We are conducting a series logistic regressions using the TYPE=COMPLEX in which dummy-coded ethnicity (9 schemes with 8 dummy codes each) predicting a dichotomous outcome. In one model when ethnic group A is the reference, results show that ethnic group B (vs. group A) significantly predicts the outcome. However, when the same model is run with ethnic group B as the reference, results show that ethnic group A (vs. group B) does not significantly predict the same outcome. Nothing about the models (excepting the reference group of ethnicity) is different. How can this be happening?