Data Accuracy Matters: Improving Highway-Rail Grade Crossings Crash Predictions through Inventory Verification

Abstract

Highway-rail grade crossing (HRGC) crash prediction models’ effectiveness hinges on the input data accuracy and precision. This paper investigates the impact of inaccurate HRGC inventory data on the modeling of HRGC crashes. Specifically, the research explores data gaps by obtaining samples of Federal Railroad Administration rail crossing inventory data. These inventory data were checked for accuracy by visiting the rail crossings and comparing the inventory elements to their field conditions. Any inaccurate records were corrected; the process created an accurate inventory of the rail crossings under consideration. The corrected inventory data was subsequently used for crash predictions using the U.S. Department of Transportation accident prediction formula (U.S. DOT APF), released in 2020. To fit for the U.S. DOT APF, the corrected inventory data from Nebraska was used for the case 1 study, which applied a multiple imputation algorithm to augment the empirical data to verify improvements in the model’s goodness of fit. The results showed that the adjusted Akaike information criterion (AIC) improved from 1,074 to 1,068 when only 7% of the total inventory dataset was corrected, and to 813 assuming all verified corrected data obtained through data imputation. In case 2, the filtered inventory data from four Midwest states (i.e., Kansas, Iowa, Missouri, and Nebraska) were utilized to address data stratification issues in the U.S. DOT APF. Results showed that the adjusted AIC improved from 1,442 to 1,431 when the latest annual average daily traffic data and properly stratified variables (i.e., road surface, traffic control) were included in the U.S. DOT APF. The findings emphasize the need for regular HRGC inventory data verification and improved data-updating processes for more accurate HRGC crash predictions.