Cluster randomized trials (CRTs) are experiments in which clusters of people, or intact social units, rather than individuals are randomized into either intervention or control groups. The key feature of CRTs is that outcomes on individuals in the same cluster tend to be correlated. Hence standard analytical methods that assume individuals within a cluster are independent will have a tendency to underestimate the standard error of the treatment difference and inflate type I error, which lead to overstate the significance of the results. In addition, for randomization at cluster lever, it is quite common to find a large imbalance on important factors between treatment groups, especially when only a few clusters are randomized. In order to reduce the imbalance risk and keep the study as effective and precise as possible, in this thesis, we employed the stratification design under CRTs, investigated and compared the performance of six methods. These methods include (1) unadjusted Mantel-Haenszel test; (2) adjusted Mantel-Haenszel test; (3) cluster level t test; (4) two-stage covariates adjusted cluster level t test; (5) generalized estimating equations (GEE) and (6) generalized linear mixed model method (GLMM). Empirical Type I errors and powers for the different methods considered are evaluated using Monte Carlo simulation over a wide range of scenarios, including very small numbers of clusters, which are commonly used in practice. Furthermore, the choice of an appropriate number of strata for the study design is also discussed. Based on our simulation results, the two-stage covariates adjusted t test outperformed other methods in terms of empirical significance and power levels when the number of clusters is small. We further illustrated these approaches by applying real life data collected from a community-based cluster randomized trial, in which the outcome measure was the knowledge of HIV acquisition.