Hello, trying the power of this wonderful research community here. Maybe somebody is better trained in statistics and can help me decide following issue.
I have posted a while ago a similar question about average response function here, now I am facing another challange.
I have a group of patients (N=36) and a group of controlls (N=93), obviously.
But in the end, I will compare them as one-vs-group, not group-vs-group. This makes some differences when compare to “typical” analysis I am experienced in. I am doing a fixel-based analysis.
My question is, how to calculate popuation template in this case.
calculation using N=20 healthy controls and N=20 patients seems weird to me, since I don’t think that making a template using subjects that won’t be included later in a statistical analysis is a right option. (Again, I will compare one patient with all controls). ?
using all HC and all patients (same issue as 1. but more time-consuming) ?
calculation using only N=20 controls/all controls and no patients, but I am not sure since the only patient that will be in a statistical analysis representing whole one group, won’t be included in a making of template? I went through this discussion with @rsmith.
Technically, one would say that speaking about one-vs-group statistics, it should be treated in the same way as group-vs-group, just with the difference that one group contains only from one subject. That would mean that If I use only HC for creating a template I would skip whole patient group - something that in a stroke case mentioned before, @rsmith did not recommend.
Caltulation of separate template for each statistical-analysis-group (Template 1: patient No.1+controls; Template2: patient No.2+controls… etc) does not seem like something I wanna calculate and store in my PC. For obvious time and space reasons.
My patient group is widely heterogeneous - epileptic patients with nonvisible pathologies and different lobes affected.
I hope I explained it well.
Thank you for your ideas!
Sorry for taking absolutely forever to get back on here (and this extends to many). Things have been pretty rough down here, so I’ve now got a very long and intimidating list of questions ahead of me…
While I’ve thought a fair bit about how to introduce N=1 capabilities, it’s not something that I’ve yet gone to the effort of reading up on (after all, it’s not a challenge that’s unique to FBA). So I can muse on the topic, but I’m not claiming authority.
There’s technically two separate considerations here:
The data that are to be used to generate a template image that will be used as a registration target;
The data that are to be used to generate a null distribution against which individual subjects will be compared.
The prior argument for use of data distributed across groups for formation of the template image is based on point 1 above. So if you do not want results in N=1 tests to be biased by inferior / superior efficacy of registration, you could still employ this strategy, without such necessitating that you do the same for point 2 above. It would however only provide such a bias reduction in the context of your specific cohort, and would limit the applicability of your pipeline to other data.
For point 2, I’m not personally convinced that the hypothesis “is this one subject different from the group (36 patients + 93 controls)” is of tremendous utility; the question “is this subject different to controls” feels far more intuitive.
There is however a further point of clarification here regarding the separation of the two points above. In your description (particularly 4.), I sense the implication is that for every individual subject, you would need to generate a new template image. This I personally don’t believe to be the case. Once you have a satisfactory template image, you can then, for every individual subject, generate the normative data against which the subject is to be compared based on selection of the appropriate fixel data. This should be very fast in comparison to generating a new template image every time. It will however be important to not bias those comparisons. So for instance, in the case of a control subject you may compare that 1 subject to the other 92 controls, whereas for a patient you may want to randomly omit one subject from the control group so that the volume of data in the control group is the same regardless of which subject you are testing.
Don’t know if any of that will still be useful given the time elapsed, but maybe there’s something useful in there.
Thanks for your time and your muse on the topic! Looking at it as at two separate points helped I have actually created a separate thread for a statistical comparison in 1-vs-group cases here (don’t get confused by number of subjects, I am experimenting with different quaility control evaluations - more or less strict) You will probably run into it after you will successfully dig through the piles of questions here . Need to mention that we all are grateful you are digging through them!
At first, yes, by “group” I meant Control Group made of helthy controls. I never wanted to compare one epileptic patient to the group of mixed epi patients and healthy controls. My intention was to compare every single one patient to the group of healthy control subjects in an individual testing. Now I am thinking maybe I will compare one-healthy-control-subject to the-rest-of-the-healthy-controls as well, to see whether and how big effect I can see on healthy subjects. Sorry for confusion
Eventually, I have decided to make a template using 20 healthy controls and 20 patients (excluding ones with huge visible structural abnormalities) for a registration purposes (point 1).
However, I am not sure If I understand this last part correctly. By normative data you mean basically
to have a right set of subjects in a control group that are used to generate a null distribution against which individuals will be statistically compared, right? For patients, using FD, FC and FDC data from 92 out of 93 healthy control subjects (one randomly omited); for healhy subjects, comparing his/hers/its FD, FC and FDC to the ones of remaining 92 (without the one in a first group)?
Yes, though admittedly cryptic. The key point was that you don’t necessarily need to be generating a new set of template fixels for every subject to be analysed. You can use the same set of template fixels. But for each N=1 subject, you can choose which subjects will contribute to the estimation of a “healthy” e.g. FD distribution against which the N=1 subject will be compared.
I have actually created a separate thread for a statistical comparison in 1-vs-group cases 