RBA scaling/standardizing

AFNI Message Board
1

Hi there,

I'm wondering what the suggested procedure is for using RBA on measures of different scales. For example, group-wise volume differences when across-region volumes can vary largely; or even, assessing differences in gray matter structure/white matter microstructure simultaneously.

Thank you!

2

group-wise volume differences when across-region volumes can vary largely; or even, assessing differences in gray matter structure/white matter microstructure simultaneously.

Could you provide some context? What is the response or outcome variable? Are you comparing the response variable across different groups at various brain regions? Do you hypothesize that region size or gray/white matter microstructure might impact the response variable?

Gang Chen

3

Sure. Consider a morphometry study between two groups, say TBI patients vs uninjured controls.

Response variable: Volume of each brain region. Predictor variable: group assignment (TBI/control). We expect TBI volumes are lower than controls.

In essence it is a t-test between each group at each ROI. Under multiple-univariate approach there is no problem because each model is independent and the response variable is on the same scale (due to being same brain region). In the RBA approach, we're combining all regions to one response variable ("volume"). Does the difference in expected volumes across brain regions (brain stem vs thalamus vs amygdala) matter? Should we normalize within brain regions first?

Thank you for a great tool and statistical framework. I believe the BML/RBA approach will be incredibly fruitful in neurology, where we expect many regions may be subtly affected by pathology.

Daniel

4

Daniel,

Does the difference in expected volumes across brain regions (brain stem vs thalamus vs amygdala) matter? Should we normalize within brain regions first?

How many regions are involved in this project? I would be hesitant to standardize region volumes because maintaining the physical meaning of those values is crucial for group comparisons and result interpretation. The cross-region differences in volumes are not necessarily problematic. As long as the volumes within each group approximately follow a centralized distribution, such as Gaussian or Student's t-distribution, it should be sufficient to use the hierarchical modeling approach implemented in RBA.

It would be great it if you could share how this method works for your data.

Gang Chen

5

Great, thanks!

This begs the question whether I can include different types of measures within the same RBA analysis.

For example, temporal lobe cortical thickness and temporal white matter fractional anisotropy which have very different scales but are themselves normally distributed. We expect such concomitant effects often and it would be interesting to model them together.

Daniel

6

Daniel,

Do you mean a bivariate or multivariate version of RBA? That would be a nice example to extend the hierarchical modeling approach. If you're interested, contact me via email as I'd like to explore the modeling capabilities further.

Gang Chen

7

i believe the model would still be univariate, as each region has only one neuroimaging outcome. However, the modality of that outcome might change across regions, such that cortical-thickness is used for cortical regions, volume is used for subcortical regions, and FA might be used for white-matter regions (all within one model, theoretically).

Perhaps as there's a region-wise intercept that it doesn't matter? I'm still not sure about how the response variable should be normalized given varying scales of this measure.

8

If I understand correctly, the three metrics (cortical thickness, volume, and FA) are measured in entirely different brain regions. Consequently, there is no shared information across these metrics and regions. Therefore, it would not be beneficial to adopt an overall hierarchical model for all the metrics/regions combined. Instead, analyze each metric separately among their respective regions.

Gang Chen