Causal mediation analysis with a partially missing mediator: exploring the effect of portable air purifier use during pregnancy on infant birth weight

Mediation analysis examines the exposure-outcome association that acts through an intermediate variable. However, mediation analysis becomes challenging when data have missing values. Although methods exist to deal with missing data and mediation analysis independently, few studies have examined how to combine the approaches, specifically, how to pool the mediation analysis results across a series of imputed datasets and compute confidence intervals for target parameters. We propose a new technique that combines multiple imputation with maximum likelihood estimation. Using computer simulations, we compare the performance of our proposed approach with a traditional bootstrap approach. Our method performs well and is more computationally efficient than other resampling methods. We apply the new method to randomized trial data on the role of cadmium exposure in mediating the effects of an environmental health intervention on birth weight.