The canonical missingness family intentionally keeps a broad structured-missingness view, where the missingness of one target column can depend on the states of any other usable column.
As a sensitivity analysis, we also evaluate a strict pairwise variant that only retains pairs of columns that both carry meaningful native missingness. Differences between the two reveal whether a model preserves general conditional missingness structure more easily than direct co-missing behavior among missing columns themselves.