How to increase the power of the test in sparse contingency tables: a simulation study

When analyzing categorical or ordinal data one often comes across sparse contingency tables. One of the biggest problems with this situation is due to the low power of tests for independencies between variables. In this paper, we propose a procedure, based on context-specific independencies, to increase the power of tests. The idea is to focus on subtables where the number of null cells is relatively low. In addition, including these kinds of independencies in the Union-Intersection procedure can provide comforting results. These results are shown in a simulative study of different scenarios.