On the empirical construction of implications between bi-valued test items