Reading focus
Syllabus: What Would Justify a Causal Claim?
Separate the assumptions needed for a total effect from those needed for an indirect effect.
Why you’re reading it. Keep a plausible developmental explanation distinct from evidence that identifies its pathway.
A question to consider. If income payments change both parental supervision and children’s symptoms, what additional design could test the supervision explanation, and what new assumption would that design introduce?
Why this reading
Bullock, Green and Ha examine the additional assumptions needed to move from an intervention’s effect to an explanation of how it worked. Their examples make clear why randomizing a treatment does not automatically identify its mediator.
Before you read
A mediator is a proposed intermediate step linking a treatment to an outcome. A total effect concerns the overall contrast; an indirect effect concerns a specified pathway. A variable measured after treatment can reflect other causes as well as the treatment itself.
Read for
- Identify which relations remain unrandomized when only the initial treatment is assigned.
- Consider why manipulating a proposed mediator can introduce other pathways or affect a different group of people.
- Follow how variation between people complicates simple products of average path coefficients.
Consider afterwards
- A randomized intervention changes both confidence and performance. What additional comparison would help distinguish confidence as a mechanism from confidence as another consequence of the intervention?
Editorial context
The critique challenges easy identification of mechanisms, not the possibility of learning about them. Measured mediators, manipulated mediators and heterogeneous effects pose different problems. The authors recommend a research program with explicit assumptions rather than a single decisive statistical routine. Bullock and Green’s later design-based mediation proposal (2021) also requires scrutiny of what the experimental contrasts identify.