On 2015 Apr 22, Clyde Francks commented:

Study authors Tulio Guadalupe and Clyde Francks reply to comment by Dorothy Bishop:

We thank Dorothy Bishop for insightful comments. We considered the HO measurement of PT grey matter volume as a region-of-interest index 'within and around the human planum temporale', rather than a direct measurement of its neuroanatomical definition. The HO measurement was weighted on the voxels that most probably belonged to the PT, based on 37 brains used in constructing the atlas. The HO definition of PT is lateralized to the left and the measurement of PT regional asymmetry in our datasets reflected this. We agree that this spatial restriction probably contributed to the lower inter-subject variability measured with the HO approach compared to FreeSurfer's parcellations. However, because of the large variability in landmarks in the superior temporal lobe, these regions are among the least reliable in Freesurfer parcellations, and also the Freesurfer-Destrieux definition of the planum temporale includes the planum parietale which is not usually recognized as part of its extent. Reassuringly, we found PT regional lateralization to be sexually dimorphic with both approaches (PT region was the most sexually dimorphic of all 44 regions using the HO atlas and the third most dimorphic of 74 regions using the FreeSurfer-Destrieux atlas). We agree that the measurement issues do not obviously explain the sex effect.

Because of the normalization pre-processing of the GM maps to the MNI template, subjects with departures from average PT lateralization are already somewhat 'pre-fitted' for the HO probabilistic atlas before it is applied (and we saw no major departures based on the random thirty normalized subjects we visually inspected, which would have been expected to contain three or four subjects with rightward PT lateralization). If normalization was perfect, someone with a larger-than-average right PT in relation to that hemisphere would have their right PT morphed into an equivalent normalized space as someone with a smaller than average right PT, prior to atlas application.

Our additional requirement with our HO PT lateralization index was to support genome-wide association meta-analysis across multiple datasets, for which a single and comparable index was required across datasets, after which individual genetic associations would be further interrogated in a voxel-based-morphometry context without use of the HO atlas. We argued that, for measuring group and individual differences, regional identification was likely to be more accurate with an asymmetrical atlas than with a left-right symmetrized atlas, for structures that were asymmetrical both in the atlas and, on average, in our datasets. Using a symmetrized atlas would affect the mean and range of lateralization in the dataset, and probably allow some individuals to be measured with rightward lateralization, but then the fit would be worse for people with leftward lateralization (the majority) than for the un-symmetrized HO atlas.

Genomewide association analysis requires thousands of participants to achieve sufficient statistical power to detect the effects of individual polymorphisms that are individually expected to account for a fraction of 1% of trait variation. Even the sex difference that we found only explained 1.2-1.6% of trait variance. The use of large and multiple datasets was required within a single collaborative study, and the creation of dataset-specific atlases, based on large numbers of participants, by each of the contributing teams matched to their particular scanning setup and population demographic, was not practical. We agree that there is an urgent need for manually created brain atlases based on larger numbers of participants, together with improved methods of automated application that are robust to dataset heterogeneities and flexible for the full range of individual differences.

On 2015 Apr 18, Dorothy V M Bishop commented:

The literature on sex differences in cerebral asymmetry does appear to be very confusing, and this paper does a thorough job in attempting to address the inconsistencies. I'm not sure it completely resolves the issue, but it sets a high methodological standard by using much larger sample sizes than previous studies and including two replication samples. It also includes a genetic analysis, which I won't discuss here.

Structural asymmetry of the planum temporale, with larger PT on the left, was described 47 years ago by Geschwind and Levitsky. Since that time it has become clear that the proportion of the population showing asymmetry varies substantially according to how it is measured. There have been claims of sex differences in PT asymmetry, but, as noted by Guadalupe et al, there have also been counterclaims, and a recent meta-analysis concluded that there is no sex difference, and the impression of one arose because of publication bias.

Guadalupe et al, however, found a reliable sex difference in PT asymmetry that replicates across three samples. However, their method of measurement does have one characteristic that I found puzzling: in a sample of 2337 adults they found nobody with reversed asymmetry (i.e., R>L planum). This contrasts with the original sample of Geschwind and Levitsky, where 11% had reversed asymmetry, and 24% had equal size PT on left and right. G&L's study was based on examination of post mortem specimens, but Watkins et al (2001) found similar proportions in a voxel-based analysis of grey matter from 142 MRI scans.

I assume the finding of universal left-biased PT in the current study had to do with the method by which this was defined: this was different from the method used by Watkins et al.

In the current study, grey matter was quantified in cortical regions based on a template taken from a average brain, then weighting each voxel by the probability that it belonged to that specific cortical region. The authors note that there are regional asymmetries in the atlas they used that will necessarily influence the mean – presumably, the left planum map is bigger than the right planum map, and so there are more voxels to count on the left. The authors argue that we already know that the left and right perisylvian regions differ systematically in their anatomy on average, and the interest here is in individual differences. They argued therefore that an averaged left-right template would not capture the systematic differences between the two sides.

I have trouble understanding how the template they used could adequately capture cases where someone had a larger than average PT. If the template represents the PT from an average brain, there will be variation around that average, with some people have larger and others having smaller PTs. If voxels were counted if they were within the template and not otherwise, anyone who had a PT that actually extended beyond the template would not have all their voxels counted. I appreciate that in practice, a more sophisticated probabilistic approach was used, but I'm not sure this would solve the problem. The template is based on 37 brains; we know that around 11% of people have R>L planum temporale, so around 3-4 contributors to the template would have that pattern. So, if I understand it correctly, voxels in the outskirts of a core region would have a low weighting, because in only a small proportion of the template subjects would this region be included in PT. But this means that for a new subject who genuinely had PT in that region, the amuont of grey matter in PT would be underestimated, because the weighting would be low. And there may be some people with rare PT configurations who may have parts of their PT not represented at all in the template. In this regard, given the known individual variation in PT, 37 people seems rather small a number on which to base probabilistic weightings.

I wondered whether this aspects of methodology explained why the volume measures for the Freesurfer analysis had standard deviations that were about 1.7 times as large as the standard deviations for the HO atlas analysis – is the HO atlas analysis in effect minimising or excluding values for those whose PTs are more extensive than the average?

I can't see an obvious reason why this should generate spurious sex differences, but I wondered if it could explain the absence of cases of reversed asymmetry in this sample.

Overall, I thank the authors for providing such comprehensive data on this vexed topic; it is remarkable how the measurement of asymmetry, which looks on the surface like such a simple issue, is exceedingly complex, with specific analytic choices leading to different patterns of findings. It is good to see examples like this where alternative approaches are used to give one the opportunity to observe their effects.

Geschwind, N., & Levitsky, W. (1968). Human brain: left-right asymmetries in temporal speech region. Science, 161, 186-187. Watkins, K. E., Paus, T., Lerch, J. P., Zijdenbos, A., Collins, D. L., Neelin, P., . . . Evans, A. C. (2001). Structural asymmetries in the human brain: a voxel-based statistical analysis of 142 MRI scans. Cerebral Cortex, 11, 868-877.

On 2015 Apr 18, Dorothy V M Bishop commented:

The literature on sex differences in cerebral asymmetry does appear to be very confusing, and this paper does a thorough job in attempting to address the inconsistencies. I'm not sure it completely resolves the issue, but it sets a high methodological standard by using much larger sample sizes than previous studies and including two replication samples. It also includes a genetic analysis, which I won't discuss here.

Structural asymmetry of the planum temporale, with larger PT on the left, was described 47 years ago by Geschwind and Levitsky. Since that time it has become clear that the proportion of the population showing asymmetry varies substantially according to how it is measured. There have been claims of sex differences in PT asymmetry, but, as noted by Guadalupe et al, there have also been counterclaims, and a recent meta-analysis concluded that there is no sex difference, and the impression of one arose because of publication bias.

Guadalupe et al, however, found a reliable sex difference in PT asymmetry that replicates across three samples. However, their method of measurement does have one characteristic that I found puzzling: in a sample of 2337 adults they found nobody with reversed asymmetry (i.e., R>L planum). This contrasts with the original sample of Geschwind and Levitsky, where 11% had reversed asymmetry, and 24% had equal size PT on left and right. G&L's study was based on examination of post mortem specimens, but Watkins et al (2001) found similar proportions in a voxel-based analysis of grey matter from 142 MRI scans.

I assume the finding of universal left-biased PT in the current study had to do with the method by which this was defined: this was different from the method used by Watkins et al.

In the current study, grey matter was quantified in cortical regions based on a template taken from a average brain, then weighting each voxel by the probability that it belonged to that specific cortical region. The authors note that there are regional asymmetries in the atlas they used that will necessarily influence the mean – presumably, the left planum map is bigger than the right planum map, and so there are more voxels to count on the left. The authors argue that we already know that the left and right perisylvian regions differ systematically in their anatomy on average, and the interest here is in individual differences. They argued therefore that an averaged left-right template would not capture the systematic differences between the two sides.

I have trouble understanding how the template they used could adequately capture cases where someone had a larger than average PT. If the template represents the PT from an average brain, there will be variation around that average, with some people have larger and others having smaller PTs. If voxels were counted if they were within the template and not otherwise, anyone who had a PT that actually extended beyond the template would not have all their voxels counted. I appreciate that in practice, a more sophisticated probabilistic approach was used, but I'm not sure this would solve the problem. The template is based on 37 brains; we know that around 11% of people have R>L planum temporale, so around 3-4 contributors to the template would have that pattern. So, if I understand it correctly, voxels in the outskirts of a core region would have a low weighting, because in only a small proportion of the template subjects would this region be included in PT. But this means that for a new subject who genuinely had PT in that region, the amuont of grey matter in PT would be underestimated, because the weighting would be low. And there may be some people with rare PT configurations who may have parts of their PT not represented at all in the template. In this regard, given the known individual variation in PT, 37 people seems rather small a number on which to base probabilistic weightings.

I wondered whether this aspects of methodology explained why the volume measures for the Freesurfer analysis had standard deviations that were about 1.7 times as large as the standard deviations for the HO atlas analysis – is the HO atlas analysis in effect minimising or excluding values for those whose PTs are more extensive than the average?

I can't see an obvious reason why this should generate spurious sex differences, but I wondered if it could explain the absence of cases of reversed asymmetry in this sample.

Overall, I thank the authors for providing such comprehensive data on this vexed topic; it is remarkable how the measurement of asymmetry, which looks on the surface like such a simple issue, is exceedingly complex, with specific analytic choices leading to different patterns of findings. It is good to see examples like this where alternative approaches are used to give one the opportunity to observe their effects.

Geschwind, N., & Levitsky, W. (1968). Human brain: left-right asymmetries in temporal speech region. Science, 161, 186-187. Watkins, K. E., Paus, T., Lerch, J. P., Zijdenbos, A., Collins, D. L., Neelin, P., . . . Evans, A. C. (2001). Structural asymmetries in the human brain: a voxel-based statistical analysis of 142 MRI scans. Cerebral Cortex, 11, 868-877.

On 2015 Apr 22, Clyde Francks commented:

Study authors Tulio Guadalupe and Clyde Francks reply to comment by Dorothy Bishop:

We thank Dorothy Bishop for insightful comments. We considered the HO measurement of PT grey matter volume as a region-of-interest index 'within and around the human planum temporale', rather than a direct measurement of its neuroanatomical definition. The HO measurement was weighted on the voxels that most probably belonged to the PT, based on 37 brains used in constructing the atlas. The HO definition of PT is lateralized to the left and the measurement of PT regional asymmetry in our datasets reflected this. We agree that this spatial restriction probably contributed to the lower inter-subject variability measured with the HO approach compared to FreeSurfer's parcellations. However, because of the large variability in landmarks in the superior temporal lobe, these regions are among the least reliable in Freesurfer parcellations, and also the Freesurfer-Destrieux definition of the planum temporale includes the planum parietale which is not usually recognized as part of its extent. Reassuringly, we found PT regional lateralization to be sexually dimorphic with both approaches (PT region was the most sexually dimorphic of all 44 regions using the HO atlas and the third most dimorphic of 74 regions using the FreeSurfer-Destrieux atlas). We agree that the measurement issues do not obviously explain the sex effect.

Because of the normalization pre-processing of the GM maps to the MNI template, subjects with departures from average PT lateralization are already somewhat 'pre-fitted' for the HO probabilistic atlas before it is applied (and we saw no major departures based on the random thirty normalized subjects we visually inspected, which would have been expected to contain three or four subjects with rightward PT lateralization). If normalization was perfect, someone with a larger-than-average right PT in relation to that hemisphere would have their right PT morphed into an equivalent normalized space as someone with a smaller than average right PT, prior to atlas application.

Our additional requirement with our HO PT lateralization index was to support genome-wide association meta-analysis across multiple datasets, for which a single and comparable index was required across datasets, after which individual genetic associations would be further interrogated in a voxel-based-morphometry context without use of the HO atlas. We argued that, for measuring group and individual differences, regional identification was likely to be more accurate with an asymmetrical atlas than with a left-right symmetrized atlas, for structures that were asymmetrical both in the atlas and, on average, in our datasets. Using a symmetrized atlas would affect the mean and range of lateralization in the dataset, and probably allow some individuals to be measured with rightward lateralization, but then the fit would be worse for people with leftward lateralization (the majority) than for the un-symmetrized HO atlas.

Genomewide association analysis requires thousands of participants to achieve sufficient statistical power to detect the effects of individual polymorphisms that are individually expected to account for a fraction of 1% of trait variation. Even the sex difference that we found only explained 1.2-1.6% of trait variance. The use of large and multiple datasets was required within a single collaborative study, and the creation of dataset-specific atlases, based on large numbers of participants, by each of the contributing teams matched to their particular scanning setup and population demographic, was not practical. We agree that there is an urgent need for manually created brain atlases based on larger numbers of participants, together with improved methods of automated application that are robust to dataset heterogeneities and flexible for the full range of individual differences.