A novel hybrid structural equation modeling–machine learning framework for identifying key predictors of pro-environmental behavior: A study of urban cycling

Nasiri, A.R., Mahjouri, N., & Yazdanpanah, M. (2026). A novel hybrid structural equation modeling–machine learning framework for identifying key predictors of pro-environmental behavior: A study of urban cycling. Results in Engineering 32 e111929. 10.1016/j.rineng.2026.111929.

Preview

Text 1-s2.0-S2590123026029518-main.pdf - Published Version Available under License Creative Commons Attribution. Download (7MB) | Preview

Abstract

This study proposes a hybrid analytical framework integrating structural equation modeling (SEM) and machine learning (ML) to investigate the socio-psychological predictors influencing citizens’ cycling behavior within a metropolitan context. The framework is grounded in the extended theory of planned behavior (ETPB), which incorporates moral norms to address limitations of the classical TPB model in explaining pro-environmental behavior. Although SEM is widely used to validate causal relationships among theoretical constructs, it has limitations in capturing non-linear associations and latent interaction effects among variables. To overcome these limitations, this study integrates ML with SEM. In addition, Shapley additive explanations (SHAP) are applied to improve the interpretability of ML outcomes by revealing the relative importance and functional impact of each predictor. Data were collected through 307 validated survey responses and subjected to systematic preprocessing and validation. The analysis employed a combined SEM and random forest (RF) approach, complemented by SHAP analysis to enhance model interpretability. SEM results indicate that behavioral intention and perceived behavioral control (PBC) are the main predictors of cycling behavior, explaining 79% of its variance (R² = 0.79). Attitude, moral norms, and PBC were also identified as key predictors of behavioral intention (R² = 0.67), demonstrating improved explanatory power compared with the classical TPB model. The RF models demonstrated robust performance, achieving test R2 values of 0.618 (CV R2=0.577±0.088) for intention and 0.761 (CV R2=0.722±0.115) for behavior. SHAP analysis revealed that core psychological constructs strongly dominate the predictions, while age and topographical slope showed unexpectedly weak predictive importance. These findings indicate that in low-cycling communities, profound psychological barriers act as the primary bottleneck, effectively overshadowing physical constraints like land slope across all age groups. Consequently, urban mobility policies must prioritize building cultural readiness alongside infrastructural improvements.

Actions (login required)

View Item