The ground level ozone concentration over the continental United States is analyzed from the point of view of modern Extreme Value Theory using ozone data from the Clean Air Status and Trends Network (CASTNET) at 25 measurement sites. First, we estimate the changes in ozone means according to the NOx SIP call policy implemented during 2000 in Northeastern U.S. The most significant change time in ozone mean is estimated in both parametric and non-parametric ways. The change in variability is also estimated but the results are not as significant as the change in mean. The results show that the policy is effective in reducing the ozone means within 2-4 years. Thus, to analyze the effects of the policy in the extreme sense, the ozone data is divided into two climate schemes. Then, the Generalized Pareto Distribution is fit to extremes of the ozone concentration by using a combination of maximum likelihood estimates (MLEs) and Hill estimates. The data is transformed prior to extreme value analysis and data in the right tail is separated from that in the middle part of the distribution. This analysis is compared to current approaches by using synthetic data. Under a variety of conditions the procedure using the MLE approach is likely to underestimate the tail of the distribution. The analysis shows that at some CASTNET locations the ozone probability distribution is not exponentially bounded, and thus can be characterized as heavy tailed. The ozone tail distributions become heavier following the NOx SIP call at most of the sites with heavy tails prior to this call. In the final part, we study the extreme dependence between temperature and ozone. We also use simulated data from existing models in the study to verify the model correctness. These models are widely used in climate analysis; however, in terms of extremes, the models usually do not well represent the relationship between temperature and ozone compared to the CASTNET measurements. The models tend to have higher extreme dependence than those from CASTNET data, and hence, we should be careful when using data from these models in extreme studies.