Lecture 3: Causal inference with observational data

Question 1

Natural Experiment Definition

Answer 1

Use of a naturally occurring event

Question 2

Advantages natural experiment

Answer 2

• Relatively high external validity and thus applicable to the real-world setting
• Internal validity can be high if assumptions are met and study is conducted
  with scientific rigor
• Results can be communicated in an intuitive manner in general (e.g., ideal for
  evidence-based interventions, practitioners)

Question 3

Assumptions Natural Experiment

Answer 3

• Treatment and control group
• Random assignment
• Researchers and subjects have no control
• Can be used in various designs

Question 4

How to identify opportunity for natural experiment?

Answer 4

• Read opportunities
• Keep yourself informed
• Intended or unintended effect
• Population level
• Confounding variables
• Map out the casual graph

Question 5

How to ensure methodological rigor?

Answer 5

• Select opportunities for natural experiment wisely
• Increase your rigor by involving comparison and time-series data
• Compare characteristics between treatment and control groups
• Use matching if necessary
• Find additional evidence in data to support research conclusions

Question 6

Instrumental variable assumptions

Answer 6

1. Observable
2. Related to but does not have a causal effect on
   the outcome
3. Causal effect on the treatment/independent variables
4. Randomly assigned
5. Monotonicity (no-defiers)
6. No confounder -> Z is related to Y only via X

Question 7

Instrument variable conclusions

Answer 7

• Instrument variable allows the estimation of causal relationship if
  assumptions are met
• Trickiest part is fulfilling the causal relationship
• Most assumptions can be assessed
• Intensively studied with machine learning method

Question 8

Regression discontinuity design features

Answer 8

• An assignment of individuals to treatment or control group based on certain
  threshold
• Only difference between them is the treatment they receive
• Most attractive when a threshold is used
• Can be randomized
• Requires larger sample size

Question 9

Assumptions of RDD

Answer 9

• The treatment is assigned by an observable variable, in which a
  discontinuity exists for the group assignment
• Outcome and assignment is continuous
• Cut-point is determined independently
• Functional form of the regression is specified properly

Question 10

Sensitivity analysis/ Robustness check

Answer 10

• Compare various models
• Scatterplot to identify patterns
• Explore various models
• Try overfitting
• Cross validate

Question 11

RDD conclusions

Answer 11

• Estimation of causal relationship
• Important to test multiple functional forms
• Generalization can be limited by the cut-off value
• Evolving rapidly overtime

Question 12

Difference-in-differences features

Answer 12

• Panel/ longitudinal data

- Requires both treatment and control groups