Measuring inequality in the adoption of ESG scores by small and medium enterprises

This paper proposes a novel methodology to assess the adoption of Environmental, Social, and Governance (ESG) criteria by Small and Medium-Sized enterprises (SMEs), measuring the inequality in ESG scores as a proxy for their uneven diffusion across firms and sectors. Rather than focusing solely on average ESG performance, we argue that inequality captures critical disparities that reveal structural barriers to adoption. To quantify these disparities, we employ the Gini index, and to explain them, we compare alternative machine learning models and simple Linear Regressions, evaluating their performance through residual analysis, including the calculation of the residual Gini indices. We apply this framework to a sample of over 1,000 Italian SMEs across various sectors. The findings reveal significant heterogeneity in ESG practices, with marked sectoral differences. This approach helps identifying where targeted policy interventions are most needed to promote more inclusive and balanced sustainable development among SMEs. • The paper shows how to measure inequality in the adoption of ESG scores by enterprises, and how to explain such inequality by means of explainable machine learning models both linear and non-linear. • The main results can be summarised as follows. • For Environmental predictions, simpler models (Linear Regression) may suffice, while Random Forest provides better explanatory power. • For Social and Governance predictions, Gradient Boosting and Random Forest should be prioritized depending on whether predictive accuracy or variance explanation is more critical. • The Linear Regression minimizes inequality in residuals (as shown by the Gini rankings) but may not always achieve the highest accuracy or explanatory power. • Gradient Boosting excels at minimizing errors (MSE and MAE), but Random Forest consistently explains more varianc.. • The findings highlights Random Forest’s utility as a more balanced model, while Gradient Boosting might be preferred when error minimization is a priority. On the other hand, Linear Models are generally more explainable.