Outlier Handling in Applied Regression: Performance Comparison Between Least Trimmed Squares and Maximum Likelihood-Type Estimators

Cinta Rizki Oktarina; Andini Setyo Anggraeni; Muhammad Arib Alwansyah; Reza Pahlepi

doi:10.21009/JKOMA.082.01

Authors

Cinta Rizki Oktarina Department of Mathematics, Faculty of Information Technology, Batam Institute of Technology, Batam, Indonesia
Andini Setyo Anggraeni Department of Mathematics, Faculty of Information Technology, Batam Institute of Technology, Batam, Indonesia
Muhammad Arib Alwansyah Department of Mathematics Education, Faculty of Mathematics and Natural Sciences, Universitas Negeri Jakarta, Indonesia
Reza Pahlepi Department of Statistics, Faculty of Mathematics and Natural Sciences, University of Bengkulu, Bengkulu, Indonesia

DOI:

https://doi.org/10.21009/JKOMA.082.01

Keywords:

Robust Regression, Outliers, M-Estimations, Least Trimmed Squares, Poverty Modeling

Abstract

Poverty analysis often relies on regression models whose performance can deteriorate in the presence of outliers, leading to biased estimates and unreliable conclusions. This study aims to evaluate the effectiveness of robust regression methods compared with Ordinary Least Squares (OLS) when modeling poverty levels across 154 regions in Sumatra. Four socioeconomic indicators were used as predictors, and outlier detection was conducted using the DFFITS approach. After identifying deviations from normality and the presence of influential observations, two robust estimation techniques M-estimation and Least Trimmed Squares (LTS) were applied to improve model stability. The results show that while all predictors significantly influence poverty, the LTS estimator provides the most accurate and robust performance, yielding the smallest Mean Squared Error (MSE) and an R-squared value of 53.37%. These findings demonstrate that LTS is better suited than OLS and M-estimation for handling data contamination and offers a more reliable approach for modeling poverty determinants