I do share the concern expressed by others about the abstract conclusion, particularly the bit I've highlighted in bold:
The misrepresentation in the abstract is worse than that. It claims that steroid ratios were altered. In fact, the text notes that steroid ratios were not different.Conclusions Despite no significant differences in absolute steroid levels, network analysis revealed profound disruptions in steroid-steroid relationships in ME/CFS compared to controls, suggesting disrupted steroid homeostasis. Collectively the results suggest dysregulation of HPA axis function and progestogen pathways, as demonstrated by altered partial correlations, centrality profiles, and steroid ratios. These findings illustrate the importance of hormone network dynamics in ME/CFS pathophysiology and underscores the need for more research into steroid metabolism.
Not necessarily. We see everywhere in biology that different attributes of the person can affect what the metabolites are, and how quickly they are processed into other things. After all, that is what is being argued here - that the ME/CFS status is altering the relationships. I think then it has to be acknowledged that, in the case of many of these hormones, for example the stage of the menstrual cycle, whether someone is menopausal, whether they have been treated with steroids including those in oral contraceptives, gosh possibly even hormonal replacement therapy which we also don't know about, are of such fundamental importance that they should not be ignored.But even with that effect, you’d still expect (for example) the relationship between progesterone and its downstream metabolites to hold up strongly.
Thanks for explaining that. But, it makes me even more concerned.I am guessing that a lot of the confusion originates from understanding of the method used. A partial spearman correlation is a rank-based (not value-based) correlation which effectively regresses out the effects of other measured variables to try to quantify the direct influence of one variable on another as much as possible.
I think the 52 number was before adjustment for FDR.However, network analysis revealed a marked reduction in direct steroid-steroid relationships in ME/CFS, with controls exhibiting 52 significant partial correlations
I’m sorry @Hutan, that’s nearly the opposite of what I’ve been trying to explain. It must be something in my explanation that is not coming across, maybe some background that I don’t realize needs to be laid out explicitly. But I don’t really know how else to explain without investing time that I don’t have at the moment.If I'm understanding you correctly @jnmaciuch, what they did find is some differences when they applied a method that makes the results for all the comparisons even more vulnerable to some chance differences in some hormones.
I think I understood you correctly, in broad terms But, removing the influence of other factors inevitably changes the factor of interest. So, I think it does make the factor of interest even more vulnerable to random oddities or problems related to the confounders that we can't properly quantify. There are not enough significant data points and too much uncertainty about them to get carried away with clever adjustments and still be certain that there is something true. It's very hard to make a silk purse out of a sow's ear.I’m sorry @Hutan, that’s nearly the opposite of what I’ve been trying to explain. It must be something in my explanation that is not coming across, maybe some background that I don’t realize needs to be laid out explicitly. But I don’t really know how else to explain without investing time that I don’t have at the moment.
Do you mean that it doesn't appear to be statistically valid? Each group on its own might have a correlation that is too small to be statistically significant when you're testing if the correlation is different from zero correlation. But the distance between the correlation in ME/CFS and the correlation in controls can be large enough that it is significant.Also, it's not just the problem with the number of features in table 8, but how that flows through to the suggestion of 57 differences between correlations the ME/CFS and control cohorts. For example, in Table 8 of significant partial correlations for the Controls, there is no steroid pair that includes DOC. And yet in Table 10, my rough count gets to 11 steroid pairs including DOC where they are claiming significant differences between steroid relationships for the Controls and ME/CFS cohort . I can't see how that can be valid.
One test is testing if either group's correlation is different from zero. The other is testing if the groups' correlations are different from each other.I think I mean something along the lines of
Table 8 tells us that, in the controls, levels of DOC are not significantly related to any other hormone
So, given that, then it doesn't seem reasonable to suggest that there is something wrong with the relationship between DOC and 11 other hormones in the ME/CFS group just because the trends of a relationship are different. The relationship between DOC and other hormones are not significant for either group.
And so on for other hormone pairings.
Yes. And I'd say with that example, if there is no statistically significant response in the ME/CFS group, and no statistically significant response in the controls, then I don't think we can say that the ME/CFS response was faulty. All we can say is that either the metabolite isn't related to the stimulation or the sample wasn't big enough to get a sufficiently strong signal. The 1 unit up in the ME/CFS group and the 1 unit down in the controls might just be noise.A similar example: You're testing change in a metabolite after stimulation. In ME/CFS it goes up 1 unit, and this is so small that it is not significant when seeing if it's different from 0 units change. Controls go down 1 unit. Again not significant. But if comparing the change in ME/CFS to the change in controls, the difference between them is 2 units, which is large enough to be significant.
I understand the 'different from zero' versus 'different from each other' tests.One test is testing if either group's correlation is different from zero. The other is testing if the groups' correlations are different from each other.
The first not being significant doesn't mean there's definitely no correlation between these metabolites, just that with the data we have (one group's data compared to zero) we can't be sure there is a non-zero correlation.
The second being significant means that there is enough of a difference between groups to say there's likely a difference between them.
(This implies that at least one of the groups must indeed actually have a non-zero correlation, since they can't both be zero and different from each other. But the nature of the data - small sample, large variance - didn't allow for it to be significant for either group in the first test.)
To explore potential associations between circulat-ing steroid levels and symptom severity, we conducted exploratory spearman correlation analyses with clinical scores from the Mental Fatigue Scale and the FibroFa- tigue Scale. None of the associations remained signifi-cant after correction for multiple comparisons, and these results should be interpreted with considerable caution due to the small sample sizes.
If by noise you mean confounders like people with ME/CFS might be more likely to take contraceptives, that's still totally possible, no disagreement here. That's not the kind of noise this kind of statistical test is testing for though.The 1 unit up in the ME/CFS group and the 1 unit down in the controls might just be noise.
They don't really seem to focus on the differences between group correlations from Table 10 anyway, from what I could see. It just got a couple passing mentions. The bulk of their conclusion seems to be based on Tables 8 and 9.There is a very poor basis for then saying that because the non-significant trends in relationships in the two groups are different, there is a faulty relationship between DOC and another hormone in the ME/CFS group. It's torturing the data.
Yes. But my point is that when you say "correlation", most of the time the steroid pairs aren't actually correlated with each other in a statistically valid way. Not in the ME/CFS group and not in the controls. For most of the steroid pairs within a group, there was no statistically significant correlation. Table 10 is mostly concerned with comparing trends.The p-value is small enough to say that the reason the ME/CFS correlation is far from the control correlation is not just because of random noise. As in factors that have nothing to do with their group status. (e.g. if everyone was tested at a totally random point in their menstrual cycle, that'd be this kind of random noise that a low p-value would rule out as the reason for the difference.)
Despite the authors claiming there were 52 significant partial correlations in the control group, Table 8 shows that there were only 27.52 significant partial correlations were demonstrated within the control group (Table 8)
So are you suggesting they made an error in the statistics? Or that the results aren't strong enough to base conclusions on? For the former, I don't see any reason to think what they got was not possible. For the latter, their conclusions have very little to do with Table 10's findings.Yes. But my point is that when you say "correlation", most of the time the steroid pairs aren't actually correlated with each other in a statistically valid way. Not in the ME/CFS group and not in the controls. For most of the steroid pairs within a group, there was no statistically significant correlation. Table 10 is mostly concerned with comparing trends.
Yeah, I see those same switch ups and mislabels.There are errors in the tables.
There are two abbreviations described as androstenedione - Andro and AED. All the other steroids presumably have unique long names.
In Table 8 Androstenedione (AED): Progesterone (P4) r=0.854 in controls
In Table 10 Androstenedione (Andro): Progesterone (P4) r=0.854 in controls
In Table 10 Androstenedione (AED): Progesterone (P4) r=0.712 in controls
In Table 10, there is nothing called pregnanolone. P5 is pregnenolone, and so is PNL (presumably that one should be pregnanolone). In Table 8 there is also a PLN - also pregnenolone.
In Table 8, the r values for the correlation of 170HP with ADT is 0.707. The correlation of 170HP with AED is 0.783.
In Table 10, the r values for the correlation of 170HP with ADT is 0.783. The correlation with AED is 0.707.
There are some differences in the r values given for the steroid pairs in controls in table 8 and for steroid pairs in controls in table 10. Some of the r values are the same, but some are different.
e.g. ADT DHEA in controls - Table 8 r=0.85, Table 10 r=0.96.
In Table 8 AED DHEA in controls r=0.96, Table 10 r=0.85
And those are just the errors I can spot because there is an overlap between the two tables, so there are values that should match to compare. (hopefully I have these right, probably good to check.)