5.1 Why can’t we always experiment?

Example: Did the way President Biden went about troop removal in Afghanistan cause the government to fail?

• Our outcome: The stability of the Afghanistan government
• Our causal effect of interest \(Y_i(1) - Y_i(0)\):
  • The Afghanistan government based on President Biden’s troop removal plan in 2021 vs.
  • The Afghanistan government based on an alternative troop removal plan.

What would be our ideal experimental design? We’d want to randomly assign the treatment: troop removal in Afghanistan?

• We probably would not want to randomize that.

Example: Do political leaders tend to matter for democracy?

• Our outcome: how democratic nations are
• Our causal effect of interest:
  • On average, how democratic nations are with their current leaders -
  • On average, how democratic nations would be with different leaders
• Possible Experimental Designs to randomly assign half of countries to receive a different political leader
  • Rig elections (I.e., Election fraud- Illegal, unethical)
  • Forcibly remove half from office (Probably illegal)
  • Assassinations (Illegal, Immoral, Unethical, etc.)

Again, we have problems!!

5.1.1 What can we do instead?

Let’s say we want to make a causal claim about the effect of one variable on an outcome, but we can’t think of an experimental design that will help us estimate this.

What do you do?