SYNTHESIS OF LINEAR REGRESSION MODEL AND EIV-MODEL

UDC 519.862.6
doi: 10.26102/2310-6018/2019.24.1.033

M. P. Bazilevskiy


This paper is devoted to a synthesis of pair-wise linear regression model and simplest EiV-model (Errors-In-Variables model), better known as the Deming regression. The EIV model is a regression in which all variables contain random errors. Such models have a number of significant drawbacks, which makes it difficult to work with them. The synthesis proposed in the paper, called the two-factor model of a fully connected linear regression, is not only devoid of these shortcomings, but also has certain advantages. The main stages of the construction and analysis of two-factor models of fully connected linear regression are considered. The proposed fully connected linear regression model has much in common with the classical multiple regression model; however, these two types are based on completely different approaches. If multiple regression is based on the principle “independent variables affect the dependent one”, then the principle of fully connected regression is “all variables influence each other”. It is established that the approximation abilities of fully connected models do not exceed the capabilities of multiple regressions, but the former have a much more diverse interpretation. The developed synthesis can be used in the construction of multiple models as a tool for solving problems of reducing the dimensionality of data, eliminating multicollinearity and selecting informative regressors.

Keywords: :regression model, ordinary least squares, total least squares, Deming regression, EIV-model, fully connected linear regression model.

Full text:
Bazilevskiy_1_19_1.pdf