I have been suggested to exclude item whose point-biserial correlation coefficient (indicates that this item fails to discriminate student competence, based on CTT) is negative before we perform item analysis (estimate item and person parameters) in IRT. However, I have found that I have no literature that support the suggestion.