When running hierarchical clustering analysis of a matrix of
individuals x samples (e.g., employee performances across different days), there are several possibilities for normalization. If one is clustering the columns (to see whether on certain days individuals perform similarly), one could
z-score normalize across the rows to make each individual employees mean and standard deviation comparable across days, or
z-score normalize across the columns to make all employees comparable within a day, or
not normalize at all and cluster the raw values
Could someone explain the relative advantages/disadvantages of each approach here? To clarify, I am using correlation distance.
Methods 1 or 2 in practice give different results but it's not clear that for the task of seeing if days cluster together, whether #1 or #2 are more appropriate if one chooses to normalize.