Abstract
<title>Abstract</title> <p> This study evaluates how response variable distribution families and smoothing techniques influence Generalized Additive Model (GAM) performance, using juvenile <italic>Chelonia mydas</italic> abundance data from southern Brazil. We systematically compared Normal, Log-normal, Poisson, and Negative Binomial distributions combined with Thin Plate, Cubic, and Penalized Regression Splines. The Negative Binomial distribution outperformed other frameworks across all information criteria and residual diagnostics, effectively mitigating overdispersion. Among smoothers, Thin Plate Regression Splines demonstrated superior efficiency in capturing complex spatial and seasonal patterns, yielding the lowest Generalized Cross Validation (GCV) and Root Mean Square Error (RMSE) scores. Ultimately, the correct specification of the response distribution family proves just as critical as smoother selection for ensuring robust ecological inferences and accurately representing species abundance and site fidelity patterns. </p>