Detecting Hidden Structure in Traffic Data by Using the New Methodology of Hybrid Structural Equation Modeling-Artificial Neural Network (SEM-ANN)

Abstract

Background and Objectives: Since that the Structural Equation Modeling (SEM) model includes both structural and measurement models, but it does not include non-linear relations, in this case the ANN model is used.So combining the two approaches gives optimal results. In PLS-SEM model, the measurement error and hidden variables can be considered even in small sample sizes and includes both formative and reflective measurement models. The objective of the present study is to investigate the predictors of injuries caused the hospitalization of motorcyclists in a case-control study, using a SEM, ANN, hybrid structural equation modeling-artificial neural network (SEM-ANN), PLS-SEM model. ANN compared to SEM and PLS with SEM, ANN, SEM-ANN. Methods: In this case-control study, 300 cases and 156 controls were selected using a cluster random sampling in Tabriz, Iran. Using of motorcycle-riding behavior questionnaire (MRBQ), Attention-deficit/hyperactivity disorder (ADHD), along with the subscales, and the motorcycle related variables (Such as having a helmet, etc.) were measured. Also, MRBQ were considered as a mediator variable for the ADHD, Demographic characteristics and MRBQ. The SEM model was used to examine the linear relationships of variables in the conceptual model, ANN model was used to examine the nonlinear relationships, interaction of variables and model, SEM-ANN model was used to examine the nonlinear relationships and interaction effect of variables and PLS-SEM model, SEM and PLS models were compared with Chi2/df and SRMR. PLS, ANN and SEM-ANN were compared with sensitivity, accuracy and specificity. Results: The results indicated significant linear and direct relationships between odds of injury and cell phone answering, hyper active child, dark hour riding and MRBQ, while significant inverse relation was observed between injury and being married,and academic education. Also, results of the ANN model showed that MRBQ, ADHD, and its subscales, with injury nonlinear relationships. Also, The results of the hybrid model with consideration of the second and third order of MRBQ, the model's optimality and the meaning of the intermediate variable with the response variable were confirmed. Furthermore, ANN showed higher sensitivity, accuracy and specificity almost was identical to SEM, As a result, the prediction of ANN was better than SEM. The results of PLS-SEM model show that there was a direct and significant relationship between injury outcome with ADHD, and motorcyclists' behavior, however there was not a significant relationship between injury and MRBQ and demographic variables. Also, there was a direct and significant relationship between ADHD and MRBQ and motorcyclists' behavior with MRBQ and ADHD with motorcyclists behavior .The comparison of the two CVSEM and PLS-SEM models in the present study showed the superiority of the CVSEM model by means of Chi2/df and SRMR. Comparison of two models of PLS and ANN is indicators of sensitivity, accuracy and specificity confirmed the superiority of the ANN model. Finally, the comparison of the two models of PLS and SEM-ANN with sensitivity, accuracy, and specificity indicators, confirmed the superiority of the SEM-ANN model. Conclusions: Based on the obtained results, since the SEM does not consider the nonlinear relations and interactions between variables, is better to use the ANN model, which has a higher predictive power compare to SEM, and considering the inclusion of an assemblage measurement model in the study, the PLS-SEM model, which includes both formative and reflective measurements, is used. Due to the significance of the variables, having intervention programs to reduce damage on those have the hyper active child, and those who answer their cell phones while driving, and dark hour riding is highly recommended.