Systematic Evaluation of Bias Associated with a Multiple Imputation Approach for Estimating Missing Exposure Data

摘要

Introduction: Occupational exposure data frequently include missing or incomplete measurements that, if not addressed, can result in an increased potential for exposure misclassification. Standard approaches for handling missing data, including complete-case analysis and mean substitution, have known limitations. This presents an opportunity to evaluate the use of more advanced statistical approaches, including multiple imputation (Ml). Methods: We systematically evaluated a Ml approach for addressing missing exposure data using a uniquely large and complete dataset of >1 million radiation measurements collected on workers in the skilled trades at three shipyards from 1975-2005. The original dataset contained no missing exposure information and thus represented the true exposure levels in the population. We performed a series of simulations in which 10-99% of the radiation measurements were randomly replaced with missing values and then imputed. To evaluate the performance of Ml, estimates of the mean, median, 25th and 75th percentiles, and variance were calculated for each imputed dataset and compared to the true values of each metric to obtain estimates of the raw and relative bias. Results: For the simulations in which 10-95% of the measurements imputed, the raw bias of the mean ranged from -3.0 to -0.3 (relative bias: 2-15%), the raw bias of the median ranged from -3.0 to 0.0 (relative bias could not be calculated), and the raw bias of the variance ranged from -7.1 to 6.6 (relative bias: 0.1-7.4%). For the simulations in which >95% of the measurements imputed, the magnitude of the biases varied widely for each metric and were not informative. Conclusion: Ml was shown to perform well in characterizing the true distribution of exposures, even when large percentages of radiation measurements were imputed. Our results, combined with the statistical advantages of model-based approaches, support the use of Ml for addressing missing occupational exposure data.