Causal models for qualitative and mixed methods inference
Exercise 2: Bayes
1 Part 1: Queries (30 mins)
Return to the model that you created in Exercise 1. Now write down at least 3 queries that can be defined for this model and that might be of substantive interest.
You should write down each query in two ways:
In plain, but maximally precise, plain language.
Formally using
CausalQueries. See the queries handout for examples. Be sure to distinguish between controlled interventions (using[X=...]) and simple conditioning (using “given…”).
2 Part II: A causal chain (30 mins)
Imagine a causal chain: \(A\) causes \(B\) and \(B\) causes \(C\)
Say you knew for the \(A\), \(B\) relationship:
- there is a 50% chance that \(A\) causes \(B\) for a given unit.
- there is a 25% chance that \(B=0\) regardless of \(A\)
- there is a 25% chance that \(B=1\) regardless of \(A\)
Say you knew for the \(B\), \(C\) relationship:
- there is a 50% chance that \(B\) causes \(C\) for a given unit.
- there is a 25% chance that \(C=0\) regardless of \(B\)
- there is a 25% chance that \(C=1\) regardless of \(B\)
Now:
- What is the probability that \(A\) causes \(C\)?
- What is the probability that \(A\) causes \(C\) for a case in which \(A=1\) and \(C=1\), when \(B\) is not observed?
- What is the probability that \(A\) causes \(C\) for a case in which \(A=1\) and \(B=1\) and \(C=1\)?
- What is the probability that \(A\) causes \(C\) for a case in which \(A=1\) and \(B=0\) and \(C=1\)?
Summarize: when is \(B\) informative for the \(A \rightarrow C\) relationship.
3 Group discussion (30 mins)
4 Additional (if you have time!): Bayesian sequencing and posterior variance (30 mins)
Say you knew the probability of \(Y\) given \(X_1\) and \(X_2\):
| X_1 = 0 | X1 = 1 | |
|---|---|---|
| X_2 = 0 | .2 | .5 |
| X_2 = 1 | .8 | .9 |
In addition you know that \(X_1 = 1\) with probability 0.5, and \(X_2 = 1\) with probability 0.5, independently.
- Say you do not know \(X_1\) or \(X_2\). What is your prior that \(Y = 1\)?
- How uncertain are you? (You can use variance as a measure, where the variance in beliefs about a bernoulli event that arises with probability \(p\) is just \(p(1-p)\))
- Say you now learn \(X_1 = 0\): what is your posterior belief now?
- What is your uncertainty now? (Again use variance, but now you should look at the posterior variance). Has it gone up or down with the additional information?
- Say in addition you now learn \(X_2 = 1\): what now is your posterior? What now is your uncertainty?
- Ex ante you cannot tell what you would observe if you looked for \(X_1\) or \(X_2\). Thinking through all the patterns you might see, what is the expected estimate you would get if you just looked for \(X_1\) or just looked for \(X_2\)? What is the expected posterior variance? If you had to choose just one of \(X_1\) or \(X_2\) to look for, which would be more informative?
Bonus: Think through how things might change if X1 and X2 were not independent?