One challenge in brain imaging studies is determining how to analyze the data. There is a large amount of variance between the subjects, and building dynamic networks to model the variance is difficult. Junning Li et al. at the University of British Columbia analyzed the various available methods to do so using a data set from individuals with Parkinson’s.

Although they already knew the “biological truth” of the matter (that L-dopa improves motor control in individuals with Parkinsons, by inhibiting the effects of bradykinesia), they were interested in which statistical method would yield the most appropriate result.

They compared three distinct dynamic Bayeisan network strategies:

1) The “virtual typical subject” approach, which pools together all of the subjects data and “learns” a mixed model for the whole group. This approach is based on the idea that variability between subjects will be low enough to ignore.

2) The “individual subject” approach, which creates a separate Bayesian network for each subject and searches for common elements of the networks. It is based upon an assumption of large variance between subjects.

3) The “common-structure” approach, which is based on the assumption that subjects will have similar *patterns* of brain connections, but that they will only differ in the amount and type of interaction between each region. Mathematically, each subject’s brain model has the same structure, but the connection coefficients for each subject are optimized individually.

For this particular study of Parkinson’s patients, the researchers found that the individual subject approach worked best to support the idea that L-dopa makes the brain connections of Parkinson’s patients more similar to normal patients. Since this was their preconceived “biological truth”, one might say that it was the most effective method.

Nonetheless, the researchers stressed that each of the approaches has limitations, and that researchers should attempt to combine pieces of statistical evidence as well as preconceived biomedical knowledge in making the decisions about which group-analysis technique to use.

**Reference**

L. Junning, Z. J. Wang, S. J. Palmer, M. J. McKeown, Dynamic Bayesian network modeling of fMRI: A comparison of group-analysis methods, *NeuroImage ***41**(2008), pp. 398-407.