Comparison of Weighted Total Least Squares and Partial Weighted Total Least Squares Methods: Linear Regression
Comparison of Weighted Total Least Squares and Partial Weighted Total Least Squares Methods: Linear Regression
Shahram Jazaeri1
1) Assistant Professor, Department of Civil Engineering and the Environment, Faculty of Engineering, Shiraz University
IOI Code
:
XKHB-AZADAZ
Publication :
Annual Conference on Hekmat-Based Architecture, Urban Development, Art, Industrial Design, Construction and Technology(icahu.com)
Abstract :
Weighted Total Least Squares (WTLS) is used when the design matrix, in addition to the observation vector, is also contaminated with random errors. As a result, Ordinary Least Squares (OLS) does not provide the desired accuracy for estimating unknown parameters. In WTLS, the design matrix includes both fixed and random variables, making the variance-covariance matrix of the design matrix non-invertible. In Partial Weighted Total Least Squares (PWTLS), the independent random variables are separated in the design matrix, allowing for the invertibility of their variance-covariance matrix. This paper compares various methods for WTLS and PWTLS in linear regression problems with different data sets. The numerical results show that it is possible to estimate the unknown parameters by using WTLS despite the singularity of the variance-covariance matrix of the design matrix and without separating the independent random variables in the design matrix. The weighted total least squares algorithm presented by (Jazaeri et al. 2012, 2014) obtains the unknown parameters with fewer iterations and less time than the partial-weighted total least squares algorithm. Therefore, based on the numerical results, there is no need to use PWTLS considering the changes in the design matrix structure and the complexity of the variance-covariance matrix. It is possible to estimate the unknown parameters for any type of problem, including linear regression, by preserving the design matrix structure and using WTLS without any changes or complexity in the design matrix.
Keywords : least squares weighted total least squares partial weighted total least squares linear regression