A Robust Shrinkage Estimator for Over dispersed Poisson Regression Using Penalized Likelihood Approaches.
Saved in:
| Title: | A Robust Shrinkage Estimator for Over dispersed Poisson Regression Using Penalized Likelihood Approaches. |
|---|---|
| Authors: | habash, Hussein kareem hswnalkrym47@gmail.com |
| Source: | Journal of Computational Analysis & Applications. 2025, Vol. 34 Issue 7, p93-107. 15p. |
| Subject Terms: | *POISSON regression, *ROBUST statistics, *MONTE Carlo method, *STATISTICAL models, *FEATURE selection |
| Abstract: | Classical Poisson regression collapses when real-world counts are simultaneously over-dispersed, collinear, and contaminated by aberrant observations. Ridge, lasso, and elastic-net penalties each cure part of the problem--variance inflation or over-parameterization--while robust Mestimators address outliers, yet no single method succeeds on all three fronts. We introduce a convex Huber--elastic-net (HEN) estimator that unifies bounded-influence scoring with mixed ℓ1ℓ2shrinkage and joint dispersion estimation. A hybrid IRLS/coordinate-descent algorithm converges rapidly, and a trimmed cross-validation scheme selects the penalty (λ), mixing (α), and robustness (δ) parameters with minimal extra computation. Monte-Carlo experiments spanning two sample sizes, two dispersion levels, and 15 % response contamination show that HEN lowers coefficient mean-squared error by ≈ 25 % and preserves variable-selection F1 scores above 0.85 compared with the best non-robust alternatives, while incurring < 5 % efficiency loss on pristine data. HEN thus delivers a practical, one-stop solution for messy count outcomes. [ABSTRACT FROM AUTHOR] |
| Database: | Academic Search Index |
Nájsť tento článok vo Web of Science | Abstract: | Classical Poisson regression collapses when real-world counts are simultaneously over-dispersed, collinear, and contaminated by aberrant observations. Ridge, lasso, and elastic-net penalties each cure part of the problem--variance inflation or over-parameterization--while robust Mestimators address outliers, yet no single method succeeds on all three fronts. We introduce a convex Huber--elastic-net (HEN) estimator that unifies bounded-influence scoring with mixed ℓ1ℓ2shrinkage and joint dispersion estimation. A hybrid IRLS/coordinate-descent algorithm converges rapidly, and a trimmed cross-validation scheme selects the penalty (λ), mixing (α), and robustness (δ) parameters with minimal extra computation. Monte-Carlo experiments spanning two sample sizes, two dispersion levels, and 15 % response contamination show that HEN lowers coefficient mean-squared error by ≈ 25 % and preserves variable-selection F1 scores above 0.85 compared with the best non-robust alternatives, while incurring < 5 % efficiency loss on pristine data. HEN thus delivers a practical, one-stop solution for messy count outcomes. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 15211398 |
| DOI: | 10.48047/jocaaa.2025.34.07.8 |