Statistical Test Method Of Observation Data Outliers

The existence of outliers reduces the quality of data to a great extend, and makes the corresponding results of data analysis significantly change more, which will lead to people's inaccurate judgments. The purpose of this dissertation was to investigate the statistical detection method for outliers in observational data. It discussed and analyzed outlier detection methods in several important distributions, such as Exponential Distribution, Normal Distribution, Weibull Distribution, Extreme Value Distribution, Log-normal Distribution, each of which plays a significant role in survival analysis and reliability theory. The main focus was establishing a new method of detection for multiple outliers in Two-parameter Log-normal Distribution. Its main results are as follows:Firstly, on the basis of leap degree and order statistic, this paper present a concept called fluctuation rate which can be used as a criterion to determine the set of suspicious abnormal data, and furthermore confirm normal data set. This step made a foundation for it to propose a new method of abnormal data detection. The method was to construct some new statistic or pivot first, and then add an adjacent observation data of order statistic to the left or right of normal data set .It utilized the statistic to make detection test for the recombining data set. Thirdly, two piece of statistic were established, and they are proved to be pivot. Forth, by means of Monte Carlo simulation, a table formed by sample fraction is proposed.Finally, by means of numerical simulation, some comparative studies for some common outlier detection methods were accomplished, too. The results showed that the new approach is feasible, and its efficiency is comparatively higher.