About the Program

Mechanisms for Social Science: Applied Tools for Causal Mediation Analysis

Overview

Mediation is central to understanding mechanisms in the social world. Yet much empirical research relies on outdated statistical methods. This workshop teaches state-of-the-art causal mediation tools with a focus on intuition and real-world applications. Through real research examples and hands-on data analysis, participants will develop the skills to apply modern mediation methods to their own research.
 

Workshop Goals

Mediation analysis is central to sociology and all social and biomedical sciences. Mediation analysis tells us how and why causal effects come about, allows us to test the causal mechanisms specified by social theories, and provides an understanding of how policies and other interventions have the effects that they do. Consider these examples:

 

 

Immigration

Negative media coverage reduces public support for immigration. But what explains this effect? Do news stories trigger diffuse anxieties, stoke economic fears, or activate racial resentment? How can the researcher disentangle these causal mechanisms empirically?

 

Health
Mothers’ smoking increases newborn mortality overall, but apparently not among low-birthweight babies.  Does this mean that mothers who are predicted to give birth to low-birthweight babies should start smoking? The answer rests on avoiding a common analytic error in conventional mediation analysis.

 

Social mobility: Intergenerational mobility is higher among college graduates than high-school graduates, leading many to label education the "great equalizer." But what assumptions does this analysis make about confounding, and how can we examine the robustness of these findings when key assumptions are violated?

 

This intensive 3-day workshop will provide participants the understanding, software, and guided practice to answer mediation questions like these and many others. The workshop will focus on modern causal mediation methods that have been developed over the last decade in the social sciences and statistics to better pose and answer research questions about mediation. These recent advances carefully define different types of mediation effects, specify the assumptions needed to draw firm conclusions about mediation, provide methods for estimating mediation effects, and tell us what to do when assumptions may be suspect.

Workshop instruction will focus on understanding key concepts, developing intuition for assumptions and effects, and providing the tools to perform modern mediation analysis effectively in one’s own research. The course starts from the classic mediation methods familiar to social scientists and develops participants’ knowledge up to the research frontier of causal mediation analysis, drawing on recent developments by VanderWeele, Pearl, Robbins, and others. The instructors will teach using real research examples, and the workshop will provide opportunities to practice applying and implementing mediation analysis in Stata (with R code also provided).

Topics

Program activities will run from 9:00am - 4:00pm with a 12:00 - 1:00pm lunch break. 

Day 1: Wednesday August 5

  • Introduction to causal inference: potential outcomes and graphical models
  • Mediation estimands: varieties of direct and indirect effects
  • Mediation vs. moderation
  • Identification assumptions, unobserved confounding, and treatment-induced confounding

Day 2: Thursday, August 6

  • Analyses with a single mediator
  • Regression-based estimators for continuous and categorical variables
  • Interaction effects in mediation analysis
  • Dealing with unobserved confounding: quasi-experimental solutions and sensitivity analysis.

Day 3: Friday, August 7

  • Analyses with multiple mediators
  • Time-varying treatments and mediators
  • Regression-based estimators, inverse probability weighting, sequential g-estimation, regression-with-residuals


Prerequisites

No prior experience with causal inference is required. To get the most out of the workshop, participants should have a good understanding of multivariate regression analysis equivalent to a first-year graduate course in quantitative methods. Instruction will focus on implementation with Stata, but examples and exercises will also be offered in R. A working knowledge of either program (but not both) will be helpful.
 

Register Now