Guiding Questions
- Let’s think about the word “cause” (and related phrases like “results in” and “leads to.”) What do you think “cause” means? How do you/people you know use it in day-to-day conversation and thinking?
- Why do we want to learn about causes? What should our goals be in learning about causes?
- “Correlation does not imply causation.” Has this saying come up in conversations occasionally? If so, how? How do you feel about the value of this saying?
- Is the definition of a potential outcome and a causal effect satisfying to you? How do these notions relate to what you had thought about causes and causation previously? (Reflect on and connect to our Day 1 activity.)
- Let’s think about exchangeability in big picture terms. Setting aside the technical definitions in terms of (conditional) independence, what is exchangeability trying to describe? Try to find a good word/phrase (not “independent”) to complete the following sentence: “If the treated and untreated groups are exchangeable, we could also describe them as being …”
- How can we use the concepts of marginal and conditional exchangeability to make statements about potential outcome distributions in the treated and in the untreated?
- Marginal and conditional exchangeability are assumptions–do you think these assumptions are testable using data? Why or why not?
Extra self-check resource: Moodle checkpoint on Causal Effects and Exchangeability
- Why are randomized experiments often called the “gold standard” for causal inference? What is the relationship between randomized experiments and exchangeability?
- Consider the quasi-experimental designs discussed in the video. How do these designs try to mimic randomization? What assumptions do they make to allow making statements about missing potential outcomes?
Extra self-check resource: Moodle checkpoint on Study Designs for Causal Inference
Causal Graphs as Statistical Models
- Why is it sensible for causal graphs to be directed and acyclic?
- How can “feedback loops” be represented in causal graphs?
- What extra information does a structural equation model encode that a causal graph alone does not?
Extra self-check resource: Moodle checkpoint on Causal Graphs as Statistical Models
Key Structures in Causal Graphs
- What marginal and conditional (in)dependence relations pertain to chains, forks, and colliders? Do these relations make sense? What examples (in the video or your own) help rationalize these relations?
Extra self-check resource: Moodle checkpoint on Key Structures in Causal Graphs
Graphical Structure of Confounding
- What is the relationship between d-separation, causal/noncausal paths, conditional exchangeability, and estimating causal effects?
Extra self-check resource: Moodle checkpoint on (Non)causal Paths and D-Separation
Graphical Structure of Selection Bias
- How is selection bias different from confounding bias?
The Smoking-Birthweight Paradox
- What is the paradox/surprising finding at the center of this paper? What data are presented that illustrate this surprising finding?
- How do the authors use causal graphs to explain the paradox? Do you find the discussion of causal graphs helpful? Why or why not?
Building Causal Graphs for Applied Analyses
- The video describes a way to think through building a causal graph for an applied analysis. Do you agree with the rationale for the thought process? Are there any other aspects of the graph-building process that you have questions about? (e.g., How would a given step work out in practice?)
Regression for Estimating Causal Effects
- What is the connection between conditional exchangeability and the variables needed to include in a regression model for estimating causal effects?
- What steps can we take to ensure that the form of our regression model is as accurate as possible?
- What concerns arise when trying to interpret all coefficients (not just the treatment coefficient) from regression output?