?
Index Construction Methodology Using Training Sample Based on Pairwise Comparisons
Most of index construction techniques combine measured features presenting different components of the index. To obtain the correct weights of the features is a matter of great importance. Widely used principal component analysis allows us to do without training sample. It settles the weight of the feature according to its variability. But this method works only with correlated features. Indicators with more or less independent components need either expert-defined weights or a training sample. We propose construction of such a sample on the base of the direct index estimation by experts. To prevent considerable bias it is reasonable to use an expert panel and convert processed information of all the experts into the training sample. Though the sense of the index is considered to be clear for the experts and the objects in the sample are also well known for them, the experts can give information about the index value for each object in qualitative rather than quantitative form even if they are asked to present it in numerical scale. That is why we propose using pairwise comparisons instead, and the number of objects involved of about a dozen or some more looks reasonable. While elements of the pairwise comparisons matrix are assigned in expert-friendly Likert scale, the eigenvector of this matrix as well as the results of its future processing with corresponding estimations of other experts is a source of data in quasi-numerical scale. We present the results of application our methodology for alternative assessing of some widely used indices.
In book
