A1514
Title: Comparing regularization strategies for $CO_2$ determinants: Classical, Bayesian and maximum entropy approaches
Authors: Maria Costa - University of Aveiro (Portugal) [presenting]
Ana Helena Tavares - University of Aveiro (Portugal)
Jorge Cabral - University of Aveiro (Portugal)
Abstract: Mozambique provides an empirical setting for comparing information-theoretic and Bayesian approaches in a small-sample regression problem. An annual macroeconomic dataset with a carbon-emissions response and three economic covariates is used to examine 18 static specifications, namely OLS, ridge, lasso, four Bayesian models with weakly informative, ridge, lasso and horseshoe priors, and eleven GME or GCE variants with alternative support widths, updating schemes and prior mass structures. Unit root tests indicate strong persistence in most series, and bounds tests find no robust evidence of cointegration, shifting the analysis from long-run interpretation to a short-run specification in first differences. The central question is not which static estimator recovers a long-run relation, but which family remains closer to the short-run effects identified by the dynamic model when temporal dependence is ignored. The results suggest both a stable component, with consistent sign patterns across specifications, and a method-sensitive component, with substantial variation in magnitude and relative proximity to the dynamic short-run effects. Classical and Bayesian specifications tend to remain closer to the dynamic short-run structure, whereas many entropy-based variants produce more compressed coefficient patterns. The comparison highlights regularization as an inferential choice rather than a purely technical adjustment.
