All of my independent variables are categorical (binary, ordinal, and nominal). If I select a model using the training subsample one of the variables, which was significant in the full sample drops out. Odds ratios vary between 1 and 80% for models applied to training and test subsample. The...