ESG and firm value persistence: a survival-based machine learning approach

Purpose This study extends prior survival analysis evidence on ESG and firm value persistence by examining whether ESG-related inference is sensitive to linear and nonlinear model specifications. Design/methodology/approach This study uses a sample of Taiwan listed firms from 2016 to 2024, measures firm value using Tobin's Q, and operationalizes firm value persistence by treating the first occurrence of value deterioration as a discrete-time event. Using discrete-time survival analysis (DTSA), the hazard is estimated with machine learning (ML) models under linear and nonlinear specifications, and SHapley Additive exPlanations (SHAP) is used to interpret ESG-related contributions to predicted deterioration risk and implied firm value persistence. Findings The association between ESG indicators and predicted firm value persistence is sensitive to model specification. Under linear specifications, ESG-related directional patterns are more heterogeneous across sector groups, particularly for environmental and governance indicators. By contrast, under nonlinear specifications, the directional patterns are more consistent. Higher environmental, social, and governance indicator scores are generally associated with longer predicted firm value persistence, with governance transparency as the main exception. Originality/value This study offers a methodological and interpretive contribution by integrating DTSA, ML, and SHAP to examine firm value persistence. It shows that model choice affects the interpretation of the association between ESG indicators and firm value persistence. This provides one possible methodological explanation for why prior evidence on ESG and firm value persistence has been mixed.