Markov Chain Monte Carlo Methods for Exact Tests in Contingency Tables

Abstract

This thesis is mainly concerned with conditional inference for contingency tables, where the MCMC method is used to take a sample of the conditional distribution. One of the most common models to be investigated in contingency tables is the independence model. Classic test statistics for testing the independence hypothesis, Pearson and likelihood chi-square statistics rely on large sample distributions. The large sample distribution does not provide a good approximation when the sample size is small. The Fisher exact test is an alternative method which enables us to compute the exact p-value for testing the independence hypothesis. For contingency tables of large dimension, the Fisher exact test is not practical as it requires counting all tables in the sample space. We will review some enumeration methods which do not require us to count all tables in the sample space. However, these methods would also fail to compute the exact p-value for contingency tables of large dimensions. introduced a method based on the Grobner basis. It is quite complicated to compute the Grobner basis for contingency tables as it is different for each individual table, not only for different sizes of table. We also review the method introduced by using the minimal Markov basis for some particular tables. provided an algorithm using the most fundamental move to make the irreducible Markov chain over the sample space, defining an extra space. The algorithm is only introduced for tables using the Rasch model. We introduce direct proof for irreducibility of the Markov chain achieved by the Bunea and Besag algorithm. This is then used to prove that approach can be applied for some tables of higher dimensions, such as and . The efficiency of the Bunea and Besag approach is extensively investigated for many different settings such as for tables of low/moderate/large dimensions, tables with special zero pattern, etc. The efficiency of algorithms is measured based on the effective sample size of the MCMC sample. We use two different metrics to penalise the effective sample size: running time of the algorithm and total number of bits used. These measures are also used to compute the efficiency of an adjustment of the Bunea and Besag algorithm which show that it outperforms the the original algorithm for some settings.