Assessing Robustness of Imputation Models Based on Data from Different Jurisdictions

Abstract

The literature indicates that many highway and transportation agencies in North America and Europe estimate missing values in their collected traffic data records. Estimating missing values is known as data imputation. Such a convention can be traced back to early traffic monitoring systems in the 1930s; however, no studies have been found to assess the accuracy of imputations carried out by transportation practitioners. The imputation methods used by highway agencies are varied and intuitive in nature. Some of them could result in large imputation errors in certain circumstances. Those errors can lead to significant deviations in the resulting operation plans and designed structures. Therefore, it is necessary to evaluate the accuracy of various imputation methods that highway agencies use. This study identifies and tests typical imputation methods on automatic traffic recorder data from Alberta and Saskatchewan in Canada. With assessment of imputation methods based on the data from different highway agencies, it is possible to evaluate their robustness and suitability for use across jurisdictions. The accuracy of individual imputation models was statistically analyzed, and comparisons and recommendations were made. Study results clearly indicate that models using additional observations as input and more sophisticated prediction techniques consistently produce better imputations. It is believed that this study would be helpful for traffic engineers in reviewing their imputation practices and, hence, in improving their data quality.