There are 4 reasons why not do this:

• 1. Splitting the table hurts the accuracy of the estimates for parameters because we are essentially making two less-accurate instead of pooling all the evidence into one estimate.
• 2. To acquire probability statements about the variable used to split the data they need to be included in a model. Otherwise, we will have a weak statistical argument.
• 3. We may want to use a different model to compare data like information criteria.
• 4. Multilevel models allow for borrowing of information which improves estimates in all categories.