Statistical and machine learning analysis of Potato damage: modeling count data with GLM and predictive algorithms

Abstract

This study evaluates statistical and machine learning approaches for modeling potato damage count data, comparing Poisson regression, Negative Binomial regression, Random Forest, and XGBoost. Initial analysis revealed significant overdispersion (dispersion parameter = 4.95) in the Poisson model, prompting the use of Negative Binomial regression, which demonstrated superior fit (AIC reduction from 6093 to 4326) despite higher prediction errors (RMSE = 8.28). Machine learning models outperformed traditional approaches, with Random Forest achieving the lowest prediction errors (RMSE = 3.13, MAE = 1.97), suggesting better capture of complex data patterns. Key factors influencing damage included genotype (G3 increasing damage by 30%, G5 reducing it by 27%), energy type (E2 increasing damage by 46%), and rodent exposure (R5-R8 showing protective effects). While Negative Binomial regression provided more reliable statistical inference for count data, machine learning offered superior predictive accuracy. These findings highlight the importance of selecting modeling approaches based on study objectives—traditional methods for understanding factor effects and machine learning for optimal prediction. The results provide actionable insights for potato storage and handling practices while demonstrating the value of multiple analytical approaches in agricultural research.