I have a longitudinal data set and age is a variable of interest in our study. Documentation states:
We strongly recommend you that you check the scale factors applied during intensity normalisation are not influenced by the variable of interest in your study.
I tested if scale factors differed with age and in fact they did, although I am under the impression the same scanning protocol was used for all samples in the study.
In this a problem? In this case would it make sense to normalize by the data from the youngest in the study?
What this suggests is that the WM b=0 signal is age-dependent – which I guess isn’t too surprising, depending on the age range we’re talking about (I expect T2 likely increases with age?).
The normalisation performed by dwiintensitynorm assumes that the WM b=0 signal is an independent measure of the overall signal, and uncorrelated with your effect of interest. Clearly, this assumption will not always hold, depending on pathology, so it’s important to verify whether it holds – hence the suggestion you refer to.
The fact that your normalisation factors change with age means that your CSD-derived metric (e.g. AFD) will be corrected for that age-dependence – but since the normalisation will divide by the scale factor, it’ll apply the inverse relationship in your metric. In most cases, I’d expect a reduction in AFD to be accompanied by an increase in the b=0 signal (e.g. due to increased extracellular content, etc). So in this case, normalising by the WM b=0 signal (which is what dwiintensittnormessentially does) will actually amplify the age dependence of the AFD metric. Obviously, the relationship between the b=0 signal and AFD is going to depend on many factors, including the pathology, etc. It might mask the effect of interest, or amplify it, or not affect it at all. But we can’t predict what this correction might do on your data. This is why we issue this warning.
Intensity normalisation is difficult to do well, but we currently recommend using mtnormalise to do this, rather than dwiintensitynorm. It avoids many of these issues, and also provides more robust bias field correction. It does require the use of multi-tissue CSD though, but if you’re in a position to use it, I’d strongly recommend you do so.
Thank you so much for your very informative reply. Unfortunately we do not have multishell data, and please correct me if I am wrong, but I think this precludes our use of mtnormalise. Just as a follow up, if we are interested in age related changes and don’t want to bias the AFD metric by amplifying age-related change, would it be reasonable to normalize the whole dataset using a scale factor derived from dwiintensitynorm applied only to subjects from the first wave of our data collection (the youngest group). Would that be a reasonable approach?
Yes and no. Even a single-shell data will normally have a few b=0 volumes, and these can be used as an additional ‘shell’ to perform a 2-tissue decomposition. While this is far from perfect, you can use this with mtnormalise, and the results seem pretty good, as far as we can tell. I’d expect the intensity normalisation to be more appropriate this way.
Nope, I’m afraid not. The need for intensity normalisation comes from the fact that different datasets will be scaled differently by the scanner – even for the same subject on the same scanner – depending on the scanner’s calibrations and other factors. It’s this arbitrary scaling of individual datasets that we need to account for, and by its very nature, that scaling needs to be estimated for each subject independently. It can’t be estimated from one (group of) subject(s) and applied to another (group of) subject(s).