestricted models equate to estimating a unique covariance structure per-group. As MRM assumes covariance homogeneity, only the restricted results are presented.
© Copyright Policy - CC BY
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4862963&req=5

f0035: Comparisons between the restricted and unrestricted models in each of the software packages. Voxels in pink indicate those that survive thresholding in both the restricted and unrestricted models. Voxels in orange are those that survive thresholding in the restricted model only, with voxels in green showing the same for the unrestricted model. The values beneath the images indicate the number of voxels that survive thresholding in the restricted and unrestricted models. For SPM, GLM FLEX, and SwE the unrestricted models equate to estimating a unique covariance structure per-group. As MRM assumes covariance homogeneity, only the restricted results are presented.

Mentions: To further compare these approaches, we estimated the models in each of the software packages using fewer restrictions. For SPM 12 and GLM Flex, this involved setting the group variances to unequal. For SwE, this involved requesting a unique covariance matrix to be estimated for each group. As previous authors have demonstrated, assuming covariance homogeneity when the reality is heterogeneity can lead to either conservative or liberal inference (Guillaume et al., 2014). It is therefore important for researchers to realise the potential limitations of making this assumption in the multivariate GLM. Fig. 7 shows the comparisons between the models estimated earlier and those estimated with fewer assumptions. Voxels in pink indicate overlaps between the previous model and the unrestricted model. Voxels in orange indicate those found in the restricted model only, with voxels in green indicating those found in the unrestricted model only. Looking across these results, it is clear that although the number of voxels surviving thresholding do differ between the restricted and unrestricted models, these are generally fringe cases on the edges of clusters that appear irrespective of the covariance assumptions. In addition, it is also clear that assuming covariance heterogeneity generally leads to more conservative inference, and while this is preferable to overly liberal inference, it will lead to a reduction in power if homogeneity can be assumed. This appears particularly true of SwE, where the reduction in surviving voxels when covariance heterogeneity is assumed is consistently the greatest. Again, SPM and GLM FLEX appear to differ due to their implementations of the non-sphericity correction, with the SPM/GLM FLEX and MRM/SwE split still apparent. This would suggest that the biggest differentiator between these methods is not their ability to accommodate a different covariance structure per group; rather, it is their use of unique vs pooled structures across an image. As such both MRM and SwE are the preferred approaches, with SwE providing more flexibility in allowing the covariance structure to differ between groups, but seemingly losing some sensitivity, particularly in the between-subject comparisons. It is also worth noting that the multivariate approach is capable of incorporating covariance heterogeneity using approximate degrees of freedom corrections such as the Welch–James and Brown–Forsythe approaches (Keselman and Lix, 1997, Lix et al., 2003, Vallejo et al., 2001). These are, generally speaking, more complex to implement than the standard multivariate test statistics, and given that they are not widely used, they will require further investigation before applying to imaging data. Presently, assumptions of covariance homogeneity can be checked in MRM at peaks of interest, allowing researchers to caution interpretation if this assumption appears violated.


Multivariate and repeated measures (MRM): A new toolbox for dependent and multimodal group-level neuroimaging data.

McFarquhar M, McKie S, Emsley R, Suckling J, Elliott R, Williams S - Neuroimage (2016)

Comparisons between the restricted and unrestricted models in each of the software packages. Voxels in pink indicate those that survive thresholding in both the restricted and unrestricted models. Voxels in orange are those that survive thresholding in the restricted model only, with voxels in green showing the same for the unrestricted model. The values beneath the images indicate the number of voxels that survive thresholding in the restricted and unrestricted models. For SPM, GLM FLEX, and SwE the unrestricted models equate to estimating a unique covariance structure per-group. As MRM assumes covariance homogeneity, only the restricted results are presented.
© Copyright Policy - CC BY
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4862963&req=5

f0035: Comparisons between the restricted and unrestricted models in each of the software packages. Voxels in pink indicate those that survive thresholding in both the restricted and unrestricted models. Voxels in orange are those that survive thresholding in the restricted model only, with voxels in green showing the same for the unrestricted model. The values beneath the images indicate the number of voxels that survive thresholding in the restricted and unrestricted models. For SPM, GLM FLEX, and SwE the unrestricted models equate to estimating a unique covariance structure per-group. As MRM assumes covariance homogeneity, only the restricted results are presented.
Mentions: To further compare these approaches, we estimated the models in each of the software packages using fewer restrictions. For SPM 12 and GLM Flex, this involved setting the group variances to unequal. For SwE, this involved requesting a unique covariance matrix to be estimated for each group. As previous authors have demonstrated, assuming covariance homogeneity when the reality is heterogeneity can lead to either conservative or liberal inference (Guillaume et al., 2014). It is therefore important for researchers to realise the potential limitations of making this assumption in the multivariate GLM. Fig. 7 shows the comparisons between the models estimated earlier and those estimated with fewer assumptions. Voxels in pink indicate overlaps between the previous model and the unrestricted model. Voxels in orange indicate those found in the restricted model only, with voxels in green indicating those found in the unrestricted model only. Looking across these results, it is clear that although the number of voxels surviving thresholding do differ between the restricted and unrestricted models, these are generally fringe cases on the edges of clusters that appear irrespective of the covariance assumptions. In addition, it is also clear that assuming covariance heterogeneity generally leads to more conservative inference, and while this is preferable to overly liberal inference, it will lead to a reduction in power if homogeneity can be assumed. This appears particularly true of SwE, where the reduction in surviving voxels when covariance heterogeneity is assumed is consistently the greatest. Again, SPM and GLM FLEX appear to differ due to their implementations of the non-sphericity correction, with the SPM/GLM FLEX and MRM/SwE split still apparent. This would suggest that the biggest differentiator between these methods is not their ability to accommodate a different covariance structure per group; rather, it is their use of unique vs pooled structures across an image. As such both MRM and SwE are the preferred approaches, with SwE providing more flexibility in allowing the covariance structure to differ between groups, but seemingly losing some sensitivity, particularly in the between-subject comparisons. It is also worth noting that the multivariate approach is capable of incorporating covariance heterogeneity using approximate degrees of freedom corrections such as the Welch–James and Brown–Forsythe approaches (Keselman and Lix, 1997, Lix et al., 2003, Vallejo et al., 2001). These are, generally speaking, more complex to implement than the standard multivariate test statistics, and given that they are not widely used, they will require further investigation before applying to imaging data. Presently, assumptions of covariance homogeneity can be checked in MRM at peaks of interest, allowing researchers to caution interpretation if this assumption appears violated.

Bottom Line: This approach has a number of drawbacks as certain designs and comparisons of interest are either not possible or complex to implement.Comparisons with existing approaches and software packages for dependent group-level neuroimaging data are made.Follow-up of these multimodal models using linear discriminant functions (LDA) is also discussed, with applications to future studies wishing to integrate multiple scanning techniques into investigating populations of interest.

View Article: PubMed Central - PubMed

Affiliation: Neuroscience & Psychiatry Unit, Stopford Building, The University of Manchester, Oxford Road, Manchester M13 9PL, UK. Electronic address: martyn.mcfarquhar@manchester.ac.uk.

No MeSH data available.