18.6 Assumptions and Limitations of randomized experiments

• SUTVA -“Stable unit treatment value assumption”

  • No interference among units (spillover)
  • No hidden versions of the treatment - all units receive same well-defined treatment

• External Validity

  • Internal validity - design is able to recover causal effects in the sample
  • External validity - extrapolating this to the greater population of interest.
  • Example “threat to validity” - Interactions between participation and the researcher or the subject.
    • Collecting a food diary effects the diet.
    • Assessor on antidepressant medication might be biased by knowledge of who has the medication
  • “Double Blind” studies can avoid some of this, but not always possible.

• Missing data

  • Missing pre-treatment variables can be ignored since they are independent of treatment assignment
  • Missing outcomes and non-compliance are more problematic as they introduce biases. More discussion (and potential solutions) of this in Chapter 20 and 21.