Re: [Icatb-discuss] Which method is recommended for estimating the number of independent components (ICA) in the SBM

[Vince D C. <vca...@gs...>]

You can try MDL (fwhm) using either the estimated smoothness of the data or the FWHM of the smoothing you applied as an approximation.  If it doesn't work well, you can make an empirical choice based on your question of interest similar to what is done for fMRI, low model orders (e.g., 20-30) tend to capture larger covarying patterns (networks) and higher model orders (e.g., 75-100) tend to capture subnodes of the larger patterns (cf. papers by Na Luo and Juith Segall for examples).

Best,

Vince


> -----Original Message-----
> From: Jian Zhang <zha...@gm...>
> Sent: Wednesday, October 4, 2023 5:39 AM
> To: ica...@li...
> Subject: [Icatb-discuss] Which method is recommended for estimating the number of
> independent components (ICA) in the SBM
>
>   [EXTERNAL]
>
> Dear experts,
>
> I am currently analyzing grey matter images using the SBM. I have two questions
> concerning the estimation of the number of independent components within the SBM:
>
> 1. Four methods are available for estimating the number of independent
> components: MDL (i.i.d sampling), MDL (FWHM), Entropy Rate (Finite Memory), and
> Entropy Rate (AR). Which method is most recommended for structural data?
>
> 2. For all four methods, we are required to input an FWHM value. Is this value specified in
> mm OR in voxel?
>
> Thanks in advance for the help.
>
> Best regards,
>
> Jian
>
> _______________________________________________
> Icatb-discuss mailing list
> Ica...@li...
> https://lists.sourceforge.net/
> %2Flists%2Flistinfo%2Ficatb-
> discuss&data=05%7C01%7Cvcalhoun%40gsu.edu%7C59c04b8ebda748f884cc08dbc4bdf
> 472%7C515ad73d8d5e4169895c9789dc742a70%7C0%7C0%7C638320537875067553%
> 7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6
> Ik1haWwiLCJXVCI6Mn0%3D%7C1000%7C%7C%7C&sdata=PVOHZiUrcw7nIcUcs%2
> FcS7dS%2B77wsRHZGrUVgX3VnD8Q%3D&reserved=0
> CAUTION: This email was sent from someone outside of the university. Do not click links
> or open attachments unless you recognize the sender and know the content is safe.