PURPOSES : Driving simulations are widely used for safety assessment because they can minimize the time and cost associated with collecting driving behavior data compared to real-world road environments. Simulator-based driving behavior data do not necessarily represent the actual driving behavior data. An evaluation must be performed to determine whether driving simulations accurately reflect road safety conditions. The main objective of this study was to establish a methodology for assessing whether simulation-based driving behavior data represent real-world safety characteristics. METHODS : A 500-m spatial window size and a 100-m moving size were used to aggregate and match the driving behavior indicators and crash data. A correlation analysis was performed to identify statistically significant indicators among the various evaluation metrics correlated with crash frequency on the road. A set of driving behavior evaluation indicators highly correlated with crash frequency was used as inputs for the negative binomial and decision tree models. Negative binomial model results revealed the indicators used to estimate the number of predicted crashes. The decision-tree model results prioritized the driving behavior indicators used to classify high-risk road segments. RESULTS : The indicators derived from the negative binomial model analysis were the standard deviation of the peak-to-peak jerk and the time-varying volatility of the yaw rate. Their importance was ranked first and fifth, respectively, using the proposed decision tree model. Each indicator has a significant importance among all indicators, suggesting that certain indicators can accurately reflect actual road safety. CONCLUSIONS : The proposed indicators are expected to enhance the reliability of driving-simulator-based road safety evaluations.