Methods for Mediation and Interaction: Application to Genetic Variants on 15q25.1, Smoking and Lung Cancer


Methods developed for causal mediation analysis with a dichotomous outcome, applicable to case-control studies via prevalence weighting will be presented. These methods generalize traditional approaches to mediation in the social sciences by allowing for interactions and non-linear models. Methods for sensitivity analysis for unmeasured confounding and measurement error in the context of mediated effects will be described. The methodology is used to resolve a question concerning direct and indirect effects in genetic epidemiology. Genome-wide association studies have identified variants on chromosome 15q25.1 that increase the risk of both lung cancer as well as nicotine dependence and associated smoking behavior. There remains debate as to whether the effect on lung is direct or operates through pathways related to smoking behavior. For two SNPs, rs8034191 and rs1051730, on 15q25.1, we estimated from 1836 cases and 1452 control the indirect effect mediated by smoking (cigarettes per day), the direct effect through other pathways and the overall proportion mediated. Analyses allowed for the possibility that the effect of smoking varied by groups defined by the genetic variant. There was substantial evidence for gene by smoking interaction on additive and multiplicative scales. Mediation analyses suggested that the effect of the variants on lung cancer mediated through smoking are very small in magnitude compared to the independent effect through other pathways. Sensitivity analysis correcting for measurement error and potential unmeasured confounding did not substantially change these conclusions. The results were replicated using data from three other lung cancer case-control studies.

Date & Time(s)


Memorial Sloan Kettering Cancer Center
307 East 63rd Street
3rd Floor Conference Room
New York, NY 10065


Memorial Sloan Kettering Cancer Center
Survivorship, Outcomes and Risk Program
Seminar Series


Tyler J. VanderWeele, PhD
Department of Epidemiology and Biostatistics
Harvard School of Public Health