Evaluasi Kinerja Regresi Lokal pada Pemodelan Hubungan Nonlinier

LOESS (Locally Estimated Scatterplot Smoothing) is a nonparametric approach for modeling nonlinear relationships without explicitly specifying a functional form. Its performance depends on the span, sample size, and error variability. This study evaluates local regression against classical linear regression and cubic polynomial regression across three types of nonlinear relationships (sinusoidal, local spike, and Runge function) through a Monte Carlo simulation with 300 replications per scenario. The factors varied include sample size (n = 30, 50, 100, 200), sigma (0.15 and 0.40), and LOESS span (0.3, 0.5, and 0.7). Model performance is evaluated using the Mean Squared Error (MSE), along with its decomposition into squared bias and variance. The results show that a narrow span (0.3) consistently yields the lowest MSE for functions with sharp curvature and local spikes, with an average MSE reduction of up to 91.7% compared with linear regression for the Runge function. For relatively smooth functions, cubic polynomial regression slightly outperforms LOESS. Larger sample sizes reduce the MSE of nonparametric methods by decreasing variance, whereas excessively wide spans increase bias due to oversmoothing. These findings highlight the importance of adaptive span selection and demonstrate the advantages of local regression in modeling complex and heterogeneous nonlinear relationships.