How does it work?
- source of variation that allows you to isolate just the front-door path
- similar to an RCT
- there: randomized assignment to treatment
- here: source of randomization without backdoors
- moves the requirement from closing back doors Treatment –> Outcome to Randomization –> Outcome (and front doors that don’t go through Treatment)
- Opposite of controlling: only variation coming from instrument
- Isolate part of treatment driven by instrument
- Use instrument to explain treatment
- Remove any treatment parts that are not explained by instrument
- Use instrument to explain outcome
- Remove any part of the outcome that is not explained by instrument
- Look at relationship b/w remaining outcome and remaining treatment
