search

open-connectome-classes / StatConn-Spring-2015-Info

introductory material
18 stars 4 forks source link

MR Connectome Weights? #6

Closed wrgr closed 9 years ago

wrgr commented 9 years ago

To measure MR graph connectivity, we have the option of counting and then binarizing, or using some sort of continuous measure (e.g., log of counts). Do we have any evidence that one is better than the other or that any are "right(er)?" I think no, and that we have to measure against repeatability or classification or I suppose an anatomical prior... But choosing a value seems like a very squishy process...

jovo commented 9 years ago

there is a bias/variance trade-off...what are the bias/variance trade-offs for the 2 different approaches?

On Thu Jan 29 2015 at 12:00:30 AM William Gray notifications@github.com wrote:

To measure MR graph connectivity, we have the option of counting and then binarizing, or using some sort of continuous measure (e.g., log of counts). Do we have any evidence that one is better than the other or that any are "right(er)?" I think no, and that we have to measure against repeatability or classification or I suppose an anatomical prior... But choosing a value seems like a very squishy process...

— Reply to this email directly or view it on GitHub https://github.com/Statistical-Connectomics-Sp15/intro/issues/6.

wrgr commented 9 years ago

Continuous data is susceptible to variance because small fluctuations in the measurements will affect results.

Binarized data is more susceptible to bias, because we could be disregarding an important aspect of the signal.

Log scale might be a somewhat balanced approach as sort of a "low pass filter."

jovo commented 9 years ago

+1