Skip to Main Content

Mediation analysis

In many public health studies, it is of interest to understand the mechanisms for how the intervention affects the outcome of interest. In September 2017, Drs. Daniel Nevo, Xiaomei Liao and I, published the paper "Estimation and inference for the mediation proportion" in the International Journal of Biostatistics. The paper describes novel methodology to estimate the mediation proportion, a quantity of prime interest for researchers carrying out mediation analysis. Rigorous methodology for estimation and statistical inference for this quantity when using the popular ``difference method'' has not previously been available. We formulated the problem using a generalized estimation equations approach, and utilize a data duplication algorithm for estimating the mediation proportion and its variance. We also studied the assumption that the same link function holds for the marginal and conditional models, a property which we term ``g-linkability''. We have shown that our approach is valid whenever g-linkability holds, exactly or approximately, and presented results from an extensive simulation study to explore finite sample properties.

Dr. Daniel Nevo and I worked with Epidemiology doctoral student Louisa Smith in the summer of 2017 comparing the asymptotic to bootstrap variance estimates for the multiplication method for calculating the mediation proportion under a variety of settings. We conducted extensive simulation studies to compare between asymptotic and bootstrap variance estimation methods and evaluate finite sample bias under the different scales.

In another mediation analysis project, fI am working with Rodrigo Zepeda Tello, Dalia Camacho-Garcia-Forment, and Dr. Tonatiuh Barrientos-Gutierrez of the Instituto Nacional de Salud Publica in Mexico City, Mexico. We are collaborating on a manuscript that demonstrates bias with common methods for calculating externally generalizable pPARs and provides several solutions to this problem. The methods are applied to an analysis of the pPAR of sugar-sweetened beverages in relation to diabetes risk. The manuscript has been drafted and is in the process of revision by myself and other co-authors.