A recursive partitioning approach for subgroup identification in brain-behaviour correlation analysis

In neural correlates studies, the goal is to understand the brain-behaviour relationship characterized by correlation between brain activation responses and human behaviour measures. Such correlation depends on subject-related covariates such as age and gender, so it is necessary to identify subgroups within the population that have different brain-behaviour correlations. The subgrouping is made by manual specification in current practice, which is inefficient and may ignore potential covariates whose effects are unknown in the literature. This study proposes a recursive partitioning approach, called correlation tree, for automatic subgroup identification in brain-behaviour correlation analysis. In constructing a correlation tree, the split variable at each node is selected through an unbiased variable selection method based on partial correlation test, and then, the optimal cutpoint of the selected split variable is determined through exhaustive search under an objective function. Three types of meaningful objective functions are considered to meet various practical needs. Results of simulation and application to real data from optical brain imaging demonstrate effectiveness of the proposed approach.