Causal mediation analysis with failure time outcome and error-prone longitudinal covariate
MetadataShow full item record
Mediation analyses are important for understanding the biological mechanisms whereby a treatment/exposure influences an outcome of interest. For example, one may be interested in whether body fat accumulation mediates an association of certain dietary patterns with cancer risk. Similarly mediation analyses may aim to achieve an understanding of which elements of a multi-faceted dietary modification intervention were most influential in affecting disease incidence. Several challenges occur in mediation analysis: (1) the longitudinal and observational nature of the dietary variables and BMI/weight; (2) the measurement error in dietary variables which are often assessed using food frequency questionnaires; (3) control of measured/unmeasured confounders. In this dissertation, we proposed a general potential outcome framework for causal mediation analysis with failure time outcome and longitudinal mediator/exposure with measurement error. We proposed a method to correct for the systematic bias in longitudinal self-reported dietary data and use the calibrated data to estimate parameters in the survival model. We also proposed a robust estimator of key survival model parameters that can accommodate the existence of certain types of unmeasured confounders. We studied the performance of regression calibration methods for multiple choices of survival models numerically. We analyzed some important epidemiologic data and provided scientific information on the interplay between dietary exposures, physical activity and BMI in relation to site-specific cancer and other chronic diseases.
- Biostatistics