I am trying to investigate Glx-GABA+ balance, but seeing inconsistent results depending on whether I use the tissue-corrected values or not.
When I use Osprey’s raw water scaled values I see a significant correlation between Glx and GABA+ (r = 0.29, p = 0.002). When I use the tissue-corrected values, the Glx-GABA+ correlation is completely gone (r = 0.1, p = 0.14).
Should I be skeptical of using tissue-corrected values for any reason? I am also incorporating the 2024 age relaxometry atlas using flags in my job files since i have subjects that range 18 to 85 (n=87). EIbalance_question.pdf (183.0 KB)
I assume the segmentation outcomes (tissue fractions) look reasonable?
With that age range, I’d expect a fairly clear trend for GM fraction, which might be driving the correlations – but unless you’re looking at alpha-corrected values I wouldn’t expect the correction to completely remove this (ref, amonst others: Impact of tissue correction strategy on GABA-edited MRS findings - PubMed).
I don’t see any other reason to be skeptical, but would consider either using alpha-correction or adding age/GM fraction as covariates.
Thanks!! Yes the tissue fraction estimates look reasonable.
I guess I am mostly confused why the correlation relationship would change if the same tissue correction if being applied during quantification to both GABA+ and Glx? Shouldn’t I see a scaled value for both gaba+ and glx but not a change in the GABA+ - Glx relationship when switching from water scaled to tissue corrected values? Does Osprey use the same T1 and T2 values for GABA+ and Glx?
Like all MRS metabolite data, your Glx values and your GABA values are ratios with a common denominator. In the situation your describe, they have either raw water signal or tissue-corrected water signal as their common denominator. The reason that the apparent correlation goes down is probably because the coefficient of variation of the denominator is lower when you use the tissue corrected than the raw water values.
If you look at the correlations between a different pairing of metabolite values, such as the correlation between GABA and NAA, it should also be more positive when raw water is used as the denominator than when tissue-corrected water is used as the denominator.
Karl Pearson warned against trying to calculate correlations between ratios with the same denominator back in 1897. He called the result “spurious correlation,” and he showed that the magnitude of the illusory positive correlation was a function of the coefficient of variation of the common denominator - higher when the coefficient of variation is higher.
The implications of spurious correlations between ratios with the same denominator is that if you try to test the statistical significance of the r value, you cannot compare it to a value of zero as the null hypothesis. Sadly, in MRS studies, it’s not possible to know precisely what the apparent r value will be when the two metabolites in the numerators (e.g. Glx and GABA) in reality have zero correlation. The spurious part of the correlation will always have some positive value. I’ve written two short articles about this problem (below), but it hasn’t caught on yet in the MRS community. The key point is that two ratios with a common denominator cannot be treated as statistically independent values in correlational analyses.
Hopefully, knowing the true correlation between Glx and GABA is not critical to your project.
Pearson K. On a Form of Spurious Correlation which May Arise when Indices are Used in the Measurement of Organs. Proceedings of the Royal Society of London. 1897;60:489-498.
Maddock R. The problem of spurious correlations between pairs of brain metabolite values measured in the same voxel with magnetic resonance spectroscopy. JAMA Psychiatry 2014; 71:338-9. doi: 10.1001/jamapsychiatry.2013.4343.
Maddock RJ. Statistical non-independence of brain metabolite concentrations whether normalized to creatine or water. Cereb Blood Flow Metab 2025; 45:196-198. doi: 10.1177/0271678X241290018.