Evaluation of alternative data imputation strategies: A case study of motor carrier safety

Abstract

This paper examines the effectiveness of different methods in imputing mileage incurred by commercial motor carriers (used as exposure measures in deriving safety indices of carriers), by using an administrative dataset on motor carriers. The objective of the paper is to assess the behavior of different imputation methods, using simulation. A series of odds-ratio tests confirm that the mileage data on motor carriers are not in fact ideally suited for data imputation approaches. However, the literature indicates that data imputation methods may be quite robust with respect to departures from the imputation model assumptions. In order to evaluate the level of robustness, samples of motor carriers are constructed by simulation to mimic different total "non-response" conditions and "item nonresponse" (with respect to mileage values of carriers). Four data imputation methods are then evaluated on these simulated samples including: Unconditional Mean Imputation (where mean values are substituted for missing values), Conditional Mean Imputation (regression-based imputation), Expectation Maximization algorithm with simulation (EMis) (an EM-based algorithm that includes a fast simulator for randomly imputing missing values based on estimated EM parameters) and a Data Augmentation based Markov Chain Monte Carlo (MCMC) method]. Conclusions are drawn regarding the use of imputation in safety research.

Keywords:

• motor carrier safety
• missing data imputation
• Markov Chain Monte Carlo
• Expectation-Maximization
• Data Augmentation
• odds ratio