Abstract

This paper addresses the persistent problem of boundary bias in kernel-assisted finite population estimation under simple random sampling without replacement. We propose an Adaptive Boosting (AdaBoost) enhanced Nadaraya-Watson estimator that iteratively reweights observations to reduce boundary bias. The method focuses learning on poorly estimated regions near domain boundaries while maintaining the design-based properties required for finite population inference. Comprehensive simulation studies and real life application across five superpopulation models (Linear, Quadratic, Exponential, Jump, and Sine) demonstrate that the boosted estimator achieves substantial improvements over standard approaches. For moderately varying functions, we observe 11-14% reductions in both bias and root mean squared error (RMSE), with confidence intervals narrowing by 31% on average. The second boosting iteration exhibits superlinear convergence ( α= -1:41), representing a 14.6% acceleration over baseline methods. While e ective for linear and moderately nonlinear functions, the method shows limited utility for highly skewed exponential patterns. These findings offer practitioners a robust tool for boundary bias correction in survey estimation with clear guidelines for application.

This work is licensed under a Creative Commons Attribution 4.0 License.