,
4 tweets,
1 min read
Read on Twitter
Happy to share a new working paper with Philip Dawid and Monica Musio on when you can use information on causal processes to make claims about "causes of effects"
(is Y due to X? / process tracing / causal attribution)
arxiv.org/abs/1907.00399
Highlights:
1/n
(is Y due to X? / process tracing / causal attribution)
arxiv.org/abs/1907.00399
Highlights:
1/n
* Experiments focus on treatment effects but we often care about whether an outcome is due to a cause
* That's a harder question and answer usually not identified by experimental data
* But don't obsess about identification: you can learn lots even if estimands are not identified
* That's a harder question and answer usually not identified by experimental data
* But don't obsess about identification: you can learn lots even if estimands are not identified
* Knowledge of mediation processes can tighten bounds even if you cannot observe the mediators.
But:
1 Getting arbitrarily "close" to a causal process does not render causal effects observable
2 Process data better for disconfirming causal relations than for confirming them
But:
1 Getting arbitrarily "close" to a causal process does not render causal effects observable
2 Process data better for disconfirming causal relations than for confirming them
3. Max learning arises from short processes (when X is a necessary condition for a sufficient condit'n for Y)
4. Understanding conditional fx can tighten bounds more than knowledge of mediation
Reminder for me how much to learn working with great people outside your discipline
4. Understanding conditional fx can tighten bounds more than knowledge of mediation
Reminder for me how much to learn working with great people outside your discipline
