EXAMINATION OF EXPERIMENTAL DATA SAMPLES FOR THE PRESENCE OF OUTLIERS: COMPARISON OF METHODS

The task of detecting outliers (misses, abnormalous values, results that stand out sharply, results that have come off) is one of the most relevant, complex and ambiguous in the experimental materialprocessing. Such values are the experiment results, which are abnormally far from other points from a series of parallel observations. The source of emissions is often measurement errors. Among these are incorrect recording of the experiment results, possible incorrect coding of data, incorrect conduct of the experiment, etc. Gross errors occur in the event of a sudden change in the conditions of conducting the research, malfunctions in the operation of the equipment, etc. At the same time, outliers may indicate an unexpected, extraordinary behavior of the measured value – a yet-to-be-explained property process manifestation. And that's why an analysis using reliable mathematical tools is needed. The methods of detecting emissions are diverse and numerous. Parametric tests are more sensitive to the sample size and to the population values probability distribution. Non-parametric tests are more flexible and can be applied if the non-normal distributon of the sample or the sample size is small; such criteria give a better result in asymmetric distributions, because they use the median instead of the mean; they can be applied to ordinal or nominal data, as well as in the situation of an aberrant outlier value. Interval analysis methods, in particular interval statistics, are an alternative flexible toolkit for obtaining a more accurate and complete analysis of experimental data in the incomplete information, noise presence, measurement outliers, and the presence of abnormalous and aberrant points. A comparison of the results of the application of parametric criteria (-criterion, -criterion, Lvovskyi) and non-parametric criteria (the box-and-whiskers-plot) for detecting emissions, as well as calculation using interval statistics methods, was carried out. One of the outliers was determined by the non-parametric criterion, the -criterion and the procedure for detecting a single outlier using interval methods. Two values are suspicious outliers using the box-whisker rule and the interval statistics recognition algorithm. The methods of detecting outliers using interval analysis methods are no less effective than the use of non-parametric tests.