| Popis: |
Many studies in health care deal with zero-inflated data sets characterized by a significant proportion of zero and highly skewed positive values. Although it is a common practice to use the median instead of the mean as the measure of central location in skewed data, many applications require the use of the mean. For instance, the mean can be used to recover the total medical cost which reflects the entire expenditure on health care in a given patient population. For testing the value of a mean, the empirical likelihood method offers the benefit of making no distributional assumptions beyond some mild moment conditions while retaining the same advantages that parametric likelihood based tests enjoyed.In this dissertation, we proposed an empirical likelihood ratio test for the difference between means of two zero-inflated samples. The proposed test was derived by jointly specifying the empirical likelihood for the mean parameter and the probability of taking zero value in the data. There are two unique features in this procedure. One is that the information contained in the zero observations is fully utilized and that the proposed test is insensitive to the skewness of the non-zero observations. We derive an asymptotic distribution that will be used to calibrate the statistic in testing the null hypothesis of no mean difference. We also extend the procedure to testing the mean equality of several independent zero-inflated populations. As a benchmark for comparison against conventional tests, we investigate the empirical type 1 error and power rates in finite sample settings. Both the proposed two sample test for the mean difference and the equality of means between three or more populations exhibits comparable if not superior finite sample performance. Another application of empirical likelihood approach that we consider is on detecting epidemic change point in a sequence of observations. Under some mild conditions, the asymptotic null distribution of the test statistic is showed to be an extreme distribution. Simulations indicate that the proposed test performs at par if not better than other available tests while enjoying less constraint on the data distribution. |
