Online Program

Return to main conference page

All Times EDT

Friday, September 25
Fri, Sep 25, 11:45 AM - 12:45 PM
Virtual
Poster Session

PS25-Estimating Causal Mediated Effects with Mediator Values Below a Limit of Detection (301130)

View Presentation

Ronald Bosch, Center for Biostatistics in AIDS Research, Harvard School of Public Health 
*Ariel Chernofsky, Boston University School of Public Health 
Judith Lok, Boston University Department of Mathematics and Statistics 

Keywords: HIV/AIDS, mediation, organic direct and indirect effects, assay limit of detection, Monte Carlo EM algorithm

The effect of a new HIV curative treatment on time to viral rebound may be mediated through the viral reservoir. Organic direct and indirect effects provide an interpretable method for quantifying the causal mediated effects. However, for many subjects' viral reservoir, the mediator, is below the assay limit detection. Since the mediator is an effect of the treatment and a putative cause of the outcome, assay lower limit presents a compounded problem in mediation analysis. We propose two general approaches to address informative missingness in mediator values. The first approach directly maximizes the observed data likelihood through a numerical optimization procedure. The second approach iteratively maximizes a challenging conditional expectation of the complete data log likelihood through implementations of the EM algorithm. The first EM implementation estimates the parameters of the mediator distribution through a simple application of the EM algorithm and extrapolates the outcome model to mediator values below the assay lower limit. The second EM implementation replaces the E step for the mediator-outcome joint likelihood with a monte carlo estimate, an adaptation known as Monte Carlo EM. A simulation study compares both approaches to ad hoc solutions, such as imputing with half limit of detection. The first approach of numerical optimization of the observed data likelihood performed best with respect to both bias and variance. We illustrate our methods with HIV interruption study data from AIDS Clinical Trials Group described in [Li et al. 2016, AIDS] to estimate the indirect effect of a curative treatment on viral rebound by week 4 after ART interruption through the viral reservoir.