Abstract

Cognitive diagnostic models show great promise in language assessment for providing rich diagnostic information. The lack of a full understanding of second language listening subskills made model selection difficult. In search of optimal CDM that could provide a better understanding of L2 listening subskills and facilitate accurate classification, this study carried a two-layer model selection. At the test level, A-CDM, LLM, and R-RUM had an acceptable and comparable model fit, suggesting mixed inter-attribute relationships of L2 listening subskills. At the item level, Mixed-CDMs were selected and confirmed the existence of mixed relationships. Mixed-CDMs had better model and person fit than G-DNIA. In addition to statistical approaches, the content analysis provided theoretical evidence to confirm and amend the item-level CDMs. It was found that semantic completeness pertaining to the attributes and item features may influence the attribute relationships. Inexplicable attribute conflicts could be a signal of suboptimal model choice. Sample size and the number of multi-attribute items should be taken into account in L2 listening cognitive diagnostic modeling studies. This study provides useful insights into the model selection and the underlying cognitive process for L2 listening tests.