Improving Precision in Mediation Analysis Through Efficient Use of Case Data Open Access

Fang, Xi (2017)

Permanent URL:


Mediation analysis refers to a set of statistical techniques designed to explore the relationship between a dependent variable and one or more independent variables of primary interest, while also accounting for one or more intermediate (or mediating) variables. A traditional method for mediation analysis is the structural equation model (SEM) which is used to test for the existence of mediators. However, this SEM model can not be used to test the significance of the indirect effect generated by the effect of exposure on the outcome through a mediator. An alternative approach is based on causal inference developments. In the method suggested by VanderWeele et al., the total effect can be decomposed to the sum of the natural indirect effect (NIE) and natural direct effect (NDE). In this thesis, a new method based on results given by Satten and Kupper[9] and Satten and Carroll[10]is introduced to estimate the indirect effects more precisely in the case-control setting by making more efficient use of data on cases. This approach is compared with the VanderWeele proposal in terms of the precision of estimated direct and indirect effects. In this new method, the multivariate delta method was used to estimate the variance of log transformed causal effect estimates based on maximum likelihood and corresponding 95% confidence intervals and assessed for coverage of the true value. After multiple simulation studies, we found that in the model without interaction between exposure and mediator, the Satten method performed well in terms of precision for estimating causal effects. If interaction between exposure and mediator exists, although new method can estimate causal effects more precisely with small sample size, the distributions of these effects were left skewed. When increasing the sample size, the distributions were closer to normal as expected, but the difference in precision of causal effect estimates between these two methods was largely decreased. The simulations suggest that the Satten proposal attains better precision when interaction was not present. When interaction was present, the methods performed similarly, with the Satten approach showing some potential benefits when sample size is relatively small.

Table of Contents

1. Introduction

1.1 Overview of Mediation analysis

1.2 Structural Equation Modeling

1.3 Causal Inference Considerations

2. Methods

2.1 Estimation of Coefficients by Maximum Likelihood (the Satten Method)

2.2 Variance of causal effect estimates based on the delta method

2.3 Simulation Studies

3. Results

4. Conclusion

5. Discussion


About this Master's Thesis

Rights statement
  • Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.
Subfield / Discipline
  • English
Research Field
Committee Chair / Thesis Advisor
Committee Members
Partnering Agencies
Last modified

Primary PDF

Supplemental Files