On the use of global covariates in analyses of regional measures of brain structure
August 2020
In our routine analyses of regional surface area, cortical thickness, and grey matter volume, we do not control for associated global measures; namely, total surface area, mean cortical thickness, total brain volume (TBV), and intracranial volume (ICV). Here we outline the reasons for this decision with the caveat that this approach is not necessarily appropriate for all analyses, and should be carefully considered on a project-specific basis. Of note, TBV and ICV are highly correlated (r>.9) in the Dunedin Study and can thus be considered interchangeable in the context of the analytic strategies described below.
First, analyses that include a global covariate follow different assumptions and ask a fundamentally different question than those that do not. Analyses including a global covariate assume that the relative amount of brain matter in a given region is what drives differences in other phenotypes of interest, while analyses that do not include a global covariate assume that the absolute amount is more important. While it remains to be empirically demonstrated which of these assumptions more closely reflects the underlying biological mechanisms, we think the absolute amount of brain matter in a given region is often the more plausible assumption.
Further, while it is often practiced in the extant literature, the addition of a global covariate is a poor sensitivity test for a model without a global covariate (or vice versa), and treating it this way is likely to only lead to confusion. For example, a model that reveals wide-spread regional associations that “disappear” with the addition of a global covariate does not indicate that those regions are, in fact, unrelated to the outcome, but rather that they contribute to a widespread non-specific global effect.
Second, because there are often very large correlations between each regional measure and its respective global measure (see e.g., Voevodskaya et al., 2014), the inclusion of global covariates can produce often difficult-to-interpret negative associations due to this collinearity. For example, an analysis of the associations between regional surface area and IQ that controls for global surface area is likely to reveal negative associations between IQ and surface area in some regions despite widespread positive associations with (uncorrected) regional surface area.
Third, including a global measure as a covariate makes the assumption that the global measure “confounds” the association between the regional measure and the exposure or outcome of interest, and thus must be “controlled.” In most cases, however, it is difficult to imagine the mechanism by which this confounding would occur. For example, in testing the association between child abuse and amygdala volume, how would smaller total brain volume lead to both child abuse and smaller amygdala grey matter volume?
We recognize, however, that for some hypotheses we wish to test, controlling for a global measure may be warranted. For example, one scenario wherein we would want to examine relative size rather than absolute size is when there is an a priori disease process that selectively targets a brain region (at least, initially), as might be true for the hippocampus in AD. Under this scenario, we would choose to covary for total brain volume in order to determine if a specific, differential association exists between AD and hippocampal grey matter volume.
Finally, in the case of both the estimation of white matter hyperintensity (WMH) volume and brainAGE, variance in ICV is accounted for directly through the analytic procedures used to derive each measure. WMH volume is estimated after normalizing each individual FLAIR image to a common template space. In this way, individual differences in ICV are inherently adjusted for because WMH volume is estimated after warping individual images to a common space, thus removing possible influence due to variance in ICV. In the case of brainAGE, ICV is a feature that is directly included in the prediction of chronological age from multiple structural MRI measures. Thus, ICV is a part of brainAGE and is utilized in the machine-learning algorithm to the extent it helps predict how old an individual's brain "looks."
In the event that a decision is made to control for global measures, there are several different strategies. These are reviewed in O’Brien et al., 2011 and Voevodskaya et al., 2014.
Also, see Dhamala et al., 2022.
References
Dhamala E, Ooi LQR, Chen J, Kong R, Anderson KM, Chin R, Yeo BTT, Holmes AJ. Proportional intracranial volume correction differentially biases behavioral predictions across neuroanatomical features, sexes, and development. Neuroimage. 2022 Oct 15;260:119485.
O'Brien LM, Ziegler DA, Deutsch CK, Frazier JA, Herbert MR, Locascio JJ. Statistical adjustments for brain size in volumetric neuroimaging studies: some practical implications in methods. Psychiatry Res. 2011;193(2):113-122. doi:10.1016/j.pscychresns.2011.01.007
Voevodskaya O, Simmons A, Nordenskjöld R, et al. The effects of intracranial volume adjustment approaches on multiple regional MRI volumes in healthy aging and Alzheimer's disease. Front Aging Neurosci. 2014;6:264. Published 2014 Oct 7. doi:10.3389/fnagi.2014.00264