Section 5 Causality with Non-Experimental Data

In this section, we continue to evaluate causal claims, but this time we will not have the benefit of experiments.

Recall: Why do we use experiments?

We want to evaluate causal claims:

• Does manipulating one factor (a “treatment”) cause a change in an outcome? (\(Y_i(1) - Y_i(0)\))
  • But we have a problem: the fundamental problem of causal inference
  • (Can’t simultaneously both be treated and untreated - e.g., you can’t simultaneously be contacted and not contacted by a campaign)
  • So instead, we randomly assign some units to receive a treatment, and some not to, and then compare their average outcomes in an experiment

And because of random assignment of the treatment, we can be confident that the groups are similar EXCEPT for the treatment

• Therefore, any difference between the two groups in average outcomes can be attributed to the treatment

But what if we can’t randomize the treatment?